The FCA’s new AI Spotlight Observatory catalogues how financial firms use AI. Each implementation record in it should also say who checks the AI’s output and what happens when the output is wrong.
On 19 November the FCA launched the Observatory, which it calls the first resource of its kind for a regulator anywhere. It draws on use cases from more than 200 organisations across 150 areas of activity, and it was built with the Centre for Finance, Technology and Entrepreneurship and the technology partner NayaOne. Jessica Rusu, the FCA’s chief data, information and intelligence officer, described it as “an open and accessible resource that brings together real-world examples of AI adoption”.
What the Observatory already does well
The FCA says it adds a common taxonomy for comparison and separates use cases from implementations. That second split matters more than the headline count. A use case says what a firm wants AI to do, such as screening payments for fraud. An implementation says how a particular firm did it. A payments or compliance team reading the catalogue will copy implementations, so the implementation record is where the catalogue shapes behaviour.
Access is open, though registration unlocks more detailed insights, including target client information, and lets users request demonstrations from the innovators listed. Firms can apply to be showcased. The FCA says the catalogue will keep growing as new use cases arrive.
The field I would add
The announcement describes adoption: what is being built and where trends are forming. It does not say whether an implementation record carries anything about validation. I think every record should answer two questions in a sentence each: who reviews the system’s output before it affects a customer or a ledger, and what the firm does when the output turns out to be wrong.
The regulators’ own cyber paper, published the same day, gives the reason. It warns that advanced AI systems are non-deterministic and “may produce different results across repeated assessments, making validation of their outputs important”. The same paper says appropriate human oversight remains essential, especially for key decisions. Those are the three regulators’ stated positions, and a catalogue of implementations is where they can become something a bank’s model risk team compares against.
Suppose a listing for agentic payments, one of the emerging areas the announcement names, says only that a firm automated a payment step. A treasury team reading it would not learn whether a person approves the instruction, how the agent’s access is limited, or what the firm does when the agent gets a payee wrong. A reader who copies the listing without those answers copies the automation and leaves out the control. The FCA has pressed on controls before: last month it published examples of AI good practice and areas of improvement, and in October it asked AI providers to add disclaimers to financial advice, as we reported in FCA Calls on AI Providers to Add Disclaimers to Financial Advice.
The strongest objection
The best argument against my proposal is that the Observatory is a discovery tool. Firms apply to be showcased, and innovators want leads. If listing requires a description of controls, some will submit less, and the FCA would end up with a thinner map of the market. The regulator already runs other programmes, including AI Live Testing, where deeper scrutiny belongs.
That objection has force on one point: the catalogue should not turn into an audit. I would answer it on cost and on audience. The request is two sentences per implementation, written by the submitter in their own words, with no evidence attached and no FCA sign-off implied. The audience is the one the FCA names, firms learning from developments across the sector. Learning from an implementation without its control is a partial lesson, and the firm that copies the listing bears the cost of it.
A voluntary field also sorts submitters usefully. Listings that answer the two questions give a buyer something to compare. Listings that leave the field blank tell the reader that nothing was offered.
What I would watch
Three things will show whether the Observatory earns its place. The first is whether implementation records gain a validation field in the next round of additions. The second is whether the registered tier, with its target client information and demonstration requests, becomes the main way vendors reach banks, because that is where a thin listing would do the most damage. The third is whether firms that read the cyber paper’s exposure-management language start asking vendors for validation details in procurement, as we discuss in UK Regulators Propose Joint Cyber Resilience Guidance for Firms.
The Observatory gives the sector a shared vocabulary for what firms are building. Two more sentences per implementation would let it carry the part a risk team needs.
