AI broker client communication drafting.

Which letters a model should draft

AI broker client communication drafting takes the letters advisers write over and over, where the facts already sit in the broking system and insurer documents, and produces a first draft in the adviser’s voice. The adviser edits, adds any advice, and sends. The letters that suit drafting are data-heavy and repetitive; the ones that do not are personal or contentious.

How the workflow runs

Inputs

Client and policy records from the broking system, current and previous schedules, insurer renewal notices and claim messages, the firm’s approved templates, and the adviser’s notes.

Systems

The broking system, Outlook or Gmail for drafts (never auto-send), a document store, and a model API on terms that exclude training on client data.

Steps

Select the letter type; pull figures verbatim; build a change table (last year, this year, difference); fill the approved template; write plain-language explanations only from insurer-supplied reasons; place the draft in the adviser’s mailbox; when the client replies, screen the reply for new facts about the risk.

Human review

The adviser reads every draft against the schedule and sends it personally. A compliance lead samples sent letters monthly and approves template changes.

Outputs

Drafts in the adviser’s mailbox, change tables, a flag on client replies that mention changes to the risk, and a record of who approved what.

Measures

Edits per draft, figure errors found before sending (target zero after send), time per renewal, and client replies flagged and passed to insurers.

Risks

Wrong client data, invented reasons for premium changes, implied advice, outdated legal wording, and promotional content in service letters.

The NZ rules that shape the drafts

The FMA says all advisers must follow the Code of Professional Conduct for Financial Advice Services, including treating clients fairly, acting with integrity and giving suitable advice. At an FMA workshop in April 2026 there was broad consensus that primary responsibility for AI use in financial advice rests with the FAP licensee, and in August 2026 the FMA opened an exploratory thematic review of AI in financial advice. Expect questions about how drafting tools are supervised.

The Contracts of Insurance Act 2024, which the FMA says takes effect on 15 November 2027, changes two things for brokers. For consumer policies, section 13 replaces the old disclosure duty with a duty to take reasonable care not to make a misrepresentation, so reminder templates need rewriting. Section 62 requires a broker who receives a relevant representation from a consumer client before a contract is entered into or varied to take reasonable steps to pass it on to the insurer. Screening client replies (“we’ve converted the garage into a sleepout”) for exactly that is the most valuable check in this workflow.

Illustrative example: a six-adviser Hawke’s Bay brokerage

Illustrative example, synthetic numbers: a brokerage handles 1,800 personal and commercial renewals a year (150 a month) and sends about 60 claim updates a month. Assumptions: a renewal summary takes 20 minutes by hand and 7 minutes to check and finish from a draft; a claim update takes 8 minutes by hand and 3 from a draft.

• Adviser time: 150 x 13 minutes + 60 x 5 minutes = about 37 hours a month back.
• Model fees: 210 drafts at 12,000 input and 900 output tokens on OpenAI’s gpt-6.1-sol, at US$2 and US$10 per million tokens (OpenAI, as listed on 19 October 2026), is about US$7 a month.

What stays with the adviser

Advisers write every recommendation, decide what to tell a client about cover gaps, handle complaints and declined claims, and pass client information to insurers. They send every letter themselves. The model never emails clients directly.

Where drafting goes wrong

• Invented reasons. “Your premium rose because of recent weather events” when the insurer gave no reason. Explanations must quote the insurer’s notice or be left for the adviser.
• Wrong client. A merge pulls another client’s schedule. Show the client name and policy number at the top of every draft for a final check.
• Advice creep. “You may want to consider increasing your sum insured” is a recommendation. Block advisory phrasing in templates and let the adviser add it.
• Marketing creep. The Unsolicited Electronic Messages Act 2007 bars unsolicited commercial messages; keep promotions out of service letters.

Our lesson on drafting emails with human review covers the approval habit, and the Microsoft 365 guide covers drafting into Outlook. Related workflows are policy document comparison support and claims document triage. We build these as AI workflow automation; see the insurance overview.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.