FCA reopens AI Input Zone, demanding evidence — not theory — from UK banks

TL;DR:

• The Financial Conduct Authority has reopened its AI Input Zone, seeking practical examples of good and poor AI practice across banking, insurance and markets ahead of a planned FCA publication later this year; the submission window closes 19 June 2026.
• Colin Payne, FCA head of innovation, has framed the request bluntly: “Not theory. Evidence.” The regulator wants specific examples spanning governance, resilience, oversight, assurance, deployment controls and consumer outcomes — not principles or policies.
• The move follows the FCA’s expanded AI Live Testing programme, which brought Barclays, UBS, Lloyds Banking Group and Experian into a supervised real-world testing environment for governance, monitoring and deployment assurance, with technical partner Advai.

The shift is significant because it moves UK financial-services AI supervision from a principles-and-attestations posture to an outcomes-and-evidence one. For the QA, model-risk and compliance functions at UK banks, the working definition of “the AI system” has expanded: FCA’s Ed Towers told industry the system now includes the model, the deployment context, governance and human-in-the-loop, evaluation techniques, and input and output controls. That is the testing perimeter, not a wishlist.

The “POC paralysis” critique is now a regulatory expectation

The FCA has repeatedly warned firms about “POC paralysis” — perpetual pilots that never reach production because firms cannot evidence sufficient assurance to deploy. The Live Testing programme is designed to break that pattern, with regulatory cover for firms experimenting in real conditions. Notably, the regulator is applying the same standard to itself: the FCA is reportedly trialling AI capabilities (in an engagement involving Palantir) on live financial-crime intelligence data, rather than synthetic test environments. The supervisor and the supervised are converging on the same evidence bar.

The contrast with European peers is becoming clearer. The EU AI Act omnibus, agreed on 7 May and shifting the high-risk compliance deadline to December 2027, has been criticised for tilting toward simplification and away from enforcement; the FCA is pushing the other way, pulling enforcement-grade evidence forward into the supervisory dialogue. For UK fintechs, third-party AI vendors and software-testing firms, that is a commercial opportunity — assurance services with audit-grade artefacts have a clearer buyer. For US AI vendors selling into UK financial services, it is a procurement constraint: a “trust us, our model is safe” pitch will not survive a Live Testing-style review.

Looking forward

Expect the FCA’s good-and-poor-practice publication later in 2026 to set the de facto UK standard for bank AI assurance, with Bank of England Prudential Regulation Authority and Bank of England Financial Policy Committee work on AI operational resilience landing alongside. For UK SMEs offering AI tools to banks, the practical test is whether your evaluation evidence answers Towers’s expanded definition — model, context, governance, human-in-the-loop, evaluation, controls — not just model accuracy. Firms with that evidence ready by 19 June will have first-mover positioning when the FCA’s examples land in the publication.