Ivan Nosov · Writing

The tasks worked. The job did not.

i.nosov9@gmail.com (Ivan Nosov) — Sat, 10 Oct 2026 00:00:00 GMT

The question I keep coming back to: how do you get AI to do the whole job, beyond the individual tasks, and do it consistently at scale?

Epoch AI published a useful reference on 8 October. The research group, which measures how fast AI capabilities grow, tested six models, including Fable 5.1 and GPT-6 Astra, on real tasks from its own work - charts, analysis, research design. The models got the context a colleague would get and no human help, and the results were graded against Epoch’s own staff standard.

The tasks worked. Coding, analysis and operating a computer were reliable, and better than Epoch expected.

The job did not. The models failed in three places:

Epoch's own standards: even with many examples, they missed the conventions a new colleague learns quickly, like the visual style and what the audience cares about.

2.. Designing experiments: good research ideas, but tests that did not measure what they claimed.

Their own mistakes: flawed setups presented as key findings.

For me this is a useful map of where the work is. The first gap is the one I think about most, because every company runs on it: how the month-end report should look, which supplier always gets an exception, what a hiring manager really looks for. People learn this by watching colleagues, while a model only works with what it is given.

So the way to the whole job goes through making those rules explicit, giving them an owner and keeping them current, plus building checks for the other two gaps, because a model will not catch its own flawed setup.

That is the organisational capability everyone talks about, broken down into work that can actually be done.

OpenAI's 722 maths papers and the shape of jobs

i.nosov9@gmail.com (Ivan Nosov) — Sat, 10 Oct 2026 00:00:00 GMT

This week mathematics showed how the shape of many jobs will change, and it happened faster than anyone planned for.

On Tuesday, 6 October, OpenAI published 722 mathematics papers written by an internal model it has not released - 372 families of results, from about 4,000 problems it gave the model.

Two days later Tristan Buckmaster, a mathematics professor at NYU, said on CNBC: "We've had whole research programs wiped out on Tuesday by their dump".

His concern is early-career mathematicians, who build a career on a few open problems. Some of those problems are now closed, by a model nobody outside OpenAI can use. Mathematics is a clear case because results are public and dated.

In most jobs the change is slower and quieter, but it works the same way: a piece of work someone expected to own becomes cheap overnight. What I find useful is where mathematicians see the value moving.

Alex Kontorovich at Rutgers said mathematics teaches people "to think clearly and deeply and for extended periods of time on big problems", and he expects more demand for that skill.

OpenAI's own advisory group called the release "the beginning, not the completion, of the process of human understanding". Someone still has to check 722 papers, decide which results matter and build on them.

This is close to how I think about reskilling. In some areas people will work with the model and do much more. In others the model will do the work, and pretending otherwise helps nobody. So the useful step is an honest look at which part of a role is which, and then putting the training time into the part that grows: choosing the problems worth solving, judging whether a result holds, connecting it to what the business needs.

For early-career mathematicians this week is painful. For the rest of us it is an early look on what to expect, while there is still time to plan for it.

The Italian Hackathon League final at OGR Torino

i.nosov9@gmail.com (Ivan Nosov) — Fri, 09 Oct 2026 00:00:00 GMT

An unusual start to the day: the final of the Italian Hackathon League, at OGR Torino during Wave.

One challenge we haven’t seen yet, five hours to turn it into something that works, and a jury evaluating the scenarios on an NVIDIA DGX Spark.

I like challenges, especially the ones well outside my day-to-day.

Fittingly, the venue is the old railway repair works: the place was built for fixing machines under time pressure.

Let’s see what we build by the end of the day.

FAIR Barcelona 2026: keynote recap

i.nosov9@gmail.com (Ivan Nosov) — Wed, 07 Oct 2026 00:00:00 GMT

Last Thursday I opened my speech at FAIR Barcelona with a question I have been solving for a while: what would it take to build JARVIS for HR?

JARVIS is helpful as it knows Tony Starks world and can interact with the tools around him. HR has all that context too, it just lives in policies, old decks and experts' heads, and every new case means someone has to stitch it back together.

So the idea I shared was a digital twin of the HR function, one that carries that context and the judgment behind it into the work itself. Done well, it takes most of the stitching off peoples plates.

We now know numerous stories of senior HRs saving days of their time on processing, and that the time went back into the conversation with the business.

That led to the challenge I left with the room. We keep hearing that AI brings no return, and maybe thats because we judge the tool as if it were the product. The product is the work. Find one friction, fix it, measure it, then help people fix the next one on their own.

Thanks to everyone I met in Barcelona for the conversations, the pushback and the good questions. And a big thank you to the Fair Cultures team for having me and for such a well-run event.

Everyone becomes a bit of an engineer

i.nosov9@gmail.com (Ivan Nosov) — Tue, 22 Sep 2026 00:00:00 GMT

Almost everyone who gets serious about AI eventually becomes a bit of an engineer.

People who never thought of themselves as technical start taking their own work apart.

What are the inputs?
Where does judgment enter?
What does a good output look like?
Which exceptions matter?
How should the process run next time?

This is also why the corporate obsession with AI use cases is approaching a dead end.

When building was expensive, selecting a small number of use cases made sense. Now almost every recurring piece of knowledge work can be redesigned. The list gets longer while telling us less.

The real capability is emerging underneath: people close to the work learning to describe it, test it and gradually engineer a better version of it. The company's job is to give them the tools, guardrails and ownership to keep going.

The industrialization of knowledge work may arrive this way - through thousands of small acts of engineering by people who never called themselves engineers.

Twenty supplier offers: where the change AI brings is hard to reach

i.nosov9@gmail.com (Ivan Nosov) — Mon, 14 Sep 2026 00:00:00 GMT

Fundamental change that AI brings might sound simple, yet it is hard to reach.

Imagine a company buying glass bottles. Twenty supplier offers arrive as PDFs, spreadsheets and emails. Some omit delivery costs, others use different units, and the contract terms conflict.

Someone in the team has to read each file, reconcile the differences and revisit the comparison whenever a supplier sends a revision.

An AI-assisted workflow could prepare that comparison across the whole set: extract prices and terms, calculate costs using checked formulas, retain links to the sources and flag missing or conflicting information for review. A revised offer would trigger the relevant calculations and checks again.

The same rules can run across every submission and revision. The team defines and improves the method, rather than reconstructing the comparison file by file.

That changes the work itself. People spend more of their attention checking exceptions, challenging assumptions, negotiating and deciding. Their expertise also becomes part of how the workflow operates.

A buyer knows that a price without delivery costs is incomplete. That becomes a check. A review exposes another omission; the correction becomes a tested rule for the next purchase. The method gets better because someone who understands the work keeps improving it.

Leadership provides the tools, data access, time and technical support. The people closest to the job define what a good result requires.

This is how AI can change a team's role: from preparing each analysis to owning a method that can handle a larger, messier workload.

Do you recognize this change happening?

The more I delegate to AI, the more distilled knowledge is worth

i.nosov9@gmail.com (Ivan Nosov) — Tue, 08 Sep 2026 00:00:00 GMT

The more work I delegate to AI, the more valuable the knowledge I have taken the time to distil becomes.

That is what connects two things I am building: a personal knowledge platform and a way of organising AI work through graph engineering.

Here, graph engineering means turning a process into an executable network of tasks. Each node defines an assignment, the knowledge and tools it needs, and the conditions for completion. The connections govern dependencies, parallel work, handoffs and routes for repair. Human decision points define where the system must stop.

The graph gives the work a structure. My personal knowledge platform supplies its substance: experience, research, decisions and corrections distilled into reusable principles, evidence, examples and standards of judgment.

Distillation takes work. It means deciding what a source actually supports, separating a durable principle from a one-off situation, preserving the reasoning behind a decision, connecting related ideas and keeping them current. Those choices determine what an agent can use without asking me to reconstruct the context each time.

The connection between the two layers matters. A researcher gets relevant sources and evidence rules. An editor gets my reasoning principles, audience context and voice. A reviewer gets criteria for challenging the result. Each assignment receives the knowledge it needs, while independent research stays open to conclusions I might not prefer.

My narrative-product workflow is one example. It develops research and a counter-case separately, compares storylines, builds the chosen direction and checks the result. I used it for this post. An unsupported claim returns to evidence; a broken layout returns to design. The correction has an address.

I keep developing both layers through use. A correction may expose missing knowledge, a weak instruction or a faulty handoff. Reviewing it can improve the knowledge base, the graph, or both. More autonomy depends on that continuing work of curation, testing and judgment.

For HR and the future of work, the implication is substantial. Expertise has to become usable beyond the person holding it, while accountability remains clear. Functional experts, HR and technical partners have connected responsibilities in making that possible.

Designing AI work increasingly means designing how professional knowledge is captured, applied, challenged and improved.

Databricks as an AI workbench for knowledge work

i.nosov9@gmail.com (Ivan Nosov) — Thu, 03 Sep 2026 00:00:00 GMT

Databricks became an end-to-end AI workbench for knowledge work. More than 100 knowledge workers now use coding-class agents on it against challenges of HR, Finance, Commercial, and Marketing.

Finance reconciliations leave private spreadsheets and remain as jobs. Commercial challenges start to find a way out of one-off decks to become repeatable jobs and autonomous agentic flows.

HRs start to productize solutions that would require lots of vendors, time, and spend, while also automating the routine.

This spring I was looking for a way to bring the power of coding agents to knowledge workers. After months on Claude Code I knew the capability was real. I did not yet have a way to give it to people inside the company without pointing a laptop at production. I went in cautiously and found the solution in Databricks, which was not the place I expected to look.

The platform is sold as a lakehouse - pipelines, SQL, dashboards. Genie Code sat on top of that stack just freshly released in March this year. The lake sits beside the work. The sandbox holds the guardrails. Specs, context, jobs, and apps live in the same environment, so a useful sequence does not die as a chat transcript. That is why wide adoption was possible.

The surprise showed up after people were on the surface.

They picked up business problems that used to wait for a specialist. Capability grew because the workbench could execute. Appetite grew with it. The first use case did not close the need. They brought the next one, then the one after that.

I went looking for a way to transfer a tool.

What scaled was the appetite for the next problem.

OpenAI limits Cursor’s access to its models

i.nosov9@gmail.com (Ivan Nosov) — Sat, 29 Aug 2026 00:00:00 GMT

Now, this is Big.

OpenAI is trying to curb Cursor's growth by limiting access to its models.

We can say anything about both Sam Altman and Elon Musk, but clearly they have a rough relationship, now it has impact on infrastructure of AI solutions.

Establishing dominance during a period of low trust between providers is a dangerous path.

For me, though it just confirms the future of Open-Weight models and minimizing dependence on specific vendors.

It appears that my dual Codex + Cursor setup makes even more sense

A watermark on AI text only hurts if AI help is something to hide

i.nosov9@gmail.com (Ivan Nosov) — Tue, 18 Aug 2026 00:00:00 GMT

A watermark on AI text only hurts if being AI-touched is something to hide.

Anthropic turned on invisible text watermarking for Claude this month, a way to solve for EU AI Act requirement in force since August 2. The feed filled with cancellation threats. Nothing about the model changed.

Now you can just be open about AI everywhere.

The draft, the edit, the summary. The same tools sit in every workflow, and have for some time. Watermarked text mostly means Claude processed it. Unmarked text probably means another model did. Neither says who owned the call, or whether the judgment behind it was any good.

Panic about the mark only appears if the assistance was supposed to stay off the record: keep the speed of the model, keep the credit of a blank page. Those two claims cannot both be true.

The technical objections arrived next.

The mark only acts where the model has free choice, a comment or a synonym, so code that has to compile is largely untouched.

Google has used the same class of watermark on Gemini for two years, tested on about 20 million responses, and reported no measurable drop in quality. The signal sits in word choice, not in the reasoning that produced the answer.

If the analysis is right and the judgment is yours, the adjective Claude picked is a detail.

Demanding that AI use stay a secret is the part that does not add up.

For research and writing for this post a mix of Grok 4.6, Gemini 3.1 Pro, Kimi K3 Max was used through Cursor. So probably no watermarks, but the AI is clearly here.

Every correction to an AI answer is a process rule you paid for

i.nosov9@gmail.com (Ivan Nosov) — Wed, 22 Jul 2026 00:00:00 GMT

You know the ritual. The first AI's output is too long. The wrong data was used. The structure misses the point. You spend 20 minutes correcting it until the result is finally useful.

A month later, you start a new chat and reconstruct the same instructions again.

Every correction in that first conversation was a process rule you paid to discover: which source takes precedence, which cases to exclude, what good looks like, and how to check the answer.

Save those rules as a specification.

This is one of the first things I teach people when they embark on the AI Capabiltiy journey.

You can write it manually. You can ask the assistant to draft it before starting. Or, often more usefully, ask for it after the task is complete:

Review the work we completed and my corrections. Turn the accepted approach into reusable instructions for a fresh chat. Capture the rules we discovered, the expected output, and how to verify it.

Then keep updating the file. Whenever a fix is needed, ask for the general rule and a test that would catch the same failure next time.

Use a referenced document for occasional repetition, project instructions for a recurring workflow, or a skill when the activity travels across projects and people.

The durable productivity gain comes from paying once for each correction.

Microsoft's biggest obstacle to AI ROI may be Microsoft Office

i.nosov9@gmail.com (Ivan Nosov) — Fri, 17 Jul 2026 00:00:00 GMT

Microsoft’s biggest obstacle to extracting ROI from AI may be Microsoft Office.

Copilot entered the world’s dominant productivity suite with a distribution advantage nobody else could replicate. More than two years later, Microsoft reported over 20 million paid Copilot seats against more than 450 million commercial Microsoft 365 seats - at most 4.4%. The denominator turns an impressive headline into weak penetration: over 95% of the installed base still had no paid Copilot license.

The architecture explains part of the gap. Copilot began as an assistant inside applications, grounded through Microsoft Graph. OpenAI and Anthropic moved toward an LLM-native workspace: a project folder containing selected sources, stable instructions and tools, with permission to execute multi-step work.

Microsoft is now catching up. Copilot Cowork brings multi-file tasks, folder access and reusable skills into its environment, while retaining two constraints created by Microsoft itself.

The first is context. A month-end close needs 20 approved files, one mapping table, explicit rules, known exceptions and a defined output. Broad retrieval across SharePoint, Teams and Outlook adds ambient context - drafts, duplicates and outdated decisions. Relevance does not establish authority.

The second is format. .docx, .xlsx and .pptx are ZIP packages containing XML. An agent must extract content, reconstruct structure, preserve formulas and formatting, then validate the rendered output. With .md, .html, .csv and .json, the content and structure are directly readable and editable. The difference is an entire conversion and rendering layer across every file, not a marginal token saving.

I removed manual Excel and PowerPoint editing from my workflows more than six months ago. I dread the rare task that forces me to touch either by hand - it is slow, brittle and unproductive. My working layer is now Markdown, HTML, CSV and JSON, with Office formats generated only when someone needs the final artifact. I have never looked back.

Microsoft therefore pays a translation tax from Office formats and an ambiguity tax from ambient organizational context. Both appear as more tokens, more actions, longer execution, higher consumption cost and more human review.

Cowork fixes the direction of travel, while the economics still run through an architecture built before LLMs.

ROI from enterprise AI will be decided by the cost of constructing trustworthy context. Everything else is noise.

Digital transformation now sits in every leadership job

i.nosov9@gmail.com (Ivan Nosov) — Tue, 14 Jul 2026 00:00:00 GMT

Digital transformation now sits in every leadership job, whether the job description admits it or not.

The boundaries between strategy, technology, operations, customer experience and organizational design keep thinning. Digital fluency has to become an organizational property. Each leader needs enough range to work across those seams and remain accountable for the choices made in their own part of the business.

Last week we've spent the Digital Week at SDA Bocconi. It moved across AI, digital sovereignty, manufacturing, customer engagement, spatial computing, blockchain and the human side of change. The subjects looked separate on the schedule; inside a company, they collide.

The Sapio case forced our group to build one coherent response across these domains. We brought together participants from Bocconi and ESADE, and the mix worked exceptionally well.

We also had excellent sessions with SDA Bocconi faculty and leaders from Sapio, Nexi, Prysmian, Accenture, Mumzworld/JADA Squad and AnotheReality, plus a day hosted at Cisco.

Completing the concentration takes me beyond halfway through the Executive MBA. The demanding case, very good company and great pacing made it a week I will remember.

Model rankings shift; my workflows stay provider-neutral

i.nosov9@gmail.com (Ivan Nosov) — Fri, 10 Jul 2026 00:00:00 GMT

ChatGPT 5.6 is fantastic, Grok 4.5 is the working horse.

In tokenomics time - Antropic loses the edge.

I was already prefering Cursor + Codex combo, and now it became even stronger.

Thankfully none of the workflows, skills, ways of working is specific provider dependant and is a simple lift-and-shift,

But clearly I would recommend Cursor for knowledge work these days for flexibility + multi model access at will.

Will it stay? Of course not.

Get used to joggling tools.

Start with the output and let it pull the data

i.nosov9@gmail.com (Ivan Nosov) — Thu, 09 Jul 2026 00:00:00 GMT

The dominant narrative in the feed repeats the same rigid prerequisite: you cannot adopt AI until your data is perfectly clean.

This premise shrinks AI to a narrow set of safe use cases and creates an artificial barrier to building real capability. We treat enterprise data like a concrete foundation that must be poured perfectly before a single brick is laid.

In reality, data behaves like a pressurized system. You do not find the leaks by inspecting dry pipes in the dark. You turn the pump on, apply pressure, and watch where it breaks. The AI output is the pressure.

You do not need "good enough" data to start, because good enough is an illusion until the data is actually used.

Until we began chatbot adoption, policy coherency was largely invisible. The moment AI tried to generate answers from those policies, contradictions and gaps surfaced immediately. The same happened when we data extracts became enabled by ad-hoc dashboarding: issues that had stayed hidden became operational blockers within days.

Because you can have rapid output, you have incentive to stabilize input till the system starts to "click".

I really like the Americana Foods example. According to Beam's case study, they started with month-end flash reporting while running a 20-30 year legacy on-prem Oracle ERP with no APIs and Excel-heavy reporting across business units and geographies. If they had waited for full architecture modernization, they would still be waiting.

Beam states the agent handled input collection, follow-ups, escalations, consolidation, and CFO report commentary, with human review before reports reached leadership. The same source says the workflow went live in production in under 30 days.

It solves the chicken-or-egg problem of AI readiness. You do not clean data to build a solution. You build a solution that makes data debt visible, then improve the foundation while the wheel is already turning.

The sequence flips: start with the idea, build the output, let the output pull the data.

When processing time collapses, the bottleneck moves to thinking

i.nosov9@gmail.com (Ivan Nosov) — Mon, 06 Jul 2026 00:00:00 GMT

When processing time collapses, the bottleneck moves to thinking.

Giac Soliman had me on the Range podcast on AI adoption in comp and HR - with Pay Equity as center showcase. The work actually changes once agents can run the middle of the workflow.

A few things we spent real time on:

the old model was input → processing → output; for most comp cycles, processing was the bulk of the job (templates, lookups, copy-paste, commentary assembly). That layer is shrinking fast

what replaces it isn't free time - it's validation, iteration, and specifying what you want before you rerun

how to get started inside a large company: prove it to yourself on synthetic data, then bring one ally, then slice the use case into steps IT and legal can approve

why domain expertise matters more, not less - you need to know what a good output looks like to catch drift

where this goes next: from individual productivity to shared team workspaces (the way engineering teams already work)

Pay equity and pay transparency were the concrete example throughout - building the analysis in-house, EU data handling, walking through the markdown repository on screen. Serious work, not a demo.

For comp and total rewards leaders figuring out adoption beyond the pilot phase.

Links in the first comment.

Thanks Giac and Range

10% × 10% × 10%: layered checks that make AI errors rare

i.nosov9@gmail.com (Ivan Nosov) — Thu, 25 Jun 2026 00:00:00 GMT

10% x 10% x 10% = 0.1% error rate.

That is the promise behind stable AI work.

AI hallucinations are real. The useful question is how to make them rare enough, visible enough, and contained enough that real work can move.

One model may miss. A second layer should catch most of what survived. A third layer should catch most of what remains.

That multiplication only works when each layer checks a different object.

That became the architecture question while building the HR Digital Twin from my last post. Once the system started producing real artifacts - executive decks, job offers, comp dashboards, policy answers - the design standard changed.

Where do we stop hallucinations before they become expensive?

The answer I've built around has two roles: Judge and Fixer.

The Judge checks the generated story before artifact generation. It asks whether the logic is grounded, complete enough, and safe to turn into a deck, dashboard, offer, or recommendation.

If the Judge fails the story, the system fixes the story and sends it back one step. The artifact is not generated from a weak narrative.

The Fixer works at the end. It checks the produced artifact against the accepted story and repairs final defects: one wrong number, one stale claim, one missing caveat, one bad wording choice.

Judge checks the story.

Fixer checks the artifact.

I see that common knowledge treats AI as a generator: prompt in, answer out, human reviews the damage. That works for low-risk drafting. It breaks when the output becomes a business artifact.

In engineering language, this maps to evaluator plus post-generation repair. For a business function, the language is simpler: do not generate the artifact until the story passes; do not deliver the artifact until the facts are fixed.

The 10% x 10% x 10% logic is not a benchmark, but the smart design intuition.

Stack different checks at the points where errors change form.

Before reasoning becomes a polished artifact.

Before the polished artifact reaches the user.

The question for every AI-first organization is now practical:

where is the Judge, where is the Fixer, and what work is allowed to leave the system without them?

Note: Infographic generated with Gemini based on my story.

What 40+ AI reports from March to June 2026 agree on

i.nosov9@gmail.com (Ivan Nosov) — Mon, 15 Jun 2026 00:00:00 GMT

I ran an analysis on 40+ AI reports and papers released March-June 2026.

The strongest consensus: AI value is moving from tool adoption to enterprise redesign.

Sources: Stanford HAI, Stanford Digital Economy Lab, Microsoft, KPMG, Accenture, BCG, Gallup, NBER, WEF, OECD and RAND/Oxford.

Different angles but same direction.

Task gains are real; workflow absorption is the bottleneck.

AI works in bounded tasks. Stanford HAI reports measured productivity gains in customer support and software development. The harder question is whether companies can absorb those gains into real workflows.

Personal productivity is rising; enterprise productivity is lagging.

Gallup reports that 65% of U.S. employees in AI-implementing organizations say AI helps their productivity. Only 12% say it has transformed how work gets done. NBER finds 69% of firms actively use AI, while 89% of executives report no labor productivity impact over the past three years.

Agentic AI exposes how poorly work is codified.

The useful question is practical: can the organization describe a bounded piece of work well enough for an AI system to execute with tools and controls? Most companies struggle here. Their work lives in emails, exceptions, tribal memory, unclear ownership and process diagrams nobody follows.

AI compresses junior work before career ladders adapt.

BCG estimates 50-55% of U.S. jobs may be reshaped over two to three years. The sharpest tension is junior work. AI removes tasks people used to learn from, while senior judgment becomes more valuable. That creates a capability pipeline problem.

Productivity gets exponentially harder across layers. Individual productivity is the easiest gain. Team productivity is harder because routines, standards and handoffs must change. Company productivity is harder again because incentives, budgets, governance and systems must move together. Managers sit directly in that gap.
AI needs connected infrastructure; silos slow it down.

Useful AI needs data, process, policy, security, legal, HR, IT and business ownership connected. Governance has to become execution infrastructure, not a late-stage approval ritual.

Models change; codified work is the durable asset.

Models will change. Vendors will change. The durable asset is codified work: clean knowledge, clear workflows, reusable controls, data definitions, evaluation patterns and escalation rules.

Employees outrun policy, training and job design.

WEF's learning research mirrors what companies see: people use AI before policy, training and job architecture catch up. The organizations that learn from real usage will move faster.

My takeaway for companies:

Stop counting pilots as progress. Pick priority workflows. Redesign them. Train managers. Make governance usable. Protect the apprenticeship layer. Build around work that can survive the next tool change.

The AI agenda is now about whether the organization can describe, govern and redesign its own work.

Two weeks testing Genie Code as a governed AI workspace

i.nosov9@gmail.com (Ivan Nosov) — Tue, 12 May 2026 00:00:00 GMT

I spent two weeks testing Genie Code for the job Databricks undersells.

The official pitch is data work: dashboards, pipelines, analytics, lakehouse productivity.

I used it as something broader: a governed AI workspace for corporate work that sits between systems and decisions.

That is where a lot of time disappears.

SuccessFactors exports. Country job files. Org charts. Capability decks. Meeting transcripts. PMO follow-ups. SharePoint folders. Excel cleanup. Knowledge bases nobody wants to maintain.

In two weeks, I used it to:

process 150 unstructured documents into capability maps across countries, job families, and org structures

refactor and maintain a functional AI agent knowledge base of 180 files from plain-English feedback

update PMO documentation and project statuses from meeting transcripts

prepare an AI Hackathon starter pack with synthetic data, instructions, flows, and participant-ready materials

work with SuccessFactors data while bypassing the transactional reporting interface

This is the part I care about. Proving that a data assistant can write SQL would be a small test.

I was testing whether a governed enterprise environment can absorb the annoying middle of business work while humans keep judgment, interpretation, and accountability.

For many use cases, yes.

The sharp HR example is SuccessFactors. Useful as a system of record. Painful as an analytics interface. If the data already sits in a governed lake, the reporting work can move closer to the place where analysis, documentation, and output generation happen.

That should make some SaaS vendors uncomfortable.

Also it has the Skills functionality I really love. A skill turns one good workflow into a reusable team method: definitions, rules, examples, output formats, and checks.

That is how you move from "one AI enthusiast did something clever" to "the function has a repeatable way to work."

My recommendation to HR leaders and non-technical AI enthusiasts:

Ask IT for a controlled Genie Code test, if you already have Microsoft's Databricks.

Use real data. Bring your definitions. Define one output. Build one reusable skill. Keep governance in the room.

The product is positioned as data work.

The wider application is company efficiency.

Personal context is the only durable asset in AI

i.nosov9@gmail.com (Ivan Nosov) — Thu, 30 Apr 2026 00:00:00 GMT

Consistent maintenance of your personal context is the only durable asset in the AI space. Everything else is a temporary tool subscription.

The rapid market shifts over the last few months proved this. Anthropic was the default assumption for serious work, but that window closed as competitors caught up and developers moved.

In April, OpenAI reported Codex growing from 3 million to 4 million weekly active developers in just two weeks, while open-source frameworks like Hermes Agent crossed 124,000 GitHub stars. The capability gap between models has shrunk to the point where adoption velocity and workflow integration matter more than fractional leads on technical benchmarks.

When vendor leadership rotates this fast, the underlying models become commoditized. What remains valuable is the structured knowledge layer you build around them - your project histories, decision logs, system rules, and explicit constraints. A clean and portable context layer turns vendor switching into a frictionless upgrade.

This portability allows you to fragment your tooling pragmatically based on the environment.

I moved my personal projects to Cursor to leverage its growing agentic infrastructure and the freedom to swap underlying LLMs depending on the task.

Simultaneously, I continue running Genie Code for enterprise work, where data governance and security architectures dictate the tool choice, and I am currently researching Hermes to map out how to best leverage its capabilities.

Operating across entirely different interfaces is seamless when you own the intelligence layer and treat the AI purely as a replaceable reasoning engine.

AI adoption is loud, but much of it still feels like extra work

i.nosov9@gmail.com (Ivan Nosov) — Wed, 22 Apr 2026 00:00:00 GMT

AI adoption is finally loud, but a lot of it still feels like extra work, not extra leverage, because most of us are paying a rework tax on generic output, then pretending that is the future of work.

The part nobody wants to name is that models execute instructions, and most of our operating logic still lives in people’s heads, not in writing.

I wrote a short, sharp take on the specification problem, starting from what we learned embedding an agent inside a real performance workflow, then moving through Whitney, Nonaka, and Taleb, with one point worth repeating.

If you want the durable advantage from AI, the binding constraint is not the model, it is whether you can specify how the organization actually runs, and keep that specification current while everything changes around it.

The Specification Problem: Why We Are Paying a 40% "Rework Tax" on AI

i.nosov9@gmail.com (Ivan Nosov) — Wed, 22 Apr 2026 00:00:00 GMT

We are putting 2025 tools into 2015 job structures. Here is the missing capability that separates the organizations capturing value from the ones just generating noise.

The defining reality of AI adoption right now is exhaustion.

According to 2026 workforce data, employees using AI report up to 32% higher productivity on individual tasks. But when you zoom out to the organizational level, those gains vanish.

The culprit is what recent research calls the "Rework Tax".

Approximately 40% of all AI-driven time savings are immediately lost to rework - correcting errors, verifying outputs, and patching up generic, context-free content. - HR Grapevine, Jan 2026

As a result, a massive value gap has opened: just 20% of companies are capturing 74% of AI's economic gains. Their differentiator? They redesign the work instead of just deploying the tool. - PwC AI Performance Study, Apr 2026

When a manager or HR Business Partner asks an AI to draft a performance calibration summary, the AI generates something plausible, professional, and entirely useless. The next hour goes into rewriting it to reflect the actual political realities, retention risks, and succession pipelines of the team.

The AI executed perfectly. The failure was ours: we asked a machine to execute a process without ever specifying the judgment required to do it.Organizations cannot tell AI what to do because they have never written down how they actually work.

The Historical Precedent

In 1798, Eli Whitney signed a contract to deliver 10,000 muskets to the U.S. government. Before Whitney, every musket was hand-crafted by a single gunsmith. Skilled labor was the bottleneck, and tacit knowledge - the "feel" of a good trigger - was irreplaceable. Manufacturing couldn't scale because nothing was explicitly defined.

Whitney's innovation was specification. He defined every dimension and tolerance precisely enough that any trained worker could produce a part, and any part could fit any musket. The bottleneck shifted from tacit skill to explicit standards. The American Manufacturing System was born.

AI faces the exact same constraint. Most organizational processes today are pre-Whitney muskets. When companies deploy AI to these processes, they are handing Whitney's machines to a craft shop. The machines work flawlessly. The shop fails, because nothing is specified.

The AI can run a pay equity regression, but it cannot apply the total rewards director's unwritten intuition about when a geographical premium justifies a variance. Organizations that specify their operating logic can plug in AI like interchangeable parts. Organizations that cannot remain unscalable craft shops.

The Knowledge Inversion

For thirty years, organizations optimized for knowledge creation. The SECI model (Nonaka & Takeuchi, 1995) explained how organizations turn tacit knowledge into explicit documents, which new humans then internalize to build their own tacit judgment over time. Six months of ramp-up for a new hire was just the cost of doing business.

AI breaks this cycle. AI executes explicit instructions literally. It never develops tacit understanding through exposure or corporate osmosis.

KEY INSIGHT: Organizations that optimized for knowledge creation - turning tacit knowledge into explicit forms for human learning - now must master knowledge extraction: surfacing tacit knowledge for machine execution. Different capability. Different discipline.

This creates the specification gap. The 20% of companies capturing the lion's share of AI value are the ones performing this extraction. The rest are paying the Rework Tax.

Where the Gap Actually Sits

The Capability in Practice

I got to this through implementation, not theory. During our shift from "ratings" to line of sight, we embedded an agent directly in the performance workflow and watched where it helped and where it failed.

The pattern was immediate. It worked when the operating context was explicit and current. It drifted when context was partial, stale, or translated differently across layers. In other words, the model quality mattered less than specification quality.

In practice, usefulness depended on whether people could trace their objectives from strategic themes to entity priorities to team contribution, then express individual goals without falling into copy-paste alignment. That is why the guidance had to encode one philosophy clearly: contribution, not cascade.

Adoption came fast because the need was real. Thousands of sessions happened in weeks, without a campaign, in a period when people were also managing changed rules and deadlines. The agent became useful only when the knowledge behind it was maintained as a live system, and not just a static FAQ.

The operational lesson: a reliable agent is a lagging indicator of organizational clarity. When operating logic is explicit and kept current, AI scales judgment. When it is not, AI scales ambiguity.

The Operational Advantage

Leaders who specify their organizations gain operational options previously impossible to execute.

Nassim Taleb's barbell strategy helps clarify why this matters. The 90% side protects downside through disciplined execution. The 10% side captures upside through bounded experimentation. AI expands both capacities, but only when the organization can specify what the core is doing and what an experiment is meant to test.

Without specification, the core remains dependent on tacit experts and the exploratory side degenerates into noise. With specification, the core becomes more scalable and the exploratory side becomes cumulative. Learning compounds because the rules, assumptions, and changes are explicit.

The Specification Audit

The audit for your organization is simple. Pick any judgment-heavy process - succession planning, employee relations escalations, executive comp design. Ask your best performer to explain how they make decisions. Try to write those rules down precisely enough that a machine with zero context could follow them.

If you can, you have found a process that can scale with AI. If you cannot, you have found your operational ceiling.

The career calculus for professionals shifts exactly the same way. Domain knowledge is the appreciating asset, but only when combined with the ability to specify it. The specialist who cannot articulate their judgment gets automated by "so-so automation" (Acemoglu, 2023). The hybrid leader - possessing deep expertise alongside specification capability - becomes indispensable.

The American Manufacturing System won because of specification, replacing the pride of unscalable craft with the economics of interchangeable scale. AI follows the identical pattern. The organizations that write themselves down are the organizations that will write the future.

Key References & Data Sources

HR Grapevine (January 2026). Corporate L&D is key to closing the AI productivity gap says report. (Source of the 40% "Rework Tax" and "2025 tools inside 2015 job structures" findings).
PwC (April 2026). 2026 AI Performance Study. (Source of the finding that 20% of companies capture 74% of AI's economic gains through workflow redesign).
Nonaka, I. & Takeuchi, H. (1995). The Knowledge-Creating Company. (SECI model of knowledge conversion).
Acemoglu, D. & Johnson, S. (2023). Power and Progress. (Critiques of "so-so automation").
Taleb, N.N. (2012). Antifragile: Things That Gain from Disorder. (Barbell strategy).
Images are generated with AI.

Celebrate

Support

Love

Insightful

Funny

Comment

Copy
LinkedIn
Facebook
X

2 Comments

Fabio Barraco

5mo

Report this comment

Spot on Ivan! The real bottleneck isn’t the AI’s capability, but our ability to frame the problem correctly. Outcomes improve dramatically when we can translate intent into well-structured prompts and break complexity down into clear, executable steps.

Reply 1 Reaction

2 Reactions

Anthropic releases Claude Opus 4.7

i.nosov9@gmail.com (Ivan Nosov) — Thu, 16 Apr 2026 00:00:00 GMT

Anthropic releases Claude Opus 4.7 today.

Three times the vision resolution of 4.6. More literal instruction following, which means existing prompts may need adjustment. Built-in cybersecurity detection, with a Cyber Verification Program for security professionals who need the full capability unlocked. And ultrareview in Claude Code, which identifies bugs and design issues rather than greenlighting whatever you wrote.

About time. Opus 4.6 was dropping IQ points faster than a GPT-4 benchmark the week after a new model drops.

Pricing holds: $5/M input, $25/M output. Available on API, Bedrock, Vertex AI, and Foundry from today.

If your agentic workflows have been bottlenecked on model reliability, this is the release to revisit.

Two Chinese AI labs as an experiment in organisational design

i.nosov9@gmail.com (Ivan Nosov) — Tue, 14 Apr 2026 00:00:00 GMT

Two Chinese AI labs are running a live experiment on organizational design.

Both eliminated middle management. Both rejected KPIs. Both have one layer between founder and individual contributor. Both outperform Western competitors on frontier benchmarks.

A Chinese journalist spent 100 hours inside Moonshot AI (the company behind Kimi). A separate investigation covered DeepSeek. What emerges is a pattern - not a culture quirk.

Moonshot: 300 employees, five co-founders each managing 40-50 people directly. No departments. No titles. 50% of employees changed roles completely since joining. They call it "two-dimensional folding" - collapsing organizational depth into a single plane.

DeepSeek: ~100 researchers, two levels total - Liang Wenfeng and everyone else. Employees leave at 6-7 PM. No timeclock. Liang's logic: "A person's high-quality work time rarely exceeds 6-8 hours daily." Meanwhile, OpenAI and xAI run 70-80 hour weeks.

Different approaches, same rejection of hierarchy. And the same enabler: AI handling what middle management used to handle.

Traditional hierarchy solves coordination problems. Who knows what. Who decides. Who talks to whom. Middle management is the routing layer - information up, decisions down, lateral sync through meetings. That layer has a cost: headcount, latency, context loss at every handoff.

These labs compressed the layer by pushing coordination into AI systems. Knowledge retrieval instead of "ask the senior person who remembers." Context synthesis instead of status meetings. Decision support that surfaces precedents and constraints instead of escalating to someone with pattern-matching experience. The routing function doesn't disappear - it moves from org chart to infrastructure.

Flat becomes viable at 300 people because the coordination tax dropped.

The obvious objection: elite research teams, not enterprise. But the principle scales even if the implementation doesn't. A 5,000-person company doesn't need to eliminate all hierarchy. It needs to ask which layers exist only for coordination that AI can now handle.

Only few have asked: what structure would we build if coordination were nearly free?

The Chinese labs asked.

Anthropic passes OpenAI in annualised revenue

i.nosov9@gmail.com (Ivan Nosov) — Tue, 07 Apr 2026 00:00:00 GMT

Anthropic just passed OpenAI in annualized revenue. $30B vs $25B ARR as of April 2026.

I've been a paying Claude subscriber since October 2024 - back when most people around me hadn't heard of Anthropic, and "AI" in the enterprise still meant Copilot procurement.

Eighteen months later, this chart exists. And the principle underneath is the one I consider #1 in every technology decision:

💎 Solve the need, don't sell the tool.

OpenAI had the brand, the distribution, the enterprise agreements, the Microsoft bundle. Everything a CIO needs to say yes. Anthropic had none of that. What this chart shows is how fast distribution advantages compress when the people doing the actual work find something that fits their workflow better. They switch, they pay out of pocket, and procurement catches up later.

This applies well beyond AI. Every technology vendor pitches integration, brand, ecosystem. The product that solves the specific daily friction wins anyway - sometimes slower, but it wins. I've seen the same pattern in HR tech, in payroll platforms, in analytics tools. The bundled option gets deployed, the better-fit option gets used.

A tough love experiment I ran recently to stress-test my own bias: moved my daily coding workflow to Cursor for a week now (after their latest release), and it lets you jump between models for different tasks. Some tasks worked better on one model, some on another. The useful discovery was that multi-model flexibility itself is the feature - picking the right model per task instead of committing to one vendor's ecosystem.

The most reliable signal in any tooling decision is where the heaviest users migrate on their own, unprompted, often paying out of pocket. More reliable than any vendor pitch or analyst quadrant.

The play most companies run with AI: hand it to IT

i.nosov9@gmail.com (Ivan Nosov) — Sat, 04 Apr 2026 00:00:00 GMT

Most companies I talk to are running the same play with AI - hand it to IT, set adoption KPIs, measure how many people opened Copilot this week. The results are predictably shallow: meeting transcriptions, advanced search, usage dashboards that measure nothing useful. Mollick calls this "de-weirding" and the framing is sharp.

Where I'd adjust is on IT as the place AI goes to die. The pattern is more specific: IT defaults to restricting what it can't fully control, functions default to waiting for someone else to figure it out, and together they produce a comfortable limbo where AI stays safely incremental. Infrastructure gets locked down before anyone knows what to build on it. Experimentation requests sit in backlogs. Nobody owns the product layer.

The Leadership-Crowd-Lab model in the article is the exact ambition I'm working to bring to life - direction from senior leadership, experimentation distributed across functions, a dedicated team connecting the two, with IT at the table as an infrastructure partner.

The structural problem is that neither side has been forced to operate this way before. AI is forcing the shift whether they're ready or not.

Distillation, licences and trust in the AI model ecosystem

i.nosov9@gmail.com (Ivan Nosov) — Sat, 21 Mar 2026 00:00:00 GMT

The AI model ecosystem right now.

Moonshot AI used 24,000 fake accounts and 3.4 million interactions to distill capabilities from Claude into their Kimi model.

Cursor, valued at $50 billion, quietly built Composer 2 on top of that same Kimi - while also selling Claude access to its users. When a developer exposed the model ID in the API, Moonshot's head of pretraining publicly called out the license violation.

Those posts were deleted hours later. By evening, everyone was congratulating each other.

Composer 1 was DeepSeek. Composer 2 is Kimi. The benchmark gaps between frontier models are narrowing quarter by quarter, while the food chain gets more circular - models training on models, everyone's output becoming everyone else's input.

This is why the model layer was never the strategic bet.

I use Claude today. It's the strongest tool I've found for the work I do. If Kimi or Gemini or something from a garage in Shenzhen pulls ahead next quarter, I'll switch without thinking twice. The knowledge I've built, the structured workflows, the infrastructure underneath... none of it is locked to a model. When your work is organized in formats any model can parse, switching costs drop to near zero.

The organizations paying the price right now are the ones who optimized for a specific tool's quirks instead of making their processes legible to any tool. Prompt libraries tuned to one model. Workflows that depend on a particular vendor's context window. Knowledge trapped in formats only one integration can read.

The model wars make for good memes. The position worth holding is the one where you don't care who wins them.

AI and the Solow paradox

i.nosov9@gmail.com (Ivan Nosov) — Wed, 18 Mar 2026 00:00:00 GMT

You can see the computer age everywhere but in the productivity statistics"

Robert Solow wrote that in 1987. Nobel laureate, one sentence, and it described the next decade of corporate IT perfectly. Companies bought computers, connected networks, installed software on every desk. Productivity didn't move.

Thirty-nine years later, same paradox, different technology.

The NBER just published one of the few AI adoption surveys fielded through central bank panels rather than paid internet samples - almost 6,000 CEOs and CFOs across the US, UK, Germany, and Australia. The findings:

69% of firms actively use AI

89% report zero productivity impact over the past three years

Executives themselves average 1.5 hours per week with AI. A quarter don't use it at all

The original Solow Paradox took about 10 years to resolve. Computers eventually showed up in the productivity data, but only after organizations redesigned workflows around the technology.

Paul David, a Stanford economic historian, documented the same pattern with electricity. Factories in the 1880s bolted electric dynamos onto steam-era layouts - same floor plan, same belt-and-pulley system, just a different motor. Modest fuel savings. Zero process change. Productivity came a generation later, when factories were physically rebuilt around individual electric motors, with single-floor layouts and production lines designed around material flow.

The pattern repeats because the mistake repeats. Most companies are at the dynamo-on-a-steam-floor stage with AI right now. Copilot sits inside Word. AI assistants summarize meetings that shouldn't have happened. Chatbots answer questions from knowledge bases nobody has updated since 2023.

Old workflows absorb new tools without changing outcomes.

Where I've seen AI actually shift output, the starting point was infrastructure - switching from .docx and .pptx to formats AI can parse, replacing static files with structured knowledge that stays current. Those are the early moves, clearing the ground.

The larger shift is what comes after. When information is structured, you can redesign how work flows - which steps disappear entirely, which move from teams to individuals, where a process that took a chain of five people compresses into one person with an AI operating on clean inputs. That's the factory floor being rebuilt, not just the motor being swapped.

Most AI budgets go to licenses. Almost none go to the sequencing that makes those licenses productive - first making work legible, then redesigning the workflows around what becomes possible. In most organizations, that's nobody's job description. The gap is visible in the NBER data, and probably recognizable in your own calendar, somewhere between the third AI demo and the first redesigned process.

Solow's paradox resolved when the factory floor was redesigned, not when better motors arrived. Same principle, different century.

When leaders build with AI themselves

i.nosov9@gmail.com (Ivan Nosov) — Mon, 16 Mar 2026 00:00:00 GMT

The AI adoption that actually stuck around me started with nobody's permission and no training program.

During an MBA assignment about a month ago, I ran a full research and full deliverable drafting exercise using AI while a General Manager from my cohort watched. Real deliverable, real time pressure. His first question was how to bring this into his organization. Within three weeks he was deep into testing what AI solves across his function, building his own workflows from scratch.

The same pattern repeated at work. And everywhere I go.

A Head of L&D saw what I'd built and started creating AI agents from scratch within days. A Rewards Regional Lead watched me run a pay equity analysis, then began doing her own through vibe coding - writing analytical logic in plain language, getting statistical output back. A Finance Director sat through one working session, then rebuilt how his team runs monthly reporting. A Commercial Head started automating market analysis workflows he'd been doing manually for years. A CHRO picked up a coding environment directly.

People whose closest relationship with code was an Excel formula, watching someone solve a real problem, then going back to their own desks and solving theirs. Each found their own use case, chose their own tools, built their own workflows... and none of them went back.

Prompt engineering didn't matter. Prompt engineering is school algebra - useful, learnable in an afternoon, absolutely ordinary. The actual barrier was seeing your own problem clearly enough to know what to demand from LLM. Domain knowledge carried them, while the tool was just the medium.

These are Managing Directors, CHROs, Heads of Functions responsible for budgets and teams across countries. When someone at that level builds with AI firsthand and brings the change into how their team operates, the organization gets structural permission to shift - how work is staffed, how budgets move, how decisions get made.

BCG surveyed 2,360 executives in early 2026 and the pattern is clear. The CEOs driving the highest AI returns - roughly 15% of the sample - spend over eight hours a week building with the tools personally. They allocate 60% of their AI budgets to workforce capability, versus 24% among those who delegate the learning.

The gap makes sense - a leader who has built something themselves knows where the value sits, knows what's real and what's vendor slides. They invest differently because they understand differently.

Most organizations start with training programs and license rollouts. What I've seen work starts with one demonstrated output that makes a leader see their own problem differently. The conviction comes from building, and the program follows the conviction.

What's your experience - have you seen leaders shift after building something themselves, or does it usually stay at the delegation level?

Microsoft Copilot Cowork at $99 a month

i.nosov9@gmail.com (Ivan Nosov) — Tue, 10 Mar 2026 00:00:00 GMT

Microsoft just launched Copilot Cowork bundled in at $99/month - AI product built on someone else's engine - then removed the three features that make it work.

Copilot Cowork runs on Anthropic's Claude. Same agentic architecture. No local file access. No project instructions. No code execution.

What remains is an M365 orchestrator. It reschedules your meetings, compiles PowerPoints from emails, assembles research memos from SharePoint. A genuine improvement over the old Copilot chat model - longer attention span, background execution, multi-step sequencing across apps.

The question is whether that improvement justifies a new product category and a new price tier.

Here's the thing about buying an engine and removing the drivetrain: you still have to push the car. Claude Cowork and Codex Desktop read your entire project folder, carry persistent instructions that define how the AI reasons about your domain, and run actual scripts in a sandbox. They compute. Copilot Cowork starts every session from a blank prompt and gives each user an isolated assistant with no shared context.

A finance controller can't point it at a cost model on their drive. A supply chain manager starts from scratch every session. An analyst can't run a regression or generate charts. Five people on the same team get five separate assistants with no common memory.

The fair counterpoint: Work IQ - the intelligence layer that reads your M365 graph - is a real capability neither Claude nor Codex have. Enterprise governance is real. For IT teams managing compliance at scale, these matter. But they matter to procurement, not to the person trying to get work done.

One tool runs a regression across four countries and produces 54 charts. The other reschedules your Tuesday standup. Both are called "Cowork."

Run the spec comparison yourself before signing E7. Pilot alternatives for your power-user cohort. Negotiate from the gap, not from the pitch deck.

Carousel below for the full breakdown.

Software engineering hiring and the Jevons paradox

i.nosov9@gmail.com (Ivan Nosov) — Mon, 09 Mar 2026 00:00:00 GMT

Software engineer job postings are up 11% year-over-year in the most AI-automated profession on the planet - a textbook case of Jevons Paradox, where making a resource cheaper to use increases total demand instead of reducing it.

Overall job postings are flat. The field with Claude Code, Codex, Cursor, and dozens of specialized coding agents... is the outlier hiring more.

Citadel Securities published a macro strategy report last month that frames why this holds. Recursive technology capability does not mean recursive adoption. Technological diffusion follows an S-curve - slow and expensive early adoption, then acceleration, then saturation. AI is still in the early phase. St Louis Fed real-time population data shows daily AI use at work is flat. No inflection.

The mechanism is specific. When coding gets 3x faster, companies don't cut engineers by two-thirds - they greenlight three times the projects. Internal tools that were never worth building suddenly have a business case.

Prototypes that took a quarter take a week. The backlog doesn't shrink, it explodes, because the organization sees what's possible at the new speed. Resistance goes down, ambition goes up, demand follows. Keynes predicted productivity gains would give us a 15-hour workweek. Societies consumed more instead.

There's also a hard economic ceiling most workforce planners ignore. Compute has a marginal cost. When the cost of automating a task exceeds the cost of a human doing it, substitution stops. That boundary exists, and we're nowhere near crossing it for most knowledge work.

This matters for anyone building a workforce strategy. The displacement question - "which roles disappear" - is the wrong frame. Roles change shape, output expectations rise, and the people who can work effectively with AI tools become more valuable.

The strategic move is capability building at speed. Finance controllers running 10x the scenario analyses. Commercial teams modeling pricing permutations they used to send to consultants. Operations managers automating reporting they used to spend Fridays assembling.

The data says complement. Build for that.

Source: Citadel Securities, "The 2026 Global Intelligence Crisis," Frank Flight, Feb 2026. Indeed Job Postings, Software Engineers vs Overall, daily + 21-day moving average.

AI Capability Framework - The Four Engineering Disciplines

i.nosov9@gmail.com (Ivan Nosov) — Tue, 03 Mar 2026 00:00:00 GMT

Most organizations treat AI as a tool problem. They buy platforms, run pilots, train people on prompting, and wait for results. Then they wonder why 60% of AI investments generate no material business value (BCG, "The Widening AI Value Gap," 2025).

The technology works. What’s missing are the engineering disciplines that make AI produce reliable, context-specific, strategically aligned outcomes. These are institutional capabilities - built across functions, embedded in how work gets specified and executed, maintained as the technology shifts underneath.

This document defines four engineering disciplines that form a progressive capability stack. Each builds on the previous. Together, they represent the organizational infrastructure required to extract durable value from AI - not as a one-time implementation, but as a permanent operating capability. The disciplines originate in AI-native companies, but the underlying work - structuring knowledge, defining objectives, codifying processes - is the same whether you sell software, spirits, or steel. Every organization that runs on institutional knowledge and human judgment faces the same gap.

────────────────────────────────────────────────────────────

The Capability Gap No One Is Naming

The discourse around AI capability has been dominated by two narratives, both incomplete.

Narrative 1: "Just adopt it." Get tools into people’s hands, train them to prompt, measure adoption rates. This produces the BCG finding that 88% of organizations use AI in at least one function, but only 5% create substantial value at scale. Adoption without capability is consumption without digestion.

Narrative 2: "Upskill for the future." The World Economic Forum’s Future of Jobs 2025 report identifies AI and big data as the most important skill, with 77% of surveyed organizations committed to reskilling. But "AI skills" as a category is uselessly broad. It conflates typing a ChatGPT prompt with architecting an agentic system - like calling both "driving a car" and "designing a powertrain" the category "automotive skills."

The organizations actually generating value tell a different story. At Shopify, CEO Tobi Lutke declared "reflexive AI usage is now a baseline expectation" - but the real operational change was requiring managers to demonstrate why a task could not be done by AI before requesting additional headcount (Lutke, April 2025). At Microsoft, Satya Nadella introduced weekly "AI accelerator" meetings that deliberately exclude senior executives, placing engineers and technical staff at the center - designed to surface the unfiltered operational reality of how AI actually works inside the organization (Microsoft, 2025). At Anthropic, engineering productivity increased 200% per engineer after building systematic AI workflows: shared specification files, multi-agent orchestration, structured context management (Cherny, 2025). These are capability architecture stories, not adoption stories. And while these examples come from technology companies - the first to formalize the practices because they build AI systems daily - the disciplines themselves describe universal organizational work. Structuring institutional knowledge, specifying how decisions get made, defining what success looks like for autonomous processes: these are challenges for any company with more than a few dozen employees, a few legacy systems, and a few decades of accumulated practice.

What’s missing is a structural framework that defines what, specifically, organizations must learn to do with AI - and in what order. The answer is foundational engineering disciplines that compound over time and create institutional capability others cannot easily replicate. Generic "AI literacy" and tool-specific training that expires with every model update do neither.

The four disciplines are:

Prompt Engineering - communicating with AI systems (2024 frontier, now commoditizing)
Context Engineering - assembling the right information for AI to produce correct results
Intent Engineering - defining objectives, constraints, and success criteria for autonomous AI
Specification Engineering - codifying organizational processes and logic into machine-executable standards

Each has an emerging body of practice, a distinct skill profile, and a clear relationship to the others. Most organizations have invested heavily in layer one while ignoring layers two through four entirely.

────────────────────────────────────────────────────────────

Layer 1: Prompt Engineering

What it is. The discipline of communicating effectively with language models through structured instructions, examples, and reasoning patterns.

Why it matters. Prompt engineering established a core insight: the quality of AI output is determined by the quality of human input. This is not obvious. Most technology before AI improved output independently of how the user communicated with it - a spreadsheet formula works the same whether you understand it or not. AI systems are different. They are input-sensitive in ways that reward human skill.

Why it is commoditizing. By 2025, the primitive era of prompt engineering - characterized by trial-and-error iteration and artisanal prompt crafting - was already fading. Techniques like few-shot learning and chain-of-thought reasoning are now, as one practitioner put it, "as fundamental as knowing SQL is to database management - table stakes, not competitive advantages" (Big Blue Academy, 2026). Models have become better at interpreting unstructured input. Interfaces have absorbed prompting patterns into templates and defaults. The Fortune headline in May 2025 was blunt: the six-figure prompt engineering role was "already obsolete."

What this means for organizations. Prompt engineering should be treated as foundational literacy - the equivalent of email etiquette or spreadsheet competence. Everyone should have it. No one should specialize in it alone. It belongs in onboarding, not in a center of excellence. Organizations that are still investing primarily in "prompt engineering training" in 2026 are building a capability that is actively depreciating.

The residual value. Prompt engineering does not disappear. It becomes absorbed into the higher disciplines. A context engineer must write effective system prompts. An intent engineer must articulate objectives in language models can parse. A specification engineer must encode process logic in structured natural language. Prompting is to the AI capability stack what grammar is to writing: necessary, invisible when done well, insufficient on its own.

────────────────────────────────────────────────────────────

Layer 2: Context Engineering

What it is. The discipline of assembling, structuring, and delivering the right information to an AI system so it can produce accurate, relevant, situation-specific outputs.

Where the term comes from. In June 2025, Andrej Karpathy - one of the founding researchers at OpenAI and former head of AI at Tesla - named the shift from "prompt engineering" to "context engineering," defining it as "the delicate art and science of filling the context window with just the right information for the next step." The term was adopted rapidly: Anthropic, LangChain, and independent practitioners each published formal frameworks within weeks (see References 2, 3, 13). By late 2025, it had its own job descriptions.

The conceptual shift. The move from prompt engineering to context engineering reflects a structural change in how AI systems work. Modern AI applications do not run on single prompts. They operate on assembled context: system instructions, retrieved documents, structured data, conversation history, tool definitions, and prior model outputs - all packed into a context window that functions as the model’s working memory. Karpathy’s analogy: the LLM is the CPU; the context window is the RAM. What you load into RAM determines what the CPU can compute.

What it involves in practice.

Knowledge retrieval (RAG). Connecting AI to organizational knowledge bases, documents, policies, and data sources so responses are grounded in actual institutional information rather than general training data.
Information architecture. Structuring organizational knowledge so it can be efficiently retrieved and assembled. This includes taxonomies, metadata, document standards, and knowledge graph design.
System prompt design. Defining the persistent instructions, constraints, and behavioral parameters that shape how the AI operates within a specific organizational context.
Memory and state management. Maintaining relevant information across multi-turn interactions so the AI retains prior context rather than starting fresh each time.
Tool and API integration. Defining what external systems the AI can access and how, expanding its capability beyond text generation into action.

Why most organizations fail here. Context engineering exposes a truth that technology purchases cannot solve: most organizational knowledge is not structured for machine consumption. It lives in email threads, in the heads of experienced employees, in political agreements that were never documented, in process exceptions known only by the people who handle them. AI systems have a hard constraint here: they can only work with the information loaded into their active memory (the "context window"), and research from Stanford and UC Berkeley has shown that performance degrades when too much information is loaded at once - critical details get buried in the middle and missed. More information is not better. The right information, well-structured, is better. The binding constraint is the organization’s ability to surface the right knowledge in the right structure at the right moment.

How the frontier organizations solve it. Boris Cherny, head of Claude Code at Anthropic, described the solution his team uses: "When Claude does something wrong, add it [to the specification] so it doesn't repeat." His team maintains a shared CLAUDE.md file checked into version control - institutional memory for AI systems, continuously refined through use. The principle is simple: every error is a missing specification. Every missing specification is a knowledge gap made visible.

The organizational capability required. Context engineering is a cross-functional discipline. It requires:

Knowledge architects who can structure organizational information for AI retrieval
Data stewards who maintain the quality and currency of knowledge bases
Integration specialists who connect AI systems to data sources and tools
Domain experts who validate that retrieved context is accurate and complete

At this layer, AI capability building is an organizational design challenge, not a training program.

────────────────────────────────────────────────────────────

Layer 3: Intent Engineering

What it is. The discipline of defining objectives, desired outcomes, constraints, decision boundaries, and success criteria for AI systems that operate with increasing autonomy.

Why it exists. As AI systems evolve from conversational assistants to autonomous agents - systems that plan, execute multi-step tasks, use tools, and make decisions - a new problem emerges: these agents fail not because they cannot reason, but because "their objectives, outcomes, and constraints are underspecified" (Product Compass, 2025). An agent that can execute perfectly will produce useless or harmful results if it does not know what success looks like, what trade-offs are acceptable, and when to stop.

The defining insight. "Intent is what determines how an agent acts when instructions run out" (Product Compass, 2025). Prompts tell an AI what to say. Context tells it what to know. Intent tells it what to accomplish - and what not to sacrifice in the process.

The framework components. An intent specification, as formalized by emerging practice, includes:

Objective. The problem being solved and why it matters. Qualitative, aspirational, guides trade-offs.
Desired outcomes. Observable states indicating success. Not activities or outputs - observable changes from the user’s perspective.
Health metrics. What must not degrade during optimization. This addresses Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure.
Strategic context. The operational system the AI is working within - organizational priorities, dependencies, constraints.
Decision types and autonomy levels. Which decisions the agent takes independently, which require human escalation.
Constraints. Both soft constraints (steering guidance) and hard constraints (enforced boundaries).
Stop rules. When to halt, when to escalate, when to declare completion.

The progression from context engineering. Context engineering improves the relevance of AI outputs but remains reactive to user queries. Intent engineering transforms the system "from a conversational interface into a structured execution engine" (Tericsoft, 2025). The shift is from "give the AI the right information" to "give the AI the right objective."

The organizational challenge. Intent engineering requires something most organizations have never done: explicitly articulate their actual objectives, actual priorities, and actual trade-off logic. The strategy deck version will not do. What’s needed is the real version - the one that exists in the heads of senior leaders and manifests in how decisions actually get made.

Guillermo Rauch, CEO of Vercel (the company behind Next.js and v0), distilled this to three essential capabilities for the AI era: "clearly defining product intent, coaching AI effectively through iteration, and quickly resolving challenges when AI gets stuck" (Rauch, 2025). The first of those three - defining intent - is the hard one. The rest follows from it.

This is uncomfortable. Organizations run on ambiguity by design. Competing priorities coexist because no one has been forced to choose between them. Intent engineering forces the choice. You cannot tell an agent to "maximize customer satisfaction while minimizing cost while maintaining quality while moving fast" and expect coherent behavior. Agents, unlike people, cannot navigate political ambiguity. They need rank-ordered objectives, explicit trade-offs, and clear escalation rules.

What this means for capability building. Intent engineering is not a technical skill alone. It requires:

Strategic clarity from leadership about what actually matters and in what order
Process owners who can articulate success criteria and failure modes for their domains
Governance specialists who can define autonomy boundaries and escalation logic
Cross-functional alignment because an agent operating across functions will expose every inconsistency in how those functions define success

────────────────────────────────────────────────────────────

Layer 4: Specification Engineering

What it is. The discipline of codifying organizational processes, decision logic, policies, and standards into structured, machine-readable specifications that AI systems can execute against reliably and consistently.

The ambition. If intent engineering tells AI what to accomplish, specification engineering tells it how the organization works. Every process. Every decision tree. Every policy. Every exception. Every escalation path. Every unwritten rule. Captured in a format precise enough for machines to execute and clear enough for humans to audit.

Why this is the hardest layer. Most organizations cannot specify their own operations with precision. Not because they choose not to, but because much of how organizations actually function has never been written down:

Tribal knowledge. The way things actually get done, passed from person to person, never documented. When key employees leave, this knowledge disappears. Research on tacit knowledge in organizations found that "institutional knowledge is often locked in the minds of subject matter experts" and cannot be reliably reconstructed from formal documentation alone (Zuin, Mastelini, Loures & Veloso, 2025).
Political agreements. Informal understandings about territory, authority, and precedence that shape how decisions actually get made. These are undocumented by design because writing them down would surface contradictions and require resolution.
Process exceptions. The "but in this case we do it differently" knowledge that accumulates in experienced teams. Often more voluminous than the official process itself.
Implicit standards. Quality thresholds, communication norms, approval dynamics, risk tolerance - all operating without formal specification, calibrated through years of organizational socialization.

The connection to spec-driven development. In software engineering, a parallel discipline has already matured. ThoughtWorks defines spec-driven development (SDD) as "a development paradigm that uses well-crafted software requirement specifications as prompts, aided by AI coding agents, to generate executable code" (ThoughtWorks, 2025). GitHub open-sourced an SDD toolkit in September 2025. Addy Osmani, engineering lead at Google Chrome, published "How to Write a Good Spec for AI Agents" (O'Reilly, January 2026), establishing that specifications should focus on "what and why, more than the nitty-gritty how" - the organizational equivalent of defining success criteria before dictating procedures.

How this differs from traditional requirements engineering. IT teams will recognize specification engineering as familiar territory - requirements gathering, business process modeling, and systems analysis have existed for decades. The resemblance is real, but the consumer is different. Traditional requirements specifications are read by human developers who fill gaps with judgment, ask clarifying questions, and infer intent from ambiguity. AI agent specifications must be complete because the machine will execute them literally. A human developer reads "handle edge cases appropriately" and makes reasonable choices. An AI agent treats it as a null instruction. The discipline is the same. The tolerance for ambiguity is zero.

The AGENTS.md standard - an open format for guiding AI coding agents - has been adopted by over 20,000 repositories and is supported natively by GitHub Copilot, OpenAI Codex, Google Gemini, Cursor, and others (InfoQ, August 2025). Each repository’s AGENTS.md encodes its build steps, testing workflows, coding conventions, and quality standards in a structured format that any AI agent can read and execute against. This is specification engineering applied to software development - and the principle transfers directly to organizational operations.

The question for any enterprise function is the same one software teams have answered: can you define the inputs, outputs, preconditions, postconditions, and constraints of your compensation review process? Your onboarding workflow? Your talent calibration? Your market entry checklist? Your supply chain escalation logic? If not, no AI system will execute them reliably. The complexity of a consumer goods company’s route-to-market process or a financial institution’s regulatory compliance workflow is no less specification-worthy than a software build pipeline. The tooling is the same. The intellectual challenge is the same. The only difference is that software teams started first.

What specification engineering produces. When an organization develops this capability, it creates a new class of artifacts:

Process specifications. Machine-readable definitions of how work flows through the organization, including decision points, routing logic, and exception handling.
Policy specifications. Structured representations of organizational rules that AI can apply consistently across cases, languages, and geographies.
Quality specifications. Explicit criteria for what "good" looks like in each context - replacing the implicit judgment of experienced employees with auditable standards.
Integration specifications. Definitions of how systems, processes, and data sources connect and depend on each other.

The organizational consequence. Specification engineering forces organizations to make explicit what has always been implicit. It surfaces contradictions, resolves ambiguities, and creates a written record of how the organization actually operates - often for the first time. The resulting clarity has independent value regardless of technology. An organization that has specified its operations can deploy AI against them, but it has also built something more durable: an auditable, transferable description of how it works.

────────────────────────────────────────────────────────────

Where Knowledge Falls

Knowledge is the raw material that each discipline progressively structures and makes actionable. The four engineering disciplines are, at their core, a framework for transforming organizational knowledge from tacit to executable:

This progression maps to the knowledge management literature’s SECI model (Nonaka & Takeuchi): Socialization (tacit to tacit), Externalization (tacit to explicit), Combination (explicit to explicit), and Internalization (explicit to tacit). The four engineering disciplines operationalize this cycle for the AI era, with one critical addition: the final transformation is not back to tacit knowledge in human heads, but to codified, executable knowledge in organizational systems.

The implication. An organization’s AI capability is ultimately bounded by its ability to codify its own knowledge. No model, no platform, no vendor can compensate for knowledge that does not exist in structured, accessible form. This is why BCG found that organizations generating value from AI are those that "redesign workflows" - they are doing the knowledge codification work, whether they call it that or not.

────────────────────────────────────────────────────────────

The Capability Stack: How the Layers Relate

The four disciplines are not independent skills to be trained in parallel. They are a stack. Each layer requires the previous one as foundation and enables the next one to function:

Why the order matters. You cannot do context engineering if your people cannot prompt effectively - they will not know what information to assemble or how to validate that it was assembled correctly. You cannot do intent engineering if your knowledge is not structured - you will define objectives against information that does not exist in accessible form. You cannot do specification engineering if your intents are ambiguous - you will codify contradictions rather than clarity.

Where most organizations are today. The vast majority of organizations are investing almost exclusively in Layer 1. "AI literacy programs" are prompt engineering programs. "AI upskilling" means teaching people to use ChatGPT more effectively. This is necessary but radically insufficient - it is teaching people to type while the real capability gap is in what they should be writing.

The stack in practice: performance management

To make the progression concrete, consider a process every leader in every organization runs: performance management. Goal-setting, feedback, calibration, development decisions. Every manager does it. Most do it badly. AI makes the gap between good and bad visible.

Layer 1 - Prompt. A manager asks an AI to "write a performance review for a team member who exceeded targets." The output is generic, plausible, and useless - it sounds professional, says nothing about actual work done. Could describe anyone in any company.

Layer 2 - Context. The AI is now connected to the goals set in January, project deliverables, peer feedback, skill assessments, and team benchmarks. The same question produces a review that references real work. This is where most organizations stall - because their goals live in one system, project data in another, feedback in a third, and nobody built the connective tissue between them.

Layer 3 - Intent. The system now calibrates against defined objectives: differentiate the top 15% with development investment, flag rating inflation across managers, ensure succession pipeline coverage for critical roles, escalate misaligned ratings before calibration meetings. The AI is no longer drafting text - it is running the review cycle against explicit success criteria, surfacing patterns humans miss, and routing decisions to the right authority level.

Layer 4 - Specification. The full cycle is codified: goal-setting templates tied to business OKRs, calibration rules by grade level, promotion criteria with evidence requirements, development plan triggers based on skill gaps and career aspirations. The AI executes the entire performance management cycle - across 25 markets, against local employment regulations - and produces an auditable output managers review and decide on, rather than draft and format from scratch.

The difference between Layer 1 and Layer 4 is the difference between an AI that helps one manager write one review and an AI that runs a global process to completion. The technology at each layer is the same. What changes is the organizational clarity required to use it.

────────────────────────────────────────────────────────────

Emerging Artifact Domains

As organizations build these capabilities, new functional domains emerge that do not map to traditional organizational structures. These are not departments - they are disciplines that must be distributed across functions:

Knowledge Architecture

Structuring organizational knowledge for both human and machine consumption. Combines traditional knowledge management with information architecture, ontology design, and retrieval system optimization. Requires deep domain expertise combined with structural thinking - this is not an IT function.

Process Specification

Defining how organizational processes work with enough precision for AI execution. Process owners move from describing "what we do" in narrative form to specifying inputs, outputs, decision logic, exceptions, and quality criteria in structured formats. This is where unspoken institutional knowledge gets transferred into documented, auditable form.

AI Governance and Intent Alignment

Maintaining the objectives, constraints, autonomy boundaries, and escalation logic for AI systems. Sits between strategy and technical implementation - requires both strategic clarity and system-level understanding.

Evaluation and Quality Engineering

Measuring whether AI systems produce correct, consistent, and useful outputs. Requires building evaluation criteria and test cases from organizational reality (not synthetic benchmarks), and maintaining quality standards as both AI systems and business processes evolve.

Integration and Orchestration

Connecting AI systems to organizational data, tools, and workflows. Includes API design, data pipeline management, and coordination of multi-agent systems that span functional boundaries.

These domains represent new forms of organizational capability that do not exist today in most enterprises. They will not emerge from training programs alone. They require deliberate organizational design, role definition, and sustained investment - the same treatment organizations gave to digital transformation capabilities a decade ago, but with a more precise target.

────────────────────────────────────────────────────────────

Implications for Capability Building

What this framework changes about AI strategy

Investment sequencing. Organizations should invest in the four disciplines in order, ensuring each foundation is solid before building the next layer. Running an intent engineering initiative without functioning context engineering is premature optimization.
Training is necessary but insufficient. Prompt engineering can be taught in a workshop. Context engineering requires organizational infrastructure. Intent engineering requires strategic clarity. Specification engineering requires institutional commitment to making implicit knowledge explicit. The difficulty and investment required escalates at each level - and so does the value created.
The technology is secondary to the discipline. Models change every quarter. Platforms consolidate and fragment. But the organizational capability to structure knowledge, define intent, and specify processes persists across all technology shifts. An organization that has codified its compensation logic into a clear specification can deploy it on any AI platform. An organization that has not done this work will fail on every platform equally.
The durable work is the specification work. The tasks that persist are the ones AI cannot do for itself: structuring knowledge, defining objectives, specifying processes, and auditing outputs. Everything else is being absorbed by the technology at an accelerating rate. The people who do this work are the ones whose roles grow in value.
Industry is irrelevant to the framework. The four disciplines describe how any organization prepares its institutional knowledge for an era of AI-augmented execution. A technology company, a beverage company, a bank, and a hospital all share the same dependency: processes that run on accumulated human judgment, documented nowhere with enough precision for a machine to replicate. The capability gap is the same. The path through it is the same. The only variable is where each organization starts.

────────────────────────────────────────────────────────────

A Note on Pace

These disciplines are not theoretical. They are already being practiced at the frontier, even if the language varies:

Anthropic runs 70-100% of its engineering output through AI agents orchestrated with specification files and sub-agent architectures. Boris Cherny’s workflow: specification first, plan second, execution third. "Coding becomes a pipeline of phases: spec, draft, simplify, verify" (Cherny, 2025).
OpenAI shipped the Agents SDK with explicit support for intent engineering primitives: tool definitions, handoff protocols, guardrails, and tracing. The Model Spec (December 2025) codifies how agents should interpret autonomy boundaries and trustworthiness signals.
Mistral AI (Paris) is building enterprise AI infrastructure with a different premise: AI as utility, not spectacle. Arthur Mensch, CEO, describes the goal as "abstracting away from lower-level routines that data scientists understand to higher-level systems that business owners can actually use" (Mensch, 2025). Their enterprise deployments focus on three use cases that map directly to the capability stack: developer productivity (context engineering), knowledge management connected to proprietary data (specification engineering), and customer service automation (intent engineering). EU-hosted, GDPR-native infrastructure.
AWS published prescriptive guidance on "clarifying agent intent and scope" as the first focus area for operationalizing agentic AI - positioning intent engineering as the prerequisite for production deployment (AWS, 2025).

The timeline of this shift is worth registering. In 2024, prompt engineering was the leading edge. By mid-2025, Karpathy and the AI engineering community had already moved to context engineering. By late 2025, intent engineering frameworks were being formalized. Specification engineering is the current frontier - and it is being standardized faster than previous disciplines because the tooling now demands it. AI agents that operate without specifications produce unpredictable results. The specification functions as the operating system of the agent.

The tech companies listed above are building these practices in public because open documentation is their culture. But the need is identical in industries where the work is less visible. A consumer goods company with 50 markets, each running localized versions of global processes, has a specification engineering challenge that dwarfs most software shops. A financial services firm with 30 years of compliance precedent embedded in institutional memory faces a context engineering problem of extraordinary depth. A manufacturing company with plant-level process variations across three continents needs intent alignment as badly as any AI lab. The disciplines are universal. The industries that build them first gain the same compounding advantage - regardless of what they produce.

Organizations that wait for these disciplines to become packaged training programs will find themselves permanently behind. The capability is built through practice - through the daily work of specifying, structuring, and iterating on how AI operates within a specific organizational context.

────────────────────────────────────────────────────────────

References

Practitioners and AI Labs

Karpathy, A. Post on context engineering. X, June 2025. "Context engineering is the delicate art and science of filling the context window with just the right information for the next step."
Willison, S. "Context Engineering." simonwillison.net, June 2025.
Anthropic. "Effective Context Engineering for AI Agents." anthropic.com/engineering, 2025.
Cherny, B. (Head of Claude Code, Anthropic). Interviews on AI workflow, specification files, and multi-agent orchestration. 2025. Via Lenny’s Newsletter, VentureBeat, DEV Community.
Lutke, T. (CEO, Shopify). Internal memo on AI as baseline expectation. April 2025. Published on X.
Nadella, S. (CEO, Microsoft). AI accelerator meetings, organizational restructuring for AI capability. Microsoft Blog, 2025.
Mensch, A. (CEO, Mistral AI). Interviews on enterprise AI as utility, knowledge management, and EU-hosted infrastructure. 2025-2026. Via CNBC, Big Technology Podcast, McKinsey.
Rauch, G. (CEO, Vercel). On defining product intent, AI-era developer skills, and spec-driven workflows. 2025. Via Lenny’s Newsletter, ChatPRD.
OpenAI. "Model Spec." model-spec.openai.com, December 2025.
Osmani, A. (Google Chrome). "How to Write a Good Spec for AI Agents." O'Reilly / addyosmani.com, January 2026.

Standards and Frameworks

AGENTS.md. "A simple, open format for guiding coding agents." agents.md. Adopted by 20,000+ repositories.
GitHub. "Spec-driven development with AI: Get started with a new open source toolkit." GitHub Blog, September 2025.
LangChain. "Context Engineering for Agents." blog.langchain.com, 2025.
ThoughtWorks. "Spec-driven development: Unpacking one of 2025's key new AI-assisted engineering practices." 2025.
AWS. "Focus Area 1: Clarify Agent Intent and Scope." AWS Prescriptive Guidance, 2025.
ISO/IEC 42001:2023. AI Management Systems.

Intent Engineering

Product Compass. "The Intent Engineering Framework for AI Agents." 2025.
Tericsoft. "Intent Engineering in AI: The Shift Beyond Context Engineering." 2025.

Knowledge and Research

Zuin, Mastelini, Loures & Veloso. "LLMs as Tacit Knowledge Excavators in Organizations." 2025.
MDPI. "Unveiling the Unspoken: A Conceptual Framework for AI-Enabled Tacit Knowledge Co-Evolution." 2025.
Nonaka, I. & Takeuchi, H. "The Knowledge-Creating Company." Oxford University Press, 1995.

Industry Analysis

BCG. "The Widening AI Value Gap: Build for the Future." September 2025.
McKinsey & Company. "The State of AI in 2025." 2025.
World Economic Forum. "The Future of Jobs Report 2025." January 2025.
Fortune. "This six-figure role was predicted to be the next big thing - it's already obsolete thanks to AI." May 2025.
InfoQ. "AGENTS.md Emerges as Open Standard for AI Coding Agents." August 2025.

Celebrate

Support

Love

Insightful

Funny

Comment

Copy
LinkedIn
Facebook
X

3 Comments

Matthew Diakonov

7mo

Report this comment

the four-layer capability stack framing is really useful. the commoditization of prompt engineering at layer 1 matches what Im seeing - the value isnt in crafting better prompts anymore, its in encoding organizational knowledge with enough precision that the agent can execute autonomously. the performance management example makes it concrete. Im seeing the same pattern in software development - at layer 1 the agent writes generic code, at layer 4 it understands your specific architecture, testing conventions, deployment constraints, and can operate across the full development lifecycle with minimal human review

Reply 1 Reaction

2 Reactions

Most AI training programmes teach people to type

i.nosov9@gmail.com (Ivan Nosov) — Tue, 03 Mar 2026 00:00:00 GMT

Most AI training programs are teaching people to type.

The capability gap is in what they should be writing.

Every organization I talk to is running prompt engineering workshops. Some have moved to RAG pilots. Almost none have asked the harder question: can you actually specify how your own processes work with enough precision for a machine to execute them?

There's a four-layer capability stack emerging from AI-native companies. Prompt engineering is the bottom layer. It's commoditizing. The three disciplines above it are where value compounds - and where almost nobody is investing yet.

Take performance management. At Layer 1, an AI drafts a generic review that could describe anyone in any company. At Layer 4, it runs the full cycle - goal-setting, calibration, development triggers - across markets and local regulations, producing auditable output managers review rather than write from scratch.

Same technology at every layer. What changes is the organizational clarity required to use it.

I wrote the full framework - what the four disciplines are, why the sequence matters, and what each one actually takes to build 👇

Anthropic’s architecture for running HR operations through an LLM

i.nosov9@gmail.com (Ivan Nosov) — Thu, 26 Feb 2026 00:00:00 GMT

Anthropic just shipped a clean architecture for running HR operations through an LLM.

Six commands, six skills, fully open-source. Worth understanding how it's built.

Just released Claude Cowork HR plugin is the great generalized example of someone leveraging LLM power through separating what the model knows (skills) from what it does (commands) from where the data lives (connectors). Three layers, each are just a readable markdown files.

Quick context if you missed Cowork - it's a tab inside the Claude desktop app, built on the same agent SDK as Claude Code but for knowledge workers. You give it access to folders, describe what you need, it plans and executes. Plugins add domain structure on top.

The architecture is more interesting than the feature list. Every command follows the same pattern - take structured inputs, apply a framework, produce a formatted output. That pattern maps perfectly onto the operational backbone of HR. Not the strategic decisions, but the repeatable work that follows rules and consumes time.

The whole thing is open-source. Every command and skill is a markdown file you can read, edit, or rewrite. If your company has a specific performance review methodology or comp philosophy, you change the file.

Walked through the full structure below - what's inside, how the layers connect, and where this fits vs. where it doesn't.

Years of Excel skill, quietly made optional

i.nosov9@gmail.com (Ivan Nosov) — Sun, 22 Feb 2026 00:00:00 GMT

Spent years getting really good at Excel. Like, genuinely good - it was one of the things that set me apart early in career.

Now I do the same analytical work through Claude in minutes ☕

The skill just... quietly stopped mattering. No one pitched me on it, no vendor demo - I just stopped opening the app.

Anyone else feeling a specific skill from earlier career becoming optional?

Not from abundant AI marketing - from actually using it daily?

HBR: AI does not reduce work, it intensifies it

i.nosov9@gmail.com (Ivan Nosov) — Sun, 15 Feb 2026 00:00:00 GMT

HBR just published a warning: "AI Doesn't Reduce Work — It Intensifies It"

I read it as a confirmation that it's working.

UC Berkeley tracked a 200-person tech company for eight months. Product managers started writing code. Researchers picked up engineering tasks. Role boundaries dissolved. People worked more, not less. By month six - burnout, anxiety, decision fatigue.

Yes, AI intensifies work. That's the whole point. The question is whether you're intensifying in the right direction - compressing process to expand what you're capable of, or spinning the same wheel at higher RPM.

Last Friday I sat down at 9 AM to prepare two pitch decks and pull together research for a third workstream. By noon I had three distinct products ready to push forward. A year ago, that's a full week. The mechanical work - compressed into hours.

What opened was the cognitive bandwidth. And cognitive bandwidth, unlike calendar time, compounds when you reinvest it.

That's the shift the Berkeley researchers documented but framed as a problem. People expanded into unfamiliar territory. They crossed role boundaries. They took on work that wasn't "theirs". The study treats this as erosion. I think it's closer to what happens when you remove artificial constraints from capable people and they find out what they can actually do.

Now apply this to a traditional organization. Not a tech startup - a multinational with established functions, reporting lines, job families built over decades. The entire architecture assumes one person, one specialty, sequential work. You stay inside your domain. You escalate to specialists when you hit the boundary of your role.

That architecture was built for a world where crossing boundaries was expensive. AI made it cheap. A finance controller building scenario models without a data analytics team. A supply chain manager running demand forecasting without a planning specialist. A commercial lead pulling together competitive intelligence and market sizing in an afternoon. A talent manager doing rewards benchmarking and HRBP-level workforce planning because the processing barrier disappeared.

People discovering their expertise was always broader than their job description allowed.

The organizational design question is now: do you build for this, or do you fight it?

The teams that restructure around expanding capability will compound. Every person brings a wider surface area to every problem. Output doesn't grow linearly. It accelerates.

The teams that keep people in boxes will get exactly what boxes produce: sequential, narrow, one-thing-at-a-time work. While someone else runs four workstreams before lunch.

We're early. But the gap between organizations that catch this and those that don't will be measurable within a year. Not in AI adoption metrics. In output, speed, and the kind of talent that chooses to stay.

Are your teams structured for who your people are becoming?

A 200-line markdown file and $285 billion off the software market

i.nosov9@gmail.com (Ivan Nosov) — Thu, 12 Feb 2026 00:00:00 GMT

A 200-line markdown file wiped $285 billion off the software market in 48 hours.

Anthropic published an open-source plugin for Claude - structured prompts, first-year law school content, some workflow logic. Thomson Reuters dropped 16%. LegalZoom cratered 20%. The selling spread to private equity. Wall Street called it the SaaSpocalypse.

The market wasn't reacting to a text file. It was repricing the assumption that enterprise software can keep charging per human who logs in, when AI agents can execute the same workflows without a seat license.

But the stock market story isn't even the interesting one.

While traders panicked, KPMG - one of the Big Four - quietly pressured its own auditor, Grant Thornton UK, to cut fees. The argument: AI makes audits faster and cheaper, so pass on the savings. Grant Thornton pushed back. KPMG's response, per the Financial Times: lower your prices or we find a new auditor. Fees dropped from $416,000 to $357,000. 14% gone.

KPMG didn't automate anything. They used the fact that everyone now knows the economics changed.

Here's what I find interesting: software developers saw this coming years ago. They already rewired how they work. AI now writes 30% of Microsoft's code and over a quarter of Google's. CNBC reporters - not engineers, reporters - built a working Monday.com clone with Claude Code in under an hour for $5-15 in compute. A Google principal engineer gave Claude Code a three-paragraph problem description and got back in one hour what her team had been building for a year.

Every other knowledge worker is now in the same position.

I tested this on my own work recently. I built a pay equity analysis framework - EU Pay Transparency Directive compliance, regression analysis multi-country - using Claude Code. Total cost: roughly EUR 30. Time: 2 hours. The consulting engagement it would have replaced? EUR 50,000 minimum.

The data I needed - compensation structures, country-specific legislation, statistical methodology - was irreplaceable. The price tag for turning that data into analysis was not.

That's the pattern everywhere. The expertise, the domain knowledge, the judgment that makes work valuable - that's worth more now, not less. It's the context that separates useful AI output from generic garbage. But the assumption that delivering on that expertise requires the same headcount, the same vendor spend, the same cost structure as two years ago - that broke.

The 200-line markdown file compressed a transition everyone expected to take 5 years into a 48-hour repricing event. And the repricing isn't just hitting software stocks. It's hitting how organizations buy, build, and staff every knowledge-intensive function.

It's a good time to be paying attention.

Where are you seeing this repricing hit first - vendor contracts, team structures, individual workflows?

Software developers rewired how they work in under two years

i.nosov9@gmail.com (Ivan Nosov) — Tue, 10 Feb 2026 00:00:00 GMT

Software developers rewired how they work in under two years.

HR hasn't even started.

HR sits on the most AI-compatible raw material in the company - policies, frameworks, guidelines, job architectures, regulatory docs. All structured text. All logic. The exact same stuff that made coding ripe for AI.

BCG surveyed 1,250 executives. 60% got nothing from AI. Not "limited results." Nothing.

The reason isn't the technology, but the fact taht that nobody changed the actual work.

Software got generative coding. HR gets generative knowledge work.

Meaning the same pattern with the same opportunity and the same narrow window before it becomes a new norm.

Knowledge Compounds

i.nosov9@gmail.com (Ivan Nosov) — Tue, 03 Feb 2026 00:00:00 GMT

Slide 1 of 10

Knowledge Compounds

What happened after we launched a company-wide AI agent

Slide 2 of 10: The Pattern

Launched our first AI agent inside the performance management system. Available to every employee and manager from day one.

Weekly highs keep climbing

Consistent user flow

Usage didn't plateau after the launch campaign ended. Word of mouth worked. People told colleagues the assistant was useful, and the data backs that up.

But the traction wasn't the interesting part.

Slide 3 of 10: Then the Surrounding Functions Noticed

All approaching with the same question

Talent Conversations

Succession Planning

Learning Guidance

Rewards Policy

Not reacting to the technology. Reacting to what happens when domain knowledge gets properly structured and fed into a system.

Slide 4 of 10

"Can we bring our knowledge domain into this?"

What caught their attention was the mechanism.

Slide 5 of 10: The Mechanism

Structured knowledge in, working solution out

Domain Knowledge

Policies, processes, decision logic, edge cases

Structured into Clean Text

Readable by humans, consumable by AI

AI System

Serves thousands of interactions

Updated Within Hours

Policies change, knowledge updates, agent reflects reality

No retraining. No development tickets. No "we'll fix it in the next release."

Slide 6 of 10: Simpler Than It Sounds

Collect whatever already exists

Policies and process guides
Presentations and handouts
Emails and decision frameworks
FAQ documents and knowledge articles
The usual materials on shared drives

1 Restructure through LLMs into clean, readable text

2 Review with domain experts for accuracy

3 Deploy

Slide 7 of 10

Explainable text.

Not a database. Not custom code. Text that humans can read and verify, that AI systems consume directly.

One source of truth, two audiences.

Slide 8 of 10: The Foundation Compounds

01 Performance Management live

02 Talent Conversations building

03 Succession Planning building

04 Learning Guidance next

05 Rewards Policy next

Same architecture, different knowledge. Every domain added strengthens the foundation.

Slide 9 of 10: What Makes This Durable

Models change

New releases, new capabilities, new providers

Interfaces change

Chat, agents, copilots, whatever comes next

Capabilities expand

Reasoning, multimodal, tool use, orchestration

The structured knowledge layer holds.

The technology is the delivery mechanism. The knowledge is the asset.

Slide 10 of 10

The materials already exist.

The extraction tool is language. You can start now.

Claude Code is for you, not only for coders

i.nosov9@gmail.com (Ivan Nosov) — Fri, 30 Jan 2026 00:00:00 GMT

ClaudeCode is for You. Not just for Coders.

The most common question: "This is great, but how do I actually set it up?"

Fair. So here's the full walkthrough.

Windows and Mac. Zero assumptions about your technical background.

What you're installing:

-- VS Code (free, takes 2 minutes)

-- Node.js (free, takes 2 minutes)

-- Claude Code extension (one command)

-- Claude Pro subscription (~20 euro/month)

After that, you open any folder, type what you need in plain English, and it reads your files, writes code, runs analysis, generates reports. No coding on your end.

Setup is the only technical step. Everything after that depends on whether you understand your problem, can ask the right questions, and can validate what you get back. That's domain expertise. You already have it.

I'm also sharing a referral link for Claude Code with one week for free - three uses only, first come first served. Link in the first comment.

Like and repost for visibility of the actually new way of working available to all!

Pay equity analysis in two hours instead of €50K

i.nosov9@gmail.com (Ivan Nosov) — Tue, 27 Jan 2026 00:00:00 GMT

Consultants charge €50K for pay equity analysis.

I built it with AI in 2 hours.

Here's what nobody tells you about the EU Pay Transparency Directive hitting June 2026:

The work isn't complex. It's just expensive.

The Experiment: I forced Claude Code to learn pay equity frameworks and build the entire compliance system from the data part. No templates. No guidance. Just the directive and a dataset.

Result:

Complete framework in <2 hours

2,500+ lines of code

Multi-country regression models (R² > 0.98)

54 automated compliance reports

Production-ready

Spain's case: 25% raw gap → -2% adjusted gap. That's the difference between freaking out and filing compliant reports.

Most €50K consulting projects are just execution. The expertise is already out there - locked behind paywalls and jargon. AI doesn't replace thinking. It democratizes execution for people willing to learn.

Your limits aren't technical anymore. They're imagination + readiness.

I'm sharing the entire framework - code, docs, sample outputs, everything. Not because I'm generous. Because I'm curious who'll actually use it.

Want it? Comment "FRAMEWORK" and I'll send you the link.

(And yes, this carousel was designed by me, compiled by Claude Code)

From Late-Night Deck Crunching to Strategic Thinking

i.nosov9@gmail.com (Ivan Nosov) — Thu, 15 Jan 2026 00:00:00 GMT

Slide 1 of 9: I stopped crunching decks at midnight

How Claude Code transformed my workflow from late-night deck slavery to strategic thinking

Slide 2 of 9: The Time Sink Problem

Large part of my job involves creating presentations, documents, and compelling stories.

70%

of time spent on formatting, not thinking

The irony? Spending most time on the least valuable part — making slides look pretty while burning midnight oil.

Slide 3 of 9: Traditional Workflow Hell

ChatGPT prompt → Copy
Paste into PowerPoint
Format. Fix layout. Adjust colors
Realize you need changes
Back to ChatGPT. Repeat.

Death by a thousand formatting tweaks. Context window limitations. Constant tool-switching torture.

Slide 4 of 9: The Claude Code Revolution

Before

Online chatbots Copy-paste hell Context limits Tool-switching Manual formatting

After

Claude Code Direct file work Full context One environment AI-driven design

Result: 90% of day-to-day materials, zero PowerPoint crunching

Slide 5 of 9: The New Workflow

Formulate workflow & basic rules for AI
Feed context: transcripts, data, requirements
Let AI generate materials directly
Your job: Focus exclusively on the story
Iterate with AI, not with PowerPoint

You become an orchestrator, not an executor

Slide 6 of 9: Real Example: MBA Project

Recent team activity — creating a chart-rich, visually attractive presentation.

Never opened PowerPoint
Fed discussion transcripts to AI
AI adjusted slides based on our conversations
Team focused on discussion & story

100%

collaboration time, 0% formatting time

Slide 7 of 9: The Honest Truth

Yes, skills are still involved You need to know what good looks like

Yes, there are compromises It won't be pixel-perfect custom design

But you get back your most precious resource: Time to actually think

Time for strategy. Time for content. Time for what matters.

Slide 8 of 9: The Meta Moment

This carousel? Made exactly this way

Story formulated first
Design through Claude Code iterations
Zero time in design tools
Direct HTML to PDF conversion

Proof of concept in action. Walking the talk.

Slide 9 of 9: Your job transforms

From

Executor Late nights Formatting Tool-switching Manual work

Orchestrator Strategic thinking Storytelling One environment AI partnership

Are you ready to make the switch?

2025: the year HR started coding

i.nosov9@gmail.com (Ivan Nosov) — Sat, 13 Dec 2025 00:00:00 GMT

2025: The Year HR Started Coding

Birds aren't supposed to swim.

HR people aren't supposed to build.

I did both anyway.

This was my year of daring to actually learn and define the journey that works.

The tools I actually learned to use:

SAP Joule (knowledge management that actually works)
Copilot Studio (in-house custom agents for HR workflows)
Claude Code (my personal favorite - writes better Python than I ever will, allows to actually ideate and make nicely looking html decks)
Azure ML & Google ML (because vendor lock-in is for cowards)
Replit (prototypes in minutes, not weeks)
n8n (automation without begging IT)
PowerApps (turns out low-code isn't a dirty word)
Python deep-dive 🐍 (finally got comfortable with IDEs and actually running solutions even for smaller tasks)

The skills mapping tool? Built it. Performance Management Q&A bot? Ready. The Job evaluator bot that used to take 3 days? 20 minutes. Your well-designed interactive decks for any purpose? Done.

The uncomfortable truth: None of this required budgets or permission from above. Just curiosity and willingness to fail repeatedly. (And I did. A lot.)

The barrier was imaginary all along. 💭

Every tool you learned make the next one easier. Knowledge compounds when you actually apply it. By December, I was solving problems I wouldn't have understood in January.

❓ But why do and learn all of this? Because you can't teach what you don't know. You can't change culture from the sidelines

Dare yourself to build

when you only know how to plan.

Write code

when you're paid for spreadsheets.

Fail publicly

when you're hired to be the expert.

Learn freely

when mastery is assumed.

They'll say: Stay in your lane.

Birds don't swim. HR doesn't code.

Dare anyway.

Swim, bird. 🦅🌊

(inspired by Swim Bird from Blue Prince)

Employees are assets: true or false?

i.nosov9@gmail.com (Ivan Nosov) — Wed, 29 Oct 2025 00:00:00 GMT

Had a moment of surprising freeze during an Accounting Basics review today. The question:

"Employees are assets: True or False?"

My finger hovered. The correct accounting answer is False.

Why? Because an asset, by definition, is a resource a company controls or owns. We (thankfully!) don't own our people. They aren't, and shouldn't be, listed as property on a balance sheet.

And yet... I hesitated.

In the real world of business, how can employees be anything but the most crucial asset? They are the engine. They create the entire value in the value-creation chain. We build entire "People Strategies" precisely because we know their skill, innovation, and dedication are critically correlated with company results.

It highlights a fascinating paradox. We've wrestled for centuries with this -from Karl Marx's critiques of capitalism using workers as tools for exploitation, to today's "war for talent" where we acknowledge human capital is the key differentiator.

But in the black-and-white world of accounting, the answer is still a firm "False."

It's a powerful reminder of the gap between a financial statement's "truth" and the living reality of what makes business succeed.

Karpathy’s march of nines

i.nosov9@gmail.com (Ivan Nosov) — Tue, 21 Oct 2025 00:00:00 GMT

One concept from Andrej Karpathy's latest interview that I take as a large takeaway is The "March of Nines".

I just listened to Andrej Karpathy's 2+ hour conversation with Dwarkesh Patel, and it's exceptional - frank, revealing, and full of insights into how frontier AI actually gets built.

Link to the video: https://lnkd.in/d-Q-spcK

For context: Karpathy founded OpenAI's computer vision research, led AI at Tesla for 5 years building their self-driving system from scratch, and now runs Eureka Labs. This is someone who's actually built the technology, not someone speculating about it.

The interview covers everything from why LLMs are "collapsed," to education post-AGI, to why we're "summoning ghosts, not building animals." But there's one concept I wanted to share because it's rarely discussed yet deeply relevant to how businesses should think about AI deployment:

The "March of Nines" 🎯

Getting a system to work 90% of the time (demo quality) is just the FIRST nine.

Getting it to 99%? That's the second nine - and requires roughly the SAME amount of work.

Then 99.9%, 99.99%, and so on. Each nine demands comparable effort.

During his 5 years at Tesla, Karpathy estimates they progressed through only 2-3 nines. Self-driving still isn't fully solved.

Why this matters for business strategy:

Most AI discussions in business focus on capability: "Can the AI do this task?" But Karpathy's framework reveals the real question: "Which nine do we need, and are we prepared for that journey?" 🤔

Software engineering is his example. The cost of failure is high - bugs expose data, vulnerabilities topple systems. A demo coding agent that works 90% might write a simple app. But production-grade code for millions of users? That requires several more nines and potentially years of iteration.

He saw this at Waymo: a "perfect" demo in 2014, then a decade to limited deployment. Demos hide in narrow conditions. Production faces "all the corners of reality" - edge cases, unusual contexts, failure modes you never anticipated. Even today, Waymo has humans in remote operation centers providing support.

How to think about AI adoption 💡

The march of nines helps you be strategic:

Identify which reliability threshold your use case actually requires

Budget time and resources for the journey between nines, not just the demo

Start with lower-stakes applications where 90-95% might be sufficient

Design human-in-the-loop systems for critical processes

Stay curious about what breaks and why - that's how you reach the next nine

Karpathy's advice throughout: don't write blog posts, don't make slides. Build it. Get it to work. That's how you discover what you're missing. That's how you push to the next nine. 🔨

It's packed with insights like this - the kind that only come from someone who's actually walked the path. 🎧

Your AI just created more work and called it productivity

i.nosov9@gmail.com (Ivan Nosov) — Fri, 17 Oct 2025 00:00:00 GMT

Congratulations! Your AI Just Created More Work (And Called It Productivity)

Let's talk about the elephant in the server room: we're drowning in AI-generated nonsense, and we're calling it innovation.

A recent study from BetterUp Labs and Stanford Social Media Lab, published in HBR, gave a name to something we've all been experiencing: "workslop" - that beautifully formatted, impressively lengthy, completely hollow output that AI spits out when someone hits "generate" instead of "think"

Link here: https://lnkd.in/dd2ws6sE

The statistics are deliciously ironic. Forty-one percent of workers have encountered this in the past month. Each instance may cost up to two hours of rework.

Fantastic example would be Delloite getting caught with ConsultantSlop: https://lnkd.in/dRpaDDvD

I'm genuinely excited about AI - when used by people who understand that technology amplifies intent, not replaces it.

Imagine two colleagues writing to a client about a project delay:

Person A thinks: "We're three weeks behind because the API integration revealed data quality issues we didn't anticipate. The client needs to know this impacts their March launch, but we can mitigate by prioritizing core features first." Then uses AI to polish that thinking into professional prose.

Person B opens a blank email, types "write professional email about project delay," and forwards whatever emerges.

Both emails look equally polished. One advances the work. The other creates a new job: deciphering what actually happened and what to do about it.

When we celebrate "80% of our emails now use AI assistance", we're conflating these wildly different behaviors. We're measuring keystrokes saved, not value created. If your KPI is "percentage of emails drafted by AI," you've fundamentally misunderstood the assignment.

The research identifies "pilots" versus "passengers" - people with high agency who use AI purposefully versus those who use it to avoid thinking altogether. Pilots generate better outcomes. Passengers hit "make longer" and forward the result, offloading their cognitive work onto colleagues.

The consequences? Half of recipients now view workslop senders as less capable, creative, and reliable. Trust erodes. Collaboration fractures. But hey, we automated seventeen tasks this quarter!

The path forward isn't less AI - it's less indiscriminate AI. It requires something quite radical: thinking about what we're trying to achieve before we type a prompt. It means accepting that some tasks genuinely benefit from AI assistance, while others just need a human to do the work.

My provocation for you: What if we stopped celebrating AI adoption rates and started measuring AI value rates? What if we acknowledged that the most productive use of technology sometimes means having an actual thought first?

Rumelt’s Good Strategy/Bad Strategy and the Nvidia example

i.nosov9@gmail.com (Ivan Nosov) — Thu, 16 Oct 2025 00:00:00 GMT

Years before the AI gold rush made it a household name, Richard Rumelt’s classic, "Good Strategy/Bad Strategy," pinpointed Nvidia as a textbook example of brilliant strategy.

I was recently reading this masterpiece, and this example, from a book written way back in 2011, struck me as incredibly fascinating. It's a powerful reminder that the fundamentals of good strategy are timeless.

So, what did Nvidia do right? Rumelt’s core concept is the "kernel" of a good strategy:

A Diagnosis: A clear understanding of the challenge.

A Guiding Policy: The overall approach to tackle the challenge.

Coherent Actions: Specific, coordinated steps that support the guiding policy.

Nvidia’s strategy was a masterclass in this framework. Their diagnosis was that general-purpose CPUs were inefficient for the massively parallel computations required by graphics. Instead of competing head-on with giants like Intel, their guiding policy was to dominate this specialized niche of graphics processing.

This led to coherent actions. They poured R&D into creating powerful GPUs, built a software ecosystem (like CUDA), and focused relentlessly on the gaming and professional visualization markets. They didn't try to be everything to everyone.

What's fascinating is that this focused strategy ran in parallel with the relentless drumbeat of Moore's Law. For decades, this created a vast reservoir of latent computational power. The current AI explosion is, in many ways, the story of humanity suddenly learning how to harness this immense surplus, and Nvidia, thanks to its long-game strategy, was perfectly positioned to provide the tools to do it.

This approach really resonates with my own way of working. It’s not just for business; it’s about personal effectiveness. By diagnosing the real challenge in a task, setting a clear policy for how to approach it, and aligning all my actions, I can cut through the noise and achieve much more. It's about finding that crucial focal point and applying concentrated effort.

It’s a fantastic lesson in the power of clear thinking. What seems like an overnight success is often the result of a brilliant, long-term strategy.

The barrier to AI transformation is not technology or budget

i.nosov9@gmail.com (Ivan Nosov) — Wed, 01 Oct 2025 00:00:00 GMT

The barrier to AI transformation isn't technology or budget. It's imagination and desire.

I recently spent a bit of time building something that helps to solve initial steps in Skills Management complexity.

Here's what I learned:

The Problem: Skills mapping traditionally takes 3-6 months of manual work. By the time you finish, half of the data is already obsolete. So many organizations just... don't do it. We've convinced ourselves it's too hard, too expensive, or requires some magical enterprise solution.

The Reality Check: I built a functional skills mapping application for $25. A generic production-ready version? $100-150.

How it works:

The Concept: You provide a brief job title. You can either add a description yourself or let the AI pre-fill one for you based on the job name.
The Process: Once the job is defined, you run an inference to map out the relevant skills.
The Output: The generated skills can easily be downloaded as JSON files, and a basic dashboard is also included (which can be customized to any format needed).

Demo: https://lnkd.in/dei9fS3V

This simple demo's point is simple: Skills-based AI is a perfect LLM use case. The technology gets you 80% there instantly, creating momentum for refinement rather than starting from zero.

The Broader Insight: We're waiting for "miracle solutions" while the tools to solve our biggest challenges are already here. The transformation doesn't require massive budgets or months of vendor evaluations.

It requires:

✅ A clear understanding of the problem

✅ Willingness to start imperfectly

✅ Basic technical curiosity

✅ Action over analysis paralysis

The Question: If skills mapping - something we've considered too complex to tackle - can be prototyped in a weekend for $25, what other "impossible" organisational and HR challenges are we overthinking?

The tools are here. The costs are negligible. The only thing standing between ideation and implementation is our willingness to experiment, which you would need to do a lot.

What seemingly "impossible" HR challenge are you ready to rethink?

P.S. Yes, the demo is fully functional. Feel free to test it and share feedback. The point isn't perfection - it's momentum.

What sits behind “AI layoffs”

i.nosov9@gmail.com (Ivan Nosov) — Sat, 13 Sep 2025 00:00:00 GMT

The phrase "AI Layoffs" sounds terrifying, doesn't it? When Salesforce CEO Marc Benioff recently said he cut 4,000 jobs because he "needed less heads", it made me wonder what's really going on behind the headlines.

So, I went on a bit of a crusade to figure it out. I dove into the numbers and official statements from companies bragging about "AI-related" changes, including giants like Meta, Microsoft, IBM, and UPS.

Is AI truly replacing us, or is it a convenient excuse for something else? The story is far more intriguing than I expected.

I’ve laid out my findings in this article. I’d love to hear your thoughts and feedback in the comments!

Is Google the dark horse of generative AI?

i.nosov9@gmail.com (Ivan Nosov) — Mon, 08 Sep 2025 00:00:00 GMT

Is Google the dark horse in the generative AI race?

I've been exploring their latest models and had a shot at a number of deep researches with Gemini 2.5Pro, and their video generation Veo. I generated the attached video with a simple prompt, and the quality is what you would expect of top tier!

It's a powerful reminder that while OpenAI often dominates the headlines, the underlying tech race is far from over. Incredible innovation is happening across the board.

While the speed of it dropped notable and it is not the same as jump from GPT 3 to GPT 4, the polishing and specialization happening as we speak

Now I have a question: Why there's a perception gap between Google's Gemini powerful AI capabilities and its market buzz compared to competitors?

Would love to hear your thoughts.

Retrieval as an answer to the knowledge crisis

i.nosov9@gmail.com (Ivan Nosov) — Mon, 01 Sep 2025 00:00:00 GMT

RAG AI: The Savior for Knowledge Crisis

Retrieval-Augmented Generation (RAG) brings one Simple Promises: Transform your existing knowledge into your favorite AI's thoughts.

No massive training, no massive infrastructure investments - just instant, accurate answers from your actual company data.

The Knowledge Mountain Problem: Modern enterprises drown in their own knowledge:

Thousands of policies, procedures, guides

Keyword search fails semantic contex

While AI LLMs excel at understanding meaning, not just words.

How RAG Works (4-Step Process):

📄 VECTORIZE: Documents → Mathematical representations

🗃️ STORE: Knowledge in searchable databases

🔍 RETRIEVE: Context relevant to queries

✨ GENERATE: Accurate answers with citations

RAG in Action - Right in platforms you already use:

🔷 Microsoft Copilot: SharePoint → Instant Q&A, Teams integration through "Knowledge" tab

🔷 Google Gemini: Gem-Bots, Grounding + Vertex AI with multimodal support

🔷 SAP Joule: Full vectorization of SharePoint folders and files (exactly what we're testing actively right now)

🔷 ChatGPT: custom GPTs and context knowledge added

Why This Approach Works: RAG succeeds because it solves universal pain points every organization faces:

Instant policy lookups without hunting through documents

Consistent compliance answers across teams

Self-service employee support that actually works

Democratized access to tribal knowledge

The No-Regrets Strategy:

Start simple: RAG builds AI confidence and adoption

Prove value: Immediate ROI on knowledge access

Scale intelligently: Add agentic layers to high-value processes

Maintain trust: Everything grounded in real company data

RAG can serve as your stepping stone to an intelligent agent that employees actually trust and use.

The 95% AI failure rate, and making the 5% count

i.nosov9@gmail.com (Ivan Nosov) — Wed, 27 Aug 2025 00:00:00 GMT

The 95% AI Failure Rate? I'll Make Those 5% Count and Scale It Forward

Fresh MIT NANDA research dropped in July with an uncomfortable truth: 95% of business GenAI initiatives deliver zero ROI. I'd say it's not a bug, but feature of new technology - this is the learning tax you pay to reach the 5% that actually transform.

I'm ready for 19 failures to get that 20th success - then 21st, 22nd, 23rd... because that's how you build knowledge around what actually works and how to make transformation happen.

Organizations succeeding today never skipped the failure phase. They found ways to accelerate through it.

The data is brutal but illuminating:

• ChatGPT adoption: 80%+ explored, 40% deployed

• Custom enterprise AI: 60% evaluated, only 5% reached production

• Shadow AI economy: 90% of employees use personal AI tools while official initiatives stall

The myth busting: GenAI isn't transforming business, only 5% of enterprises integrate AI in workflows at scale, with 7 of 9 sectors showing no real structural change. Current solutions lack continuous learning, making them "digital amnesiacs" asking for full context every single time.

The architectural reality: GenAI isn't deus ex machina. It has specific use cases. You wouldn't use GenAI for math or predictions, that's Machine Learning territory. You wouldn't put Agentic AI on constant workflows, but to orchestrate pre-set workflows.

Research shows internal builds fail twice as often as external partnerships (33% vs 66% success). Despite my belief in internal capabilities, data suggests strategic partnerships over pure building everything in-house

I'm placing my bets to use Knowledge Management as a stepping stone to build trust, and scale up with advanced and agentic capabilities. Knowledge is often overlooked, but I believe it's the stepping stone for trust, adoption, and experience. Vectorized data and personalized answers leverage exactly what LLMs do best - find and contextualize information that can be actively updated, realizing continuous learning.

Skeptical note: This research has agenda, NANDA pushes the "Agentic Web" forward. But deep research like this helps navigate hype cycles.

Bottom line:

Bridging GenAI to realization requires abandoning familiar failures and embracing the discomfort of learning, adapting, evolving.

The question isn't whether AI will transform business, it's whether business will transform to leverage new technologies beyond just the LLM/Gen part.

Time to get comfortable with being uncomfortable.

ChatGPT 5: why the launch fell flat

i.nosov9@gmail.com (Ivan Nosov) — Fri, 08 Aug 2025 00:00:00 GMT

ChatGPT 5: The Hype Fell Flat - Here’s Why

Is ChatGPT 5 the game-changer we were sold, or just a shiny letdown? The data and my own (brief) experience point to the latter. Despite the hype across the board, it’s not living up to expectations

Here’s the rundown:

Kalshi "Best AI this Month?" Bet: ChatGPT’s odds of being the top AI tanked from 75% to 19% this month, while Gemini raised to 83%. The market’s vote is in - ChatGPT 5 isn’t king

ARC-AGI Leaderboard: Its reasoning skills are underwhelming, scoring just 9,9% compared to Grok 4’s thinking at 16%. No revolutionary leap here.

METR Task Length: Task completion stretched from GPT-4’s 15 minutes to 1-4.5 hours with GPT-5. Better, yes, but barely enough to notice

Social media echoes the disappointment: users call it “incremental at best.” My own tests align (sleep is overrated anyway) - I see slight improvements, but it will certainly not transform how I work. For tricky tasks, I’d still juggle multiple AI tools to get more insights to figure out how to proceed and which route to take

The silver lining? This AI race keeps pushing boundaries, forcing us to stay sharp and adaptable. ChatGPT 5 might not impress, but the competition itself is a win for innovation and for us, end users

What do you think? Has ChatGPT 5 impressed you, or are you feeling the same letdown? Drop your thoughts below - I’d love to hear your take!

Fifteen minutes to build what HR has been overthinking for years

i.nosov9@gmail.com (Ivan Nosov) — Wed, 06 Aug 2025 00:00:00 GMT

15 Minutes to Build as HR What You've Been Overthinking for Years

I've heard countless variations of "I wish we had a tool that could..." followed by resigned acceptance that IT priorities don't align with HR dreams.

Now you can build mock-ups yourself. Today. In about 15 minutes, during your morning coffee time.

I spent yesterday afternoon tinkering with Replit (think of it as a playground where code becomes as simple as Excel formulas with limits, but still). The mission? Create a variable pay modeling tool that actually shows cost implications in real-time. You know, that thing we usually cobble together across a couple of spreadsheets that no one except for you (the designer) can use.

What took 15 minutes:

✅ Setting up flexible goal categories with different weights or multiplication methods

✅ Designer for individual payout scales

✅ Running Monte Carlo simulations (yes, really)

✅ Calculating total cost impact with various performance distributions

💡Check it out:

lnkd.in

The result isn't production-ready. It won't win design awards. But it works. And more importantly, it forced me to think through the logic in ways that Excel never demanded.

The point isn't to become a developer. It's to stop waiting for permission to solve your own problems.

During these August weeks, while emails slow down and meetings mysteriously disappear, why not experiment? Pick that one process that's been irritating you since 2019. That calibration matrix everyone complains about. That benefits calculator gathering dust in someone's desktop folder.

You don't need a real client. Build for yourself. For your team. For that one colleague who still uses printed tables from 2015.

Three reality checks before you start:

💡Keep it simple - complexity is where good ideas go to die

❌Never input real company or employee data (ethics and confidentiality aren't optional)

⚡Your first attempt will be terrible. So was everyone else's

✅Iterate, and discover how it works

The barriers between "technical" and "non-technical" are dissolving faster than sugar in hot tea. Tools like Replit, Claude, ChatGPT - they're not making programmers obsolete. They're making programming accessible to those of us who think in business logic rather than syntax.

Time for upskilling has come, and it's never been easier. Though be mindful of how you use these tools - with great automation comes great responsibility to not accidentally create Skynet in your compensation matrix. 😊

Who else has been secretly building things they were told required "proper IT resources"?

Three and a half charts for a total rewards career

i.nosov9@gmail.com (Ivan Nosov) — Tue, 08 Jul 2025 00:00:00 GMT

The 3.5 Charts That Will Save Your Total Rewards Career (And Your Sanity)

After years of drowning in compensation analytics, I've made a simple discovery: 90% of our questions can be answered with just 3.5 chart types.

Yes, really.

While everyone else is creating 34-slide PowerPoints with violin plots that nobody understands, you can become the office hero with these simple visuals:

📈 Line Chart - "The Time Traveler" Shows where you're going. Headcount trends, salary progression, costs over time. If it moves through time, it gets a line. Simple.

📊 Bar Chart - "The Battle Royale" Department vs department. This year vs last year. Performance ratings. Bars = comparisons. When you need to know who's winning, bars never lie.

📊 Histogram - "Bars with Attitude" It's still bars, but now we're looking at distributions. How many people earn what? Where do performance ratings cluster? It's frequency analysis dressed up to look sophisticated.

🎯 Scatter Plot - "The Detective" The correlation hunter. Turnover vs pay. Performance vs compensation. When you see high pay + high turnover in the same spot, case closed. Each dot should represent a group, not individuals (unless you want privacy nightmares).

That's it. That's the entire toolkit.

While others are drowning in heat maps and waterfall charts, you'll be the person who actually answers the question: "So what does this mean?"

The dirty secret? Complex charts often hide simple truths. The best insights come from asking basic questions with basic visuals.

📈 Lines show trends 📊 Bars show comparisons

📊 Histograms show distributions 🎯 Scatter plots show relationships

Master these four, and you'll solve 90% of compensation puzzles while everyone else is still figuring out how to read their violin plot.

What's your go-to chart for cutting through the noise?

Why centres of excellence keep failing

i.nosov9@gmail.com (Ivan Nosov) — Fri, 27 Jun 2025 00:00:00 GMT

Centers of Excellence solutions consistently fail - why is that? I think I found some ground for the approach I've been trying to implement with MIT's help.

This week's MIT Sloan module on "Harnessing the Crowd" enabled the concept I've been trying to apply: involving stakeholders in active design, rather than passive acceptance. There's a powerful wave of collaborative change happening, and we can leverage it so much further.

After years leading COEs in HR, I'm excited about evolving it through the change of role we can place. While the traditional COE approach isn't broken, it's just ready for transformation.

The Shift: From Expert to Enabler

Instead of asking "What should we build?" I've started asking:

"Who will actually use this?"

"What do our HR partners see in the field?"

"How do managers really make decisions?"

"What would employees design if they could?"

The answers are always better than anything I could create alone.

Breaking Down the Silos

My current AI project exemplifies this shift. Rather than approaching it as a "COE initiative", we've dropped that concept entirely. The focus is simple: products and deliverables that solve real problems.

The result? Solutions that people actually want to use.

The Question That Keeps Me Up

How do we scale this approach? How do we consistently harness collective intelligence without falling back into expert-driven design patterns?

I'm convinced the future of HR isn't about having all the answers - it's about asking better questions and creating spaces where the crowd can solve complex challenges together.

What's your experience with collaborative design in HR? Have you seen traditional COE models evolve?

Why the average employee can never get average rewards

i.nosov9@gmail.com (Ivan Nosov) — Wed, 18 Jun 2025 00:00:00 GMT

The Zero-Sum Confession: Why Your "Average" Employee Can Never Get Average Rewards

There's a moment in every merit cycle when the math stops being friendly. You're staring at a spreadsheet where the numbers simply refuse to add up to everyone's expectations. This is when you discover compensation's cruelest paradox: in a world of budgets, average employees cannot receive average rewards.

The Mathematical Betrayal

Merit Budget: 3% average increase

Leadership: "Take care of our top performers"

The Math: If 20% get 6%, the remaining 80% get 2.25%

Bonus Pool: 100% target payout

Business Reality: "Reward exceptional performance"

The Math: If 15% get 140%, the remaining 85% get 89%

The median employee - literally the middle of your workforce- will receive below-budget increases and below-target bonuses. Not because they performed poorly, but because math is cruel and budgets are finite.

Good Times vs. Hard Times

In good times

Even with a 4.5% merit budget and 120% bonus funding, median employees still got 3.2% increases and 98% bonuses. The zero-sum game isn't about scarcity, but the human nature's need for differentiation.

When Budgets Tightens

With 2% merit budget and 85% bonus pool:

Stars got 4% increases (feeling punished)

Everyone else got 1.5% (below inflation)

Median performers got 75% bonuses (a disguised pay cut)

The Two Paradoxes That Break People's Brains

Paradox #1: The Average Bonus Myth

When we say "100% target bonus," employees hear "expected bonus."

But:

High performers: 120-150%

Median employee: 85-95%

Half your workforce will always receive below-target bonuses.

Paradox #2: The Merit Budget Deception

Finance announces "3% merit budget," but:

Top talent retention: 6-8%

Equity corrections: 5-12%

Everyone else: 1-2%

The "average" employee subsidizes every exception we make.

The Hard Conversations We Avoid

What We Say: "Performance drives rewards"

What We Mean: "Someone's performance drives someone else's disappointment"

What We Say: "Merit budget reflects market conditions"

What We Mean: "Half of you will fall further behind market rates"

The Confession

Your "average" performance will never earn average rewards because average rewards don't exist in a differentiated system. The average is a mathematical artifact created by the exceptional getting more and everyone else getting less.

The system isn't broken - it's working exactly as intended. We just never explain the design honestly.

Instead of selling the illusion of merit-based fairness, what if we admitted:

"We have limited resources. Excellence will be rewarded disproportionately. Meeting expectations will be rewarded minimally. This creates winners and losers by mathematical necessity, not moral judgment."

The zero-sum game isn't going anywhere. The question is whether we'll keep pretending it doesn't exist or start acknowledging that in a world of finite budgets, fairness and equality are mutually exclusive concepts.

The most useful tool I have found for decision making

i.nosov9@gmail.com (Ivan Nosov) — Thu, 05 Jun 2025 00:00:00 GMT

The most powerful tool I've discovered for improving decision making? Transparency.

Every transparency initiative I've led - sharing grades globally, adding compa-ratios to reward statements, even creating open salary boards - has revealed the same pattern: transparency forces better decision-making.

The dirty secret of implementing pay transparency initiatives is in that the the biggest resistance never comes from employees. It comes from those, who aren't ready to explain their decisions.

When you know your team will see everyone's compensation, you suddenly can't justify that random $5K difference based on "they asked nicely" or "good vibes." You have to think systematically, apply criteria consistently, and, most importantly, be able to explain your logic

I'll never forget distributing new pay scales to field workers in Iraq, sharing individual contracts one by one and the approach that we took. Within 5 minutes, they already shared everything with each other anyway. The next 30 minutes? Intense discussion about the pay drivers. Then complete silence. No more questions, no problems since.

Why? Because the criteria were clear, culturally grounded, and consistently applied.

When people understand the "why", they can accept the "what"

Even the USSR figured this out with their skills-based compensation systems (fascinating recent article in The Journal of Total Rewards) - transparent pay levels per job and skills created clarity that actually worked (to some degree)

Transparency doesn't just improve fairness - it improves decision quality. When managers know they'll need to explain pay differences to the entire team, they suddenly become much more thoughtful about those differences. The efficient tension of accountability elevates everyone's game.

I remember sharing Compa-ratios openly in one CIS country. Traditional culture, strong HRBP resistance to sharing compa-ratios. Result? Eight questions in the first month, then it became a natural part of manager-employee conversations.

Here's the thing: you don't need permission to start this change. Just do it.

Yes, we in Total Rewards have some natural leverage—we control the data, we design the processes. But transparency can be driven from anywhere. Manager? Start sharing your team's pay philosophy openly. Employee? Ask direct questions about compensation criteria. Leader? Model the accountability you want to see.

The question isn't whether employees can handle transparency. The question is: are we ready to be accountable for our decisions?

Because once you make that shift, you'll find that transparency doesn't create problems - it solves them

Nobody actually wants data

i.nosov9@gmail.com (Ivan Nosov) — Mon, 26 May 2025 00:00:00 GMT

Nobody actually wants data. They want permission with a spreadsheet attached.

Continuing my 15-year compensation confessions...

Somewhere along my Total Rewards journey, I had a revelation that helped me a lot.

When someone asks "What should we pay for this role?" they're not actually asking for data. They're asking for permission - permission wrapped in a spreadsheet with impressive-looking market analysis.

This realization led me to understand that every compensation professional eventually becomes one of three archetypes when faced with requests to bend the rules:

🚔 The Policeman - "That exceeds our range maximum. Denied."

⚖️ The Judge - "Let me review the facts and make a fair decision"

🤝 The Politician - "Let's explore what we're really trying to accomplish here"

Most of us start as Policemen, evolve into Judges when reality hits, then transform into Politicians to survive. It's like professional evolution, except instead of growing deep expertise, we develop the ability to turn any "no" into "let's find creative alternatives"

I've been all three. The Policeman phase was mercifully brief (turns out being the "fun police" isn't sustainable). The Judge phase felt noble - weighing evidence, making thoughtful decisions - until I realized people still resented having to justify everything to me.

Now? I'm firmly in Politician territory, and honestly, it feels natural. Turns out I genuinely enjoy creating and solving puzzles - and compensation challenges are exactly that kind of puzzle. Can't pay $120K for that role? How about $95K base plus sign-on bonus, flexible working arrangement, and a title that sounds more impressive than it actually is?

The bonus? Your business partners actually start appreciating you for your thoughtful approach rather than seeing you as the person who just says "no" with data attached.

The uncomfortable truth? Politics wins. Every. Single. Time.

Not because it's right or wrong, but because compensation isn't really about perfect systems - it's about managing human expectations in imperfect organizations. And humans prefer creative collaboration over rigid rule enforcement.

The irony? The more you understand compensation theory and best practices, the more Political you become. Your beautiful merit matrices and precisely calibrated salary ranges become just tools in a much larger game of organizational psychology.

Nobody plans to become a compensation professional

i.nosov9@gmail.com (Ivan Nosov) — Thu, 22 May 2025 00:00:00 GMT

Nobody plans to become a compensation professional. It's something that happens to you, like getting caught in the rain or developing strong opinions about Excel formulas (the look of horror when seeing VLOOKUP instead of XLOOKUP in 2025... 😱).

This is a try to reflect on my 15-year journey through the strange parallel universe of Total Rewards. Not a how-to guide—more like field notes from someone who wandered into this world and somehow never found the exit.

After years of reading sanitized case studies and polished best practices, I realized what's missing from most compensation discussions: the messy reality. The contradictions. The moments when your perfectly designed system meets human psychology and organizational politics - and loses spectacularly.

Maybe it's time to talk about what compensation work actually looks like, not what we pretend it looks like in conference presentations.

If we were to plot my compensation expertise over time, it wouldn't be a nice learning curve. It would look more like a heartbeat monitor during a horror movie - wild spikes of confidence followed by devastating crashes of reality.

In the beginning, compensation seemed wonderfully mathematical. The premise was simple: gather market data, design structures, apply policies consistently. My early career was filled with beautiful color-coded spreadsheets calculating market positions to the second decimal place.

I even had special formatting rules for cells exceeding range maximums: bold red with yellow highlight. Because nothing says "this is a problem" quite like making your spreadsheet look like a fast-food advertisement.

Little did I know, I was about to learn that perfect compensation systems are like unicorns - widely discussed, never actually seen in the wild.

Does building Rewards frameworks get easy over time? Not really. But it does get easier as you accept that imperfections of human nature, and that person can adapt and abuse any perfection of a system.

There's an oddly satisfying challenge in creating just enough structure to keep things from falling apart, while leaving just enough flexibility to deal with reality.

Data-Driven Total Rewards: Powering Strategic HR Decisions

i.nosov9@gmail.com (Ivan Nosov) — Thu, 13 Mar 2025 00:00:00 GMT

I. The Fundamentals: Why Total Rewards Exist

Total rewards programs form the foundation of talent management, extending beyond mere compensation to strategically support business objectives by attracting, motivating, and retaining top talent. These systems represent a critical investment in human capital that, when aligned with organizational goals, creates a powerful connection between individual contributions and company success, fostering an environment where employees feel valued and inspired to do their best work.

The Data Connection

While total rewards strategies have always been important, what's continuously revolutionizing the field is the integration of robust data analytics. Today's decisions can't rely solely on intuition or industry conventions—they must be informed by meaningful insights extracted from organizational data. This connection transforms how HR professionals approach rewards strategy:

Raw numbers become compelling narratives about organizational health
Isolated metrics transform into interconnected indicators of performance
Historical patterns evolve into predictive models for future planning

Storytelling with Data

Behind every key performance indicator lies a narrative waiting to be discovered. When a metric like turnover rate increases from 15% to 20%, the number itself matters less than understanding why this shift occurred. Was it concentrated in specific departments? Did it coincide with changes to the rewards structure? Did market conditions play a role?

Effective rewards professionals develop the ability to look beyond surface-level data to uncover these stories. By understanding the "why" behind metrics like engagement scores, compensation ratios, or productivity measures, they can craft more targeted interventions that address root causes rather than symptoms.

II. Addressing Business Challenges with Total Rewards

Organizations frequently encounter several persistent challenges that well-designed total rewards strategies can help address:

Talent Retention Issues: Critical roles experiencing high turnover, resulting in loss of organizational knowledge
Compensation Misalignment: Pay structures that fail to reflect internal equity or market realities, undermining motivation
Cost Management Pressure: Rising market pressure, healthcare and benefits costs straining organizational budgets
Employee Disengagement: Disconnection from organizational purpose leading to decreased discretionary effort

Strategic Metric Groups to Address Challenges

To address these challenges effectively, rewards professionals rely on specific data points that provide objective measures of program effectiveness:

Market Positioning Metrics - compensation positioning and benchmarking, benefits competitiveness, total rewards mix
Retention Analytics - turnover rates across job levels, functions and demographics, correlation between compensation and retention, impact of specific elements on attrition
Financial Efficiency Measures - labor cost to revenue ratios, various productivity metrics
Business Impact Indicators - Connection between retention and customer satisfaction, incentive and revenue, benefits and productivity

Operational Impact

The true power of a data-driven approach emerges when rewards metrics connect directly to broader business outcomes. For example:

Reducing turnover in customer-facing roles might directly impact customer satisfaction scores
Aligning incentive structures with sales targets can drive revenue growth
Optimizing healthcare benefits might reduce absenteeism and boost productivity

By establishing these connections, rewards professionals elevate their work from an administrative function to a strategic driver of business performance.

III. Targeting Investments: Necessity, Not Option

Perhaps the most persistent challenge for rewards professionals is the reality of finite resources. Merit increase budgets, bonus pools, and benefits funding all have limits that require strategic allocation decisions. The art lies in distributing these limited resources in ways that maximize return on investment while maintaining perceptions of fairness.

Using Macro-Level Data

Simple yet powerful visualizations can illuminate the magnitude of these challenges:

Graphs showing the gap between labor cost growth and market rate increases
Comparisons of performance distribution versus reward distribution
Visualizations of budget constraints against ideal market positioning

These tools help stakeholders understand the constraints within which rewards programs must operate and set realistic expectations for what can be achieved.

Targeting Critical Cohorts

Data analysis allows for the identification of segments within the workforce where resource allocation will create the greatest impact:

High-performing employees in roles critical to business strategy
Positions with specialized skills facing acute market pressure
Demographic groups where equity adjustments may be needed
Roles with highest correlation between compensation and performance

By targeting resources toward these strategic segments rather than applying uniform approaches across the organization, rewards professionals can maximize the impact of limited budgets.

IV. Analytical Approaches

Basic Reporting Techniques: The First Step in Data-Driven Decisions

Building data fluency begins with mastering fundamental analytical tools that form the foundation for more sophisticated approaches

Histograms

These frequency distribution charts help visualize how values are distributed across a population, answering questions like:

How are performance ratings distributed across the organization?
What is the distribution of compensation ratios relative to market?
How evenly are merit increases allocated across different departments?

Moving Averages

By calculating averages over rolling time periods, this technique smooths out short-term fluctuations to reveal underlying trends:

Tracking turnover rates over rolling 12-month periods
Monitoring compensation-to-market ratios across quarters
Observing engagement scores over sequential survey periods

XY Scatter Plots

These visualizations reveal relationships between two variables, helping identify patterns and correlations:

Plotting performance ratings against compensation levels
Comparing tenure with promotion velocity
Visualizing the relationship between engagement scores and turnover rates

Mastering these foundational techniques builds both technical capability and analytical confidence. This foundation prepares rewards professionals to advance to more sophisticated approaches while establishing credibility with stakeholders through data-informed insights.

Advancing to Predictive Analytics and Machine Learning

While basic reporting tells us what happened in the past, predictive analytics leverages historical patterns to forecast future outcomes. This shift from retrospective to prospective analysis represents a quantum leap in the strategic value of rewards analytics.

V. Enabling the Success

Data Management Processes

Even the most sophisticated analytical techniques can't compensate for poor-quality data. Establishing robust data management practices is essential:

Implementing validation rules to ensure data accuracy
Standardizing definitions and taxonomies across systems
Establishing clear data governance and ownership
Creating regular audit and cleaning processes

These foundational practices ensure that insights reflect reality rather than data artifacts.

Building Organizational Awareness

For data to drive decisions, it must be understood and embraced by stakeholders throughout the organization:

Translating complex analyses into accessible visualizations
Developing data literacy through training and communication
Creating cross-functional partnerships to contextualize findings
Establishing regular reporting rhythms that align with decision cycles

When data becomes a shared language across the organization, it catalyzes more informed discussions about rewards strategy.

Continuous Improvement

Data-driven rewards management isn't a one-time project but an ongoing journey of refinement:

Regularly evaluating the relevance and accuracy of metrics
Incorporating new data sources as they become available
Refining analytical models based on observed outcomes
Adapting to changing business priorities and market conditions

This commitment to continuous improvement ensures that rewards strategies remain responsive to evolving organizational needs.

VI. Looking Ahead: Challenges and Future Trends

The AI Revolution

Artificial intelligence is rapidly transforming the rewards landscape:

Natural language processing enables analysis of qualitative feedback
Machine learning algorithms can detect subtle patterns in workforce data
Predictive models become increasingly accurate as they learn from outcomes
Automation of routine analyses frees professionals to focus on strategy

These advances create opportunities to develop more sophisticated, personalized approaches to total rewards.

Rapidly Changing Talent Markets

Global talent dynamics continue to evolve at unprecedented speed:

Remote work has expanded geographic talent pools
Skill premiums shift as technology transforms job requirements
Generational preferences reshape expectations for rewards packages
Economic volatility demands more agile compensation approaches

These shifts require rewards professionals to continuously recalibrate their strategies to remain competitive.

Preparing for the Unknown

In a landscape defined by constant change, adaptability becomes the most valuable capability:

Developing scenario planning for different market conditions
Building flexible rewards frameworks that can quickly adjust
Maintaining robust feedback mechanisms to detect emerging trends
Cultivating curiosity and continuous learning

By embracing these forward-thinking approaches, rewards professionals can position their organizations to thrive amidst uncertainty.

Conclusion

The evolution from intuition-based to data-driven total rewards represents a fundamental shift in how organizations leverage their investment in people. For today's master's students, developing strong data literacy alongside traditional rewards expertise is essential.

Numbers alone don't tell the full story. The most effective rewards professionals combine analytical rigor with human insight, using data to illuminate patterns while never losing sight of the people behind the metrics.

The future of total rewards belongs to those who can harness data power while maintaining human connection. This challenging but rewarding path puts you at the intersection of organizational strategy, financial stewardship, and human potential.

Celebrate

Support

Love

Insightful

Funny

Comment

Copy
LinkedIn
Facebook
X

8 Comments

Karolina Cardoso

Report this comment

Just came across your article.. it's really great! It's not about having "new info" in this space necessarily but I find it really useful to see it ALL together in one post. Thanks for sharing!

Reply 1 Reaction

2 Reactions

Reimagining HR Tech Strategy: Where Purpose Meets Innovation

i.nosov9@gmail.com (Ivan Nosov) — Fri, 10 Jan 2025 00:00:00 GMT

HR Tech is evolving quickly, yet many discussions remain focused on digitalization and systems without exploring the deeper purpose behind them. It’s surprising how often the conversation ends at implementing core systems, overlooking the potential for technology to truly serve people and drive organizational success. With AI on the rise and a growing emphasis on meaningful connections, there’s a significant opportunity to view HR Tech as a catalyst for value creation rather than merely a means to automate processes.

In my many years leading Centers of Excellence in various HR organizations, particularly with a focus on Total Rewards, I’ve frequently taken the helm on HR Tech initiatives. More than once, I’ve noticed that the most successful projects follow a similar framework: they align technology with strategic goals and prioritize the needs of their users.

Drawing from these experiences and connecting various ongoing trends, I want to share my perspective on how leaders can develop an HR Tech strategy that not only meets operational demands but also drives meaningful impact. In this article, I’ll outline the principles and steps I believe can guide the transition from basic functionality to a forward-thinking framework—one that helps organizations thrive while remaining attentive to the evolving needs of both their people and their business.

Purpose-Driven HR Tech

Aligning HR technology with organizational goals but understanding how HR Tech can genuinely add value requires a clear and strategic approach. Too often, stakeholders assume that implementing a new system will automatically resolve existing issues. However, technology alone rarely addresses the root causes of problems. Instead, HR Tech should be used as a catalyst to refine processes and unlock true value.

Three Pillars of Purpose-Driven HR Technology can be defined to ensure HR Tech initiatives deliver substantial value:

I. End-to-End Process View

Purpose: By creating efficiencies through a comprehensive view of processes, organizations can secure resources, ensure coherency, and achieve strategic benefits that extend beyond immediate improvements.
Redefine Workflows: Start by understanding and redesigning end-to-end processes to eliminate inefficiencies and align them with strategic goals.
Continuous Improvement: Implement mechanisms for ongoing evaluation and refinement to keep processes aligned with evolving business needs.
How It Works: when reviewing the talent acquisition process, it generally starts way before with defining organization structure, required capabilities, positions, and the outcome impacts the HR Administration and On-boarding directly, while sourcing other employee life-cycle processes.

II. Enhanced Experience

Purpose: Emphasizing a positive feedback loop—improving experiences leads to engaged users, who provide feedback that further enhances experiences, driving engagement and trust in future projects.
User-Centric Design: Involve key stakeholders—employees, managers, and HR teams—in the design and implementation phases to ensure the technology meets their needs.
Engagement and Adoption: Foster a positive user experience through intuitive interfaces, comprehensive training, and continuous support to drive adoption and satisfaction.
How It Works: Creating continuous feedback loop of user experience to facilitate positive changes and building for trust for more feedback and more improvements till the state of the targeted area is self-sufficient

III. Organizational Objectives

Purpose: Linking HR Tech initiatives directly to organizational objectives ensures that each project has a solid business case, reinforcing the role of HR Tech in impacting the P&L and overall business performance.
Strategic Goals Alignment: Clearly define the objectives of each HR Tech initiative and ensure they are directly linked to broader organizational goals.
Measurable Outcomes: Set specific, measurable KPIs to track the success of technology implementations and their impact on the organization.
How It Works: Take the best business-related metrics to track, be it cost per hire, turnover rates for specific cohorts, budgeting accuracy, FTE required to perform process, and strive to achieve them.

Purpose-driven HR Technology is about more than adopting the latest tools; it’s about strategically aligning technology with your organization’s mission and the needs of your people. By emphasizing process optimization, enhancing user experiences, and setting clear objectives, leaders can ensure their HR Tech initiatives deliver meaningful and sustainable value.

Case Study: Self-Service Transformation Success

A real-world example illustrates these principles in action. During a significant HR Tech transformation focused on implementing self-service processes, we applied all three pillars systematically:

End-to-End Process View: We conducted comprehensive analysis of all processes impacted by self-service implementation, identifying optimization opportunities and ensuring seamless integration across different HR functions and tools.

Enhanced Experience: We prioritized user experience by engaging HR transactional teams and employees as champions in testing new workflows. This approach helped moderate initial resistance by demonstrating the benefits of the new system through hands-on experience.

Organizational Objectives: The initiative was anchored to clear business outcomes, specifically a target reduction in HR function headcount. Through careful implementation and stakeholder management, we achieved ~15% FTE reduction while maintaining service quality.

A key success factor was our commitment to transparency, making HR processes and policies more accessible to all stakeholders. This openness built trust and facilitated change adoption, ultimately enabling us to exceed our organizational objectives while improving overall process efficiency.

Ivan Nosov · Writing

The tasks worked. The job did not.

OpenAI's 722 maths papers and the shape of jobs

The Italian Hackathon League final at OGR Torino

FAIR Barcelona 2026: keynote recap

Everyone becomes a bit of an engineer

Twenty supplier offers: where the change AI brings is hard to reach

The more I delegate to AI, the more distilled knowledge is worth

Databricks as an AI workbench for knowledge work

OpenAI limits Cursor’s access to its models

A watermark on AI text only hurts if AI help is something to hide

Every correction to an AI answer is a process rule you paid for

Microsoft's biggest obstacle to AI ROI may be Microsoft Office

Digital transformation now sits in every leadership job

Model rankings shift; my workflows stay provider-neutral

Start with the output and let it pull the data

When processing time collapses, the bottleneck moves to thinking

10% × 10% × 10%: layered checks that make AI errors rare

What 40+ AI reports from March to June 2026 agree on

Two weeks testing Genie Code as a governed AI workspace

Personal context is the only durable asset in AI

AI adoption is loud, but much of it still feels like extra work

The Specification Problem: Why We Are Paying a 40% "Rework Tax" on AI

The Historical Precedent

The Knowledge Inversion

Where the Gap Actually Sits

Recommended by LinkedIn

The Capability in Practice

The Operational Advantage

The Specification Audit

Key References & Data Sources

Anthropic releases Claude Opus 4.7

Two Chinese AI labs as an experiment in organisational design

Anthropic passes OpenAI in annualised revenue

The play most companies run with AI: hand it to IT

Distillation, licences and trust in the AI model ecosystem

AI and the Solow paradox

When leaders build with AI themselves

Microsoft Copilot Cowork at $99 a month

Software engineering hiring and the Jevons paradox

AI Capability Framework - The Four Engineering Disciplines

The Capability Gap No One Is Naming

Layer 1: Prompt Engineering

Layer 2: Context Engineering

Layer 3: Intent Engineering

Layer 4: Specification Engineering

Where Knowledge Falls

Recommended by LinkedIn

The Capability Stack: How the Layers Relate

The stack in practice: performance management

Emerging Artifact Domains

Implications for Capability Building

A Note on Pace

References

Most AI training programmes teach people to type

Anthropic’s architecture for running HR operations through an LLM

Years of Excel skill, quietly made optional

HBR: AI does not reduce work, it intensifies it

A 200-line markdown file and $285 billion off the software market

Software developers rewired how they work in under two years

Knowledge Compounds

Slide 1 of 10

Slide 2 of 10: The Pattern

Slide 3 of 10: Then the Surrounding Functions Noticed

Slide 4 of 10

Slide 5 of 10: The Mechanism

Slide 6 of 10: Simpler Than It Sounds

Slide 7 of 10

Slide 8 of 10: The Foundation Compounds

Slide 9 of 10: What Makes This Durable

Slide 10 of 10

Claude Code is for you, not only for coders

Pay equity analysis in two hours instead of €50K

From Late-Night Deck Crunching to Strategic Thinking

Slide 1 of 9: I stopped crunching decks at midnight

Slide 2 of 9: The Time Sink Problem

Slide 3 of 9: Traditional Workflow Hell

Slide 4 of 9: The Claude Code Revolution

Slide 5 of 9: The New Workflow

Slide 6 of 9: Real Example: MBA Project