The pain
Renewal season at a HK broker firm goes like this. Every quarter, a CSR pulls last year's policy schedule. They open the new carrier quote PDFs. They compare line by line: limits, deductibles, exclusions, premium deltas. They write a side-by-side EN summary for the broker of record. Then, because the client reads in 繁體, they re-write it as a one-pager in Traditional Chinese. Then the broker of record reviews and signs off. Then it goes to the client.
Per renewal: four hours. Per year, with hundreds of renewals across producers: a full week of CSR time, every quarter, that nobody had to spend.
What we shipped
A workflow with three stages, all running on infrastructure the broker already had.
Stage 1: Ingest. The CSR drops the previous policy plus the new carrier quotes into a folder. A watcher picks them up, extracts structured data with Claude's reasoning model: limits, deductibles, exclusions, endorsements, premium delta versus last year, cross-sell triggers (VHIS top-up gaps, missing CI, MPF over the threshold).
Stage 2: Compose. The structured data flows through a prompt that knows the broker's voice and the client's audience. Out comes two artefacts in parallel: an EN broker-view summary with the cross-sell triggers flagged, and a 繁體 one-pager for the client with coverage clearly explained, clauses cited.
Stage 3: Sign-off. Both artefacts land in the broker of record's queue. They review, edit if needed, sign off. Only then does the brief route to the client. The audit log captures every step: what the AI wrote, what the broker changed, who signed when.
What surprised us
Two things.
First: the bilingual quality. Mixed EN and 繁體 generation usually produces stilted Chinese. The trick was not to translate the EN draft, but to compose the two artefacts in parallel from the same structured data, with two different prompts tuned to two different audiences. The 繁體 reads like a HK broker wrote it, because we trained the prompt on the broker's own past renewals.
Second: the cross-sell triggers were where the broker found unexpected value. Spotting that a client's MPF balance crossed a threshold, or that their VHIS plan no longer matches their tax-deductible cap, was previously a CSR-eyeball-and-flag job. Now the system flags it, the broker confirms, and the cross-sell hits the same renewal cycle. Two of the first ten renewals turned into a cross-sell conversation that probably would not have happened.
The playbook
Three steps you can copy.
One. Pick the artefact, not the role. "Build me an AI assistant" is too vague. "Build the bilingual one-pager I send every renewal" is sharp. The artefact has a definition, a quality bar, and a measurable outcome. The role does not.
Two. Compose in parallel, do not translate. If your output is bilingual, treat each language as a first-class artefact, generated from the same structured data via two different prompts. Translation introduces lag and lossy phrasing. Parallel composition is faster and reads native.
Three. The audit log is the trust layer. Insurance work is regulated. Anything client-facing routes through broker of record before send. The AI does not get the last word, your licensed human does. Build that gate from day one. PCPD, IA, your sign-off process, all tied together in one log.
Where this goes next
This broker is rolling the workflow across all their P&C lines next quarter, and their group medical book the quarter after that. We are also stress-testing the same pattern on quote prep (same shape, different inputs) for a second broker firm.
If your team is doing the same four-hour artefact every quarter, it is probably automatable. The first step is the same as ours: a forty-minute discovery call to see whether AI moves the needle for your specific shop, or whether you do not need us yet.