TL;DR
A new report from the Chartered Insurance Institute, “Responsible AI: from policy to practice”, argues that insurers and personal-finance advisers are rolling out AI quicker than their staff can interrogate what it produces. Its sharpest claim is that a human signature on an output is not, by itself, a control. The work grew out of a June roundtable of chartered firms, vendors, academics and other professional bodies.
What a reviewer must actually be able to do
The CII sets a demanding test for oversight. Someone glancing at a recommendation before a customer ever sees it counts for nothing unless that person is engaged, properly briefed and personally answerable for where the decision lands — a far higher bar than the sign-off the phrase normally covers.
Training follows from that. Directors and advisers need sceptical habits of mind and enough standing to reverse a machine’s conclusion, rather than fluency in a vendor’s interface. Firms are also urged to state the problem they are trying to fix before shopping for a tool, and to satisfy themselves that whatever they buy suits the client, the objective and the risk in play.
Brokers face the concrete version of this. Where a carrier or managing general agent leans on AI for pricing, underwriting or claims, the question worth asking is not whether somebody reviewed the result. It is whether that reviewer can see the reasoning, possesses the authority to overturn it, and has been trained to notice when an automated answer is wrong for one particular client.
Supervision has not caught up
The warning lands in a gap. Back in January, the Commons Treasury Committee took the Financial Conduct Authority, the Bank of England and the Treasury to task for a “wait-and-see” stance, observing that better than three-quarters of British financial firms are already running the technology. It asked the FCA to spell out, before the year closes, how consumer protection and senior-manager accountability should operate where AI causes harm.
Grant Thornton’s 2026 poll of 100 insurance executives reinforces the point from outside. Governance or compliance trouble had already derailed or blunted an AI project for 44% of them, and only about a quarter felt certain of clearing an independent governance review within three months. Asked who most needs help adapting, 39% named frontline employees.
Looking forward
The shape of the problem is now familiar across UK financial services: paperwork first, competence later. Lloyd’s market firms illustrate it — 72% of those surveyed have frameworks in place, yet live deployment stays parked in back-office productivity rather than the underwriting and claims judgements that carry customer risk.
Matthew Connell, who runs policy and public affairs at the institute, put the aim as adoption that builds trust instead of eroding professional judgement. Set beside the controls Barclays described this week — telemetry, scoped permissions, kill switches written into the system — the conclusion is the same from both directions. Oversight is something you engineer, not something you assert.