Rolling It Out: Cohorts Beat Mandates

Structured, hands-on enablement in small groups outperforms both the broad rollout and the mandate. Booking.com took 3,000+ developers from under 10% AI adoption to 70% by running three-day workshops (DX). Coinbase mandated Copilot and Cursor onboarding within a week, fired engineers who refused, and CEO Brian Armstrong later called the move "heavy-handed" (secondary coverage; no first-party engineering post on the mandate exists). Both companies got adoption. Booking.com got it by teaching people the workflow; Coinbase got it by firing the engineers who refused, and Coinbase is the one on record regretting the method.

The pattern across the published rollouts is consistent. A rollout puts the intake, ownership, gates, and takeover described elsewhere in front of the team, and the evidence base here is rollout stories, not controlled studies.

Booking.com: the workshop is the unit of adoption

Booking.com's vehicle was a three-day "experience-based accelerator" workshop. The number that explains why it worked: roughly 70% of the code written during those workshops was AI-assisted (DX, DX newsletter).

Engineers left the workshop having already shipped real work the new way, not having watched a demo of it. A license grant transfers access; a workshop transfers the workflow. The gap between "has the tool installed" and "has restructured how they work around it" is where the under-10% number came from, and a seat assignment does nothing to close it.

Monzo: broad rollouts dilute the feedback you need

Monzo tried the broad rollout and walked it back. Their stated reason: large rollouts diluted feedback, so they switched to small cohorts (DX podcast).

A cohort is small enough that the rollout team can actually hear what is failing. When 50 engineers hit the same broken workflow, the signal drowns in volume; when eight do, you can sit with them and fix it before the next cohort starts. Monzo's results under this approach: PRs per engineer up 10–20%, roughly 20% of new code AI-generated, and a clear-eyed naming of code review as the new bottleneck. Enablement that increases output volume creates a downstream capacity problem, which is why this section treats review as a gate and review capacity as their own decisions, not afterthoughts of the rollout.

Coinbase: the mandate worked, at a cost the CEO named

Coinbase's adoption numbers are real. Armstrong reported roughly 40% of daily code AI-generated by September 2025, up from 33% a month earlier, with a target above 50%. The mandate produced adoption.

The cost was that the company's own CEO publicly characterized his method as "heavy-handed" after firing engineers who declined a one-week onboarding window. Whatever the retention and trust effects were, they are not published; the story itself exists only in secondary coverage. What the record does show is the contrast case: Ramp's Inspect carries roughly 30% of merged PRs in its frontend and backend repos with no mandate at all (InfoQ). The units differ (share of merged PRs versus share of code written), but the magnitudes are close, and one was achieved by firing people while the other came from a tool engineers reached for. See the Ramp case study for what made Inspect pull-worthy: the agent gets the same databases, CI, Sentry, and feature flags the engineers have.

Why mandates underdeliver: AI amplifies what you already are

DORA 2025 finds roughly 90% of developers using AI and concludes that AI amplifies existing engineering conditions rather than fixing them. A mandate amplifies your dysfunction at exactly the same rate it amplifies your strengths. Forcing tools onto a team with weak feedback loops and no working agreements gets you more code through the same broken pipes. The enablement path, cohort by cohort, gives you a checkpoint at each step to fix the conditions before scaling the amplifier. That is also the argument for writing down the team's working agreements before the rollout, not after the first incident.

The adoption ceiling is high either way. The Pragmatic Engineer survey of 906 engineers found 55% regularly using agents by March 2026; Spotify reports 99% of engineers using AI tools weekly (Spotify). Engineers are not refusing these tools at scale.

The rollout evidence stops at adoption

Three gaps sit on top of these numbers:

• Durability is unmeasured. Nobody publishes whether Booking's 70% held six months later, so the retention curve after enablement is unknown in public.
• The mandate is secondary. The Coinbase mandate has no first-party account, so its details rest on secondary reporting.
• Compliance is unconnected. None of these rollout stories tie adoption to the compliance properties this handbook cares about.

A cohort program teaches the workflow but says nothing about what gets measured afterward, which is where proving the rollout actually holds begins.