AI policy document comparison support.

What comparison support produces

AI policy document comparison support turns several long wordings into one table a qualified person can check. For each topic (insuring clause, key definitions, exclusions, sums insured and sub-limits, excesses, conditions, claims obligations) the table shows what each wording says, quoted or closely paraphrased, with the page and clause number. Where a wording says nothing on a topic, the cell reads “not found in documents provided”, never a guess.

That last rule matters most. “Not found” is not “not covered”. Cover often changes in an endorsement, a schedule or a separate definitions booklet, so the workflow lists every document it read and flags when a wording refers to one it did not receive.

Who uses it

How the workflow runs

Inputs

Policy wordings, schedules, endorsements and definitions documents as PDFs; a comparison template listing topics; the client’s or product’s context where relevant.

Systems

A document store with version history, a long-context model (whole wordings fit within current 1M-token context windows), and a review screen showing each cell beside its source page.

Steps

Identify each document and version; extract text and tables; map clauses to template topics; quote or paraphrase each; cite page and clause; mark gaps; flag cross-references to missing documents; highlight material differences.

Human review

The adviser or product specialist checks every cell they rely on against its citation, corrects errors, and writes the interpretation. A senior reviews comparisons used for client recommendations.

Outputs

A cited comparison table, a list of documents read and missing, and a short list of material differences for discussion.

Measures

Cell-level accuracy on past comparisons, corrections per table, time per comparison, and differences missed that a reviewer later found.

Risks

Missed endorsements, borrowed definitions, smoothed wording, stale versions, and the table being sent to a client as advice.

The NZ rules in play

The Contracts of Insurance Act 2024 was not yet in force on 19 October 2026. The FMA has told insurers it comes into effect on 15 November 2027 and that it expects gap and impact assessments now. One concrete use: under the Act, home and contents policies must not contain a pro rata condition of average (section 72), and in most other policies such a condition has no effect unless the insurer explained it in writing before the contract (section 73). A comparison run across a book of wordings can list every average clause, with citations, for legal review.

Under the CoFI regime, banks, insurers and non-bank deposit takers must treat consumers fairly, which includes assisting them to make informed decisions. For brokers, the FMA says advisers must follow the Code of Professional Conduct for Financial Advice Services, including treating clients fairly and giving suitable advice. A comparison table supports that; it does not discharge it.

Illustrative example: a Wellington commercial broker

Illustrative example, synthetic numbers: a broking firm compares commercial material damage and business interruption wordings from three insurers about 40 times a month. Assumptions: each wording is about 60 pages (roughly 35,000 tokens); a manual comparison takes 90 minutes; checking a cited AI table and writing the interpretation takes 35 minutes.

• Adviser time: 40 x 55 minutes = about 37 hours a month back.
• Model fees: 110,000 input and 4,000 output tokens per comparison on Claude Opus 5.5, at US$4 and US$20 per million tokens (Anthropic, as listed on 19 October 2026), is about US$21 a month. The same wordings recur, so prompt caching (cache reads at US$0.20 per million on that model) cuts the input cost sharply.

What stays with people

Advisers decide what to recommend and explain it to the client. Product and legal teams decide what a clause means and whether to change it. Claims handlers decide coverage on a claim. Nobody sends the raw table to a client as if it were advice.

Where comparisons go wrong

• Missing endorsements. The quote’s endorsement removes an exclusion; the table, built from the base wording, still shows it. List documents read on every table.
• Borrowed definitions. “Flood” defined in one wording is applied to another. Cite the definition used in every relevant cell.
• Smoothing. Two differently worded exclusions are summarised identically. Quote the operative words where the difference may matter.
• Wrong version. A 2024 wording compared against a 2026 one by mistake. Show version dates in the header.

The citation discipline is the same as in contract clause review with source citations, and the retrieval choice is covered in RAG versus long context. Related insurance workflows are claims document triage and broker client communication drafting. Our AI document processing service builds comparison tools, and the insurance overview covers the sector.

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