How to draft final response letters to complaints with an AI agent

The decision comes first and the letter second. Once a complaint handler has decided the outcome and any redress, the agent drafts the final response from the case file: what the customer complained about, what was found, the decision and its reasons, the redress, and the customer's referral rights. The handler edits and signs it, and a person sends it.

What is a final response, and who owns it?

A final response is the firm's formal answer to a complaint it could not resolve informally. In regulated retail financial services it usually has to do several things at once: address every point the customer raised, explain the decision in terms the customer can follow, set out any redress and how it was calculated, and tell the customer about their right to refer the complaint to an ombudsman, with the time limit for doing so. It is also the document the ombudsman reads first if the customer refers.

The complaint handler owns the letter; the complaints manager owns the standard; quality assurance samples letters after they are sent. The handler's judgement is the substance of the letter, which is why whether an agent can send the final response has a clear answer: it drafts, a person sends.

What does the agent prepare at each step?

1. Person, handler: records the decision. Upheld, partly upheld or not upheld, for each point, with the reasons and any redress.
2. Agent, depth 1: reads the case file. The complaint in the customer's words, the classification, the investigation notes, the evidence, call transcripts and prior correspondence.
3. Agent, depth 1: lists every point the customer raised, quoted, and checks that the handler's decision addresses each one. Any point without a decision goes back to the handler before drafting.
4. Agent, depth 2: drafts the letter using the approved final response letter template: a summary of the complaint, what the firm looked at, the finding and reasons for each point, an apology where the firm got something wrong, the redress and how it was worked out, and the referral rights paragraph exactly as approved.
5. Agent, depth 1: checks the draft. Reading level, every point answered, redress figures matching the case record, no internal jargon, no personal data about anyone other than the customer.
6. Person, handler: edits and signs. The handler makes the letter theirs, particularly the reasons.
7. Person, handler or team: sends, through the channel the customer prefers. The agent may queue the letter, but release is a held action.
8. Person, QA: samples sent letters against the standard, including a sample of agent-drafted ones.

Which controls apply, and what evidence does each leave?

Decision before draft. The agent cannot start a draft until the handler's decision is recorded. This stops the letter from shaping the outcome. Evidence: a decision timestamp earlier than the draft timestamp on every case.

Every point answered. The point list, with the decision against each, is filed with the letter. Evidence: a reviewer, or the ombudsman, can see that nothing was ignored.

Locked paragraphs. The referral rights wording, and any other text your procedure fixes, is inserted verbatim and cannot be edited by the agent. Evidence: the template version and the locked text, checked automatically before release.

Redress reconciled. Every figure in the letter matches the case record and the payment instruction. Evidence: the reconciliation result attached to the case.

Draft and final both kept. The handler's edits are visible. Evidence: if letters go out unedited case after case, QA can see it and ask whether the handler is really deciding.

What should you measure?

• Points missed, found by QA or in referrals.
• Handler edits per letter, by section. Heavy edits to reasons are expected; heavy edits to summaries mean the agent is misreading files.
• Readability scores against your target.
• Referrals where the ombudsman's view differs from the letter, and why.
• Time from decision to letter sent.

Where does it usually go wrong?

Letters that sound fair but say little. Fluent drafting can cover thin reasoning. The reasons section must point to specific evidence: what the call recording shows, what the terms say, what the account history shows.

Templated empathy. An apology that appears in every letter means nothing. The agent should include one only where the handler found the firm at fault, and say what for.

Redress explained badly. A figure without its workings invites a second complaint. Show what the customer lost, what interest or compensation was added, and the total.

Wrong customer details. Merged files and joint accounts lead to letters containing another person's information, which is a personal data breach. The self-check should look for names and identifiers that do not belong to the complainant.

Complaints about the agent. If the complaint concerns an agent's earlier action, the letter must explain what happened without hiding behind the technology. This is where your AI incident response record helps.

Letters go better when the classification upstream is right; see classifying complaints against regulatory definitions. Where complaint data is processed in new ways, a data protection impact assessment may be needed first. The stance on compliance agents overall is set out in AI compliance agents.