An insurer governs AI-drafted customer communications

The firm

Marlow Vale Insurance is a fictional UK general insurer selling home and motor policies to retail customers. It handles a high volume of customer letters: claim updates, renewal notices and complaint responses.

The challenge

Customer letters were slow to produce and inconsistent in tone and clarity. Complaint-handling teams wanted to use generative AI to draft first versions. The Chief Customer Officer supported the idea, but the Risk function raised Consumer Duty concerns: AI-drafted letters must support customer understanding, and errors could cause real harm, particularly for customers in vulnerable circumstances.

What the firm decided

Start from the customer problem

The firm defined the problem first: customers were receiving letters that were late, hard to understand, or did not answer their question. The goal was clearer, faster letters — not fewer staff. This framing shaped the Consumer Duty assessment.

Register and assess the use case

The use case — “AI drafts first versions of complaint responses and claim updates; a trained handler reviews and edits every letter before it is sent” — was entered in the AI use case register as high risk because outputs reach customers. The assessment covered customer harm, consumer understanding, vulnerable customers, accuracy, data sensitivity and the model provider.

Test before deployment

Before going live, the firm tested AI drafts on a sample of past cases:

• Did drafts accurately reflect the facts and the policy wording?
• Were they clear, fair and not misleading?
• How did quality vary across case types and customer groups?
• What happened with cases involving customers who had disclosed vulnerability?

Testing showed that drafts for complex and vulnerable-customer cases needed more substantial editing, so the firm decided those cases would be flagged for senior handler review. It did not rely on the AI to detect vulnerability — the FCA’s good practice publication notes this as poor practice — but on the case data and the handler.

Design meaningful review

Every AI-drafted letter is reviewed before it is sent. Handlers see the draft, the case facts and the sources the draft relied on, work through a short checklist, and record a decision. Handlers were trained on automation bias. See human oversight of AI.

Tell customers and keep a human route

Letters make clear how to contact a person, and the firm’s chat channel says when a customer is talking to an automated assistant.

How it is monitored

The firm monitors:

• Complaint and re-contact rates for AI-assisted letters compared with others.
• Edit rates — how much handlers change drafts — by case type.
• Review times, with very fast approvals flagged for a quality check.
• Outcomes for cases involving customers with characteristics of vulnerability.

Thresholds trigger a review of the use case, and results feed the board’s Consumer Duty outcomes reporting.

What the evidence looks like

• The use case assessment, approval and conditions.
• Pre-deployment test results and the decisions they led to.
• For each letter: the request, the draft, the checks applied, the reviewer, the decision and the final version.
• Monitoring results against thresholds, and actions taken.

What to take from it

• Customer problem first. It kept the use case focused on outcomes.
• Test for different customers. Average quality hides the cases that matter most.
• Review must be real. Training, checklists and monitoring of review quality made oversight meaningful.
• Records serve two purposes: investigating individual complaints and monitoring outcomes overall.

See Policy & Control or read Consumer Duty and AI.