Want AI but operate in a regulated environment? Look at Revolut and Affirm

"We'd love to use AI, but we're regulated" is one of the most common things we hear. It's a real constraint, but it's not the wall most teams assume it is. Two fintechs have published concrete playbooks for shipping AI inside heavily regulated environments. The numbers are striking — and the approach is more interesting than the numbers.

Revolut: governance as a velocity enabler

The hard part: Revolut got a 99.7% call-handling success rate using voice agents, with 8x faster resolution of tickets.

The harder part: they serve builders, operators, researchers, and compliance at the same time — inside a regulated financial product.

Nikolay Donets has publicly outlined Revolut's framework for launching GenAI products in 90 days under regulatory constraints. The headline insight: governance, designed well, is a velocity enabler, not a brake. They moved governance into the development environment, with clear tiers, predictable review cycles, and regulation treated as a technical requirement with a defined path to production.

This is the opposite of how most regulated companies treat compliance — as a gate at the end of the process that surprises engineering teams late. When governance is a known input with a clear interface, teams stop guessing and start shipping.

Affirm: 58% more PRs merged per week

Affirm (the US equivalent of Klarna) retooled their engineering org for agentic development and reported 58% more PRs merged per week, year-over-year.

How they got there, in their own words:

A working group of 9 engineers built a repeatable agentic workflow over two weeks before the org-wide rollout.
They built context files at multiple levels in the codebase — conventions, domain knowledge, team decisions. Agents need structured context, not just repo access.
CLIs were often more robust than MCPs for the same internal integrations.

Two observations worth sitting with. First: the rollout worked because a small group de-risked it before scaling. The pattern is "prove it with 9 people, then turn it on for everyone" — not "buy seats for the whole org and hope." Second: the context-file insight matches what we see everywhere. Agents are bottlenecked on context, not capability. Investing in structured, layered context documentation pays back faster than almost any other agent investment.