HR evaluation and pay for the AI era: why Ubie removed job-family walls
AI時代のあるべき人事評価・報酬制度とは？Ubieが職種の壁をなくした理由
B · Useful, limited Consultancy-hosted interview with the person who leads AI rollout, HR system design and back office at Japanese health-tech firm Ubie; unusually concrete on AI-drafted performance evaluation and merged roles, but all figures are the guest's estimates.
Summary based on the full transcript·Original: about 45 minutes to listen
In short
A 45-minute Japanese interview (AI Jidai channel, run by Liv Consulting, 27 October 2026) with Makito Hashiyama of Ubie, a health-tech company organised as a holacracy. He holds several roles: company AI rollout, HR and evaluation design, and since April 2026 the back office. After giving Claude Code to all staff in February 2026, Ubie rewrote its talent and hiring requirements, evaluation system and org structure in about 1.5-2 months. Job families such as engineer and product manager were replaced by two archetypes. Performance evaluation is now drafted by AI from activity logs, and pay will be the same for the same grade regardless of job family.
Key numbers
- 200-400"AI partner" personas staff built on the in-house chat tool in spring-summer 2025
- about 2 weekstime to decide in February 2026 to give Claude Code to all staff, management included
- about 90%first-line infrastructure questions and setup now closed by an internal infra agent
- about 70%share of meetings that disappeared, by the guest's own feeling, after transcripts became agent-readable
- 70-80%share of evaluations the target end phase would complete by AI with no human in the process
As reported by the source.
Japanese·Video·opens youtube.com
What it reports
- Spring-summer 2025: an in-house wrapper over ChatGPT and Gemini gained company context and characters; staff built 200-400 "AI partners", including role-model personas, digital twins of busy leaders and an AI scrum master.
- February 2026: Claude Code for everyone, after engineers saw coding tasks "almost disappear" by late 2025 and the tool proved useful for documents, sparring and roadmaps.
- Controls: a weekly 2-hour Friday "AI PC classroom"; two security tiers set at tenant level, a flexible mode for engineers on self-responsibility and a restricted sandbox for less experienced users.
- Merged roles: recruiting and labour administration are now one person; accounting, administration and internal control one person. The argument is that hire-to-exit is one flow and splitting it creates coordination cost.
- An infra agent built from documentation by the SRE team now closes about 90% of first-line infra questions, also for business staff.
- All Google Meet sessions are recorded and transcribed into BigQuery (sensitive meetings tagged and excluded); each person's Claude Code reads them and flags what was said about them while absent. Meetings were cut top-down and did not come back. Culture is kept through at least weekly office days and a company lunch.
- New talent model: humans keep "connecting with people" (emotion, accountability, apology, closed first-hand information) and "putting will into choices". Two archetypes: "dancers" (close to customers, ship fast, FDE-like) and "stage machinery" (platforms, guardrails, policy and industry work).
- Back-office staff are told their new job is to codify decision rules, such as approval thresholds, as company assets that AI can apply.
- Evaluation in three phases: AI drafts and humans calibrate (now entering provisional use); calibration narrowed with the CEO alone deciding; then 70-80% of evaluations done by AI alone. Inputs are a per-person activity report from Slack, Notion and meeting transcripts, a role and expectation database and the codified talent requirements.
- Accuracy is poor for managers and for people whose work runs through 1-on-1s, HR cases or confidential alliance work, because that data cannot be processed the same way.
How they got it
A recorded interview; the guest describes his own company's changes. The digest is based on cleaned auto-captions, which garble names (the company name is captioned several ways).
Limits
- Every figure is the guest's estimate or feeling, not measured.
- Hosted by a consultancy that sells AI-implementation services to executives.
- The pay link and AI evaluation are at an early, provisional stage; he says the pay question is still being debated inside the company.
- Holacracy with role-based work is unusual; merged roles may not transfer to firms with fixed job families or works councils.
- No information on how staff can contest an AI-drafted evaluation, or on legal review of the meeting-transcript pipeline.
Why it matters for companies
This is one of the few public accounts of a company redesigning evaluation and pay, not just workflows, around agent use. The pieces are concrete enough to test: activity logs as the evaluation base, codified decision rules as the back-office deliverable, and job families collapsed into two archetypes. It also shows where it breaks: confidential and people-related work leaves no processable trail, so exactly the roles with the most judgement get the least accurate AI drafts. HR leaders considering AI-assisted evaluation need a policy on which data feeds it, how sensitive meetings are excluded, and how people challenge the result.
Key passages in translation
Japanese original
人間に残るのはこう人と繋がることと、意思を込めて、同じを持つことていう風に、分けてそれを2つだろうということを決めた
We decided that what remains for humans is two things: connecting with people, and putting will into choices.
Japanese original
例えば7割8割ぐらいはもう全てAIに完結する、つまり人間があのプロスに関わらないところ
For example, about 70-80% would be completed entirely by AI, meaning no human is involved in the process.
Japanese original
7割ぐらいミーティングって減った感覚があります。
My sense is that meetings went down by about 70%.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com