---
id: "lib-2026-03-00-enterprise-ai-playbook"
type: "library"
title: "Enterprise AI Playbook"
lang: "EN"
org: "Stanford Digital Economy Lab"
source_type: "Report"
date: "2026-03-01"
week: "2026-W09"
url: "https://nosov.ai/library/lib-2026-03-00-enterprise-ai-playbook/"
original_url: "https://digitaleconomy.stanford.edu/app/uploads/2026/03/EnterpriseAIPlaybook_PereiraGraylinBrynjolfsson.pdf"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Erik Brynjolfsson's Stanford Digital Economy Lab documents 51 live enterprise deployments with measured outcomes, but the sample is selected for success and based on interviews."
summary_basis: "full-text"
original_minutes: "150"
tags: ["enterprise ai", "productivity", "adoption", "management"]
frames: ["solow-paradox"]
---

# Enterprise AI Playbook

Tier A (Strong evidence): Erik Brynjolfsson's Stanford Digital Economy Lab documents 51 live enterprise deployments with measured outcomes, but the sample is selected for success and based on interviews.

Summary based on the full text.

## In short

"The Enterprise AI Playbook: Lessons from 51 Successful Deployments" (Stanford Digital Economy Lab, April 2026) by Elisa Pereira, Alvin Wang Graylin and Erik Brynjolfsson. It studies only AI projects that moved past pilot and deliver measurable value, and asks what those organisations did differently. The main finding is that the hard part is organisational: sponsorship, process redesign, data work and resistance from staff functions. Similar use cases took weeks at one company and years at another.

## Key numbers

- 77%: hardest challenges that were invisible costs: change management, data quality, process redesign
- 71% vs 30%: median productivity gain of escalation-based models vs approval-based models
- 45%: deployments where headcount reduction was the largest outcome
- 42%: implementations where foundation model choice was fully interchangeable

## What it reports

- Sample: 51 cases from 41 organisations in 7 countries and 5 regions, representing over a million employees, studied over 5 months.
- 77% of the hardest challenges were invisible costs (change management, data quality, process redesign). 61% of successful projects had at least one prior failure whose cost does not appear in final ROI.
- Escalation models (AI handles 80%+ autonomously, humans review exceptions) delivered a 71% median productivity gain vs 30% for approval models; the authors note task mix may explain part of this.
- Agentic implementations showed 71% median gains vs 40% for high-automation, but made up only 20% of cases.
- Legal, HR, Risk and Compliance were the most frequent source of resistance (35%), ahead of internal end users (23%).
- Headcount reduction was the largest outcome in 45% of deployments; hiring avoided, redeployment or no reduction accounted for 55%.
- 73% of implementations started small deliberately. Security did not kill a project in any case studied.
- For 42% of implementations the model was fully interchangeable; the authors place durable advantage in the orchestration layer.

## How they got it

In-depth interviews with executives and project leaders at organisations whose AI systems are live, integrated into workflows, used consistently and producing measured value. Interviewees were also asked about failures and abandoned pilots before the current result. Cases are coded across eleven themes (investment, oversight, sponsorship, resistance, headcount, revenue, agents, data, security, model choice) and listed anonymously in the appendix.

## Limits

- The authors state the selection bias: only successes are studied, so the report says nothing about how common success is.
- Productivity figures are self-reported by the deploying organisations and not independently audited.
- Brynjolfsson co-founded Workhelix, a company that measures AI benefits for firms.

## Why it matters for companies

Budget lines for AI should include process redesign, data preparation and change management, which the cases show dominate cost and time. Expect resistance from legal, HR, risk and compliance, and involve them early because some later become enablers. Designing for humans to handle exceptions rather than approve every output produced larger gains in these cases. Spending effort on choosing the best model matters less than building the workflow and integration around it.

Original: https://digitaleconomy.stanford.edu/app/uploads/2026/03/EnterpriseAIPlaybook_PereiraGraylinBrynjolfsson.pdf