FCA good practice examples split AI tasks into automate or abstain

Analysis | By CoverWire, created using AI

The Financial Conduct Authority has published good practice examples on AI, which it says create no new requirements. One describes firms deciding which tasks AI should support, which it should automate, and where it should abstain and hand over to a person. For insurers, the hard part is that AI ability is uneven. A model that does one claims task well may fail at a neighbouring one. CoverWire's analysis is that the line has to be drawn task by task, not system by system.

Key points

• The FCA's good practice examples on AI, published on 19 October, include "choosing where AI should support, automate or abstain".
• The FCA describes this as building technical checks and monitoring into the system, rather than only relying on a person to review outputs.
• Two FCA case studies are about claims: one tested for bias in fraud suspicion and claim acceptance, the other checked AI-written claims information.
• CoverWire's analysis: a model's skill at one task says little about the next, so each task needs its own test.

What the FCA has said

The FCA's page sets out good practice and areas for improvement. It says the examples create no new requirements. Under good practice, firms "considered which tasks AI can support or automate, which should remain staff-led, and where the AI tool should abstain, redirect or escalate to a person."

The same passage says this means building proportionate technical checks and continuous monitoring into the system, "rather than only relying on a person to review, approve, challenge or override its outputs". That is a caution against treating human review as the whole control.

Two case studies come from claims. In one, a firm tested whether signs of vulnerability, such as accessibility needs, influenced fraud suspicion or claim acceptance. In another, a firm checked AI-generated claims information against criteria including completeness, faithfulness to the source and adequacy of reasoning, using both staff and automated checks.

Why uneven ability matters

This part is CoverWire's interpretation. The idea that AI skills transfer unevenly, so that a model can be strong at one job and weak at a closely related one, was raised by Nate B. Jones in a post on narrow superintelligence.

In claims, that would mean a model that summarises a motor claim file well cannot be assumed to judge liability well. In underwriting, a model that extracts data from a submission cannot be assumed to price it well. Each task is a separate decision about whether to automate, support or abstain.

Banks already have a framework that works this way. The Prudential Regulation Authority's model risk principles, SS1/23, apply to banks, building societies and some investment firms, not insurers. They include AI and vendor models. They are useful only as a comparison for how a regulator expects models to be tested and limited to their intended use.

The other explanation

There is another reading. Firms may already draw these lines well, and the FCA page may simply describe existing practice. The case studies show firms testing narrow uses before scaling, which suggests some already work task by task. On that view, nothing new is being asked.

The two readings can both be partly true. The FCA's choice of examples shows what it noticed in its work, not how widespread good practice is.

What it suggests

CoverWire's analysis is that claims and underwriting leaders may want a list of AI tasks, each with a stated decision to automate, support or abstain, and the test result behind it. That turns a general policy into something a reviewer, an auditor or the FCA could check.

Related coverage: FCA says senior managers are accountable for their firms' AI.

Sources

• FCA, AI in financial services: good practice and areas for improvement, 19 October 2026
• Bank of England, SS1/23 model risk management principles for banks
• Idea credit: Nate B. Jones, "Narrow Superintelligence", 13 October 2026

CoverWire is a news and information service. Nothing here is investment advice.