AI for New Zealand insurance.

Where AI helps an insurer or broker

AI helps insurers and brokers wherever staff read documents to decide what happens next: sorting a claims inbox, checking what a renewal changed, pulling facts out of an assessor’s report or drafting the update a customer is waiting for. The value is speed to the right person with the right file. The risk is that speed turns into automated decisions about cover, which is where customers, regulators and the Privacy Act all push back.

Public tolerance is low. In the Privacy Commissioner’s 2026 annual privacy survey (1,245 adults, fieldwork March to April 2026), 67% of New Zealanders were concerned about government agencies and businesses using AI to make decisions about people from their personal data, up five points on 2025. Insurers are already in the conversation with regulators: three were among the 13 respondents to the FMA’s 2024 survey on AI in financial services.

Insurance workflow map

The first three rows have worked guides: claims document triage, policy document comparison and drafting broker client communications.

The conduct and contract rules that shape insurance AI

We found no AI-specific guidance from the FMA, the Reserve Bank or the Insurance Council of New Zealand as at 19 October 2026. These existing and coming rules matter most:

• Fair conduct. Since 31 March 2025, licensed insurers have operated under the CoFI regime; 46 insurers were among the 77 institutions the FMA licensed at launch. Each insurer’s fair conduct programme must cover how customers are treated, so an AI system in a claims or sales journey needs to fit inside it.
• Contracts of Insurance Act 2024. Not yet in force as at 19 October 2026. Any part not commenced earlier takes effect on the third anniversary of Royal assent, 15 November 2027. It requires consumer policyholders to take reasonable care not to make a misrepresentation (s 13), implies a term that insurers pay sums due on a claim within a reasonable time (s 66), and prohibits pro rata conditions of average in home and contents policies (s 72), according to the Act as enacted. Templates and AI-drafted letters will need updating.
• Prudential and cyber expectations. The joint RBNZ and FMA expectations of 15 October 2026 apply to regulated entities, including insurers: boards should know where AI is used and who is accountable.
• Brokers giving advice. The FMA’s exploratory review of AI in financial advice opened on 6 August 2026 and covers advice businesses using AI.
• Privacy and discrimination. Health insurers fall under the Health Information Privacy Code 2020. Any scoring that treats people differently on a prohibited ground risks indirect discrimination under section 65 of the Human Rights Act 1993.

This is general information; take legal advice on product, claims and conduct obligations.

Systems insurers and brokers run

Insurers typically run a policy administration and claims platform, a document management system, telephony with call recording, and Microsoft 365. Brokers run a broker management system, email and a CRM. AI connects best at the document layer: reading attachments as they arrive, writing structured fields back to the claim or client record, and drafting into the user’s outbox rather than sending. Check where each provider processes data, because medical and financial details are often in the same claim file; read our data residency guide before approving a provider.

A first project: claims document triage

Triage is a good start because every claim already gets sorted by hand, mistakes show up quickly as misroutes, and nothing about cover is decided.

1. Sample real claims. Take 300 closed claims across motor, home and contents, with their attachments and the team each was routed to.
2. Define the labels. Document types, the fields to extract (policy number, date of loss, cause, amounts) and the routing rules.
3. Score against history. Measure document classification accuracy, field accuracy and agreement with the original routing.
4. Shadow live claims. Run alongside handlers for four weeks; handlers accept or correct every suggestion.
5. Switch on routing, not decisions. Let the system route and request missing documents; keep all cover decisions with handlers.

Measure misroute rate, time from lodgement to first meaningful contact, and the share of claims where missing documents were requested at first contact.

Illustrative example: an insurer receives 1,200 home and motor claims a month, and a triage handler spends about 6 minutes sorting attachments and routing each one, about 120 hours a month. If a reviewed AI pre-sort cuts that to 1.5 minutes per claim, the work falls to 30 hours, freeing roughly 90 hours for customer contact. Our AI document processing service runs this test, and LLM evaluation covers how to score it.

Where AI fails in insurance

Workflows we map in this sector

• Policy document comparison supportTwo or more policy wordings laid side by side, clause by clause, with every cell cited to its page.
• Claims document triageClaim documents identified, filed to the right claim and checked off, so handlers know what is missing on day one.
• Broker client communication draftingRenewal summaries, claim updates and disclosure reminders drafted from your own records, sent only by an adviser.

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