This week, Chris Knight published his first blog post as the FCA's Director of Insurance. It is about conflicts of interest in vertically integrated insurance models. It is not about AI. Not one mention of it.
Read it anyway, because the standard he sets applies to every consequential decision an insurer makes, however that decision is produced.
Three lines stand out.
First: firms must "actively identify, manage and evidence" conflicts. Not assess. Not disclose. Evidence.
Second: "disclosure alone is not enough. Simply telling customers about a conflict doesn't remove your obligation to manage it properly."
Third: firms need "controls that actually work in practice, not just on paper," with "clear senior management accountability," and should expect to "show us how your arrangements deliver good outcomes."
Identify. Manage. Evidence. Show, don't assert. Named accountability. Controls proven in practice, not described in policy documents.
Now apply that standard to a different corner of the same industry.
A growing share of underwriting, pricing and claims decisions are AI-assisted. Ask a firm running them three questions in Knight's own register. Can you evidence where every figure in this decision came from? Can you show that the record of the decision has not been quietly altered since it was made? Can you name the individual who signed it off, and prove it was them?
For most firms the honest answers are: partially, no, and it depends who you ask.
That is not because insurers are careless. It is because AI systems produce fluent, confident output, and fluent, confident output is precisely the thing that is hardest to evidence after the fact. A model states that a claimant's declared income was £48,200. Where did that number come from? Was it in the source documents, or was it plausibly invented? By the time a regulator, an ombudsman or a complainant asks, the answer has to be reconstructed. Reconstruction is the paper-based control Knight is warning about: it exists in policy, and fails in practice.
The alternative is to produce the evidence at the moment the decision is made. Verify each cited figure against the source document before it passes. Write the decision into a record that proves it has not been tampered with. Generate the customer explanation from verified facts, so nothing adverse can be quietly left out. Require a named person to sign off, with their identity committed to the record in a way that cannot be rewritten later.
None of this is a new regulatory demand. It is the same Consumer Duty and the same senior management accountability that already apply. What Knight's post signals, in his first week in the role, is where the supervisory emphasis sits: the FCA does not want to hear that your governance exists. It wants you to be able to show it working, decision by decision.
Here is the detail worth sitting with: no UK insurer has yet been fined by the FCA for AI-assisted decisioning. Not one. In the US, the same questions are already in court, as ongoing lawsuits over claims and coverage algorithms, allegations rather than findings, but discovery is underway. The UK is in the window where the supervisory expectation is published and the first enforcement case has not yet happened. That window is where evidence architecture is cheap. After the first case, it is retrofit.
Banks deploying AI in credit decisions are already confronting this. My view is that insurance is next, not because the FCA has said so about AI specifically (it has not, and I am not claiming otherwise), but because "evidence, don't disclose" is not a conflicts-of-interest principle. It is a philosophy of supervision. Philosophies of supervision do not stay in one product line.
If your firm's AI-assisted decisions had to meet Knight's standard tomorrow, could they?
Source: Chris Knight, "Managing conflicts of interest in insurance," FCA blog, July 2026.