A specialty insurer put AI in underwriting. Finance wrote the guardrails.

A specialty insurer with roughly £180m in gross written premium has deployed a model that triages submissions and drafts an initial risk assessment. Underwriters retain the decision. The submission throughput per underwriter is up meaningfully, which is what the press release would say. The part worth reading is how the guardrails were set.

The exception policy — which cases must go to a human regardless of model confidence — was written by the finance function rather than by underwriting or IT. The reasoning was that finance already owns materiality thresholds, already reports on loss ratio by segment, and had no incentive to over-claim the model's performance. The chief underwriting officer describes this as the single decision that made the deployment governable.

Measurement followed the same logic. The team explicitly refused to report accuracy percentages to the board, on the grounds that nobody could interpret them. Reporting is in financial terms: loss ratio on model-triaged business versus the control group, expense ratio, and the volume of cases escalated. Those are numbers a board can challenge.

There is a kill switch and, unusually, it has been tested. A quarterly exercise reverts a segment to fully manual handling for a week. It costs throughput and it has twice surfaced dependencies nobody had documented. The CFO's position is that a control you have never exercised is a control you do not have.

The candid part: the first attempt failed. An earlier version was trained on submission data with inconsistent broker coding and produced confident nonsense on a segment representing eight per cent of the book. It was caught in parallel running, not in production, because the finance function insisted on ninety days of parallel run against the wishes of everyone who wanted the benefit sooner.