AI for Insurance and Claims Businesses

AI pays first in insurance claims on the paperwork around each decision. That means summarising first notice of loss, sorting documents into the file, drafting customer updates and spotting complaints. Keep the claim decision, settlement amounts and anything that tells a customer no with a person. The duty to handle claims fairly stays with the insurer.

This is general information, not legal advice.

Where does AI pay first in a claims business?

It pays first on the high-volume work that surrounds each decision. That means summarising a first notice of loss, sorting incoming documents, listing what is missing, drafting status updates and writing up calls. Each job is frequent, follows a pattern and takes a claims officer a minute or two to check.

None of these decides anything. Instead, they give claims officers back the time to make decisions properly. They also make files more complete, which helps every later step from review to dispute resolution.

What should stay with people?

Keep every decision that affects what a customer receives: accepting or declining a claim, the settlement amount, any finding of non-disclosure or fraud, and how a vulnerable customer is treated. AI can gather the evidence and draft the letter. A person with the authority makes the decision and signs it.

These are the moments a claim becomes a complaint, a dispute or a headline. They also carry the legal duty. According to ASIC, from 1 January 2022 insurers authorised to provide claims handling and settling services must act efficiently, honestly and fairly.

Keep the approval explicit, too. A draft decline letter that a person merely glances at is still an AI decision in practice. The guide to human in the loop explains how to make sign-off a recorded act rather than a formality.

What did ASIC find about claims handling?

ASIC's 2023 review of home insurance claims found five areas to improve: communication, oversight of third parties, recognising complaints, identifying vulnerable customers, and resourcing. Several are exactly where AI can help, by drafting timely updates and flagging possible complaints, provided a person still decides what happens next.

The review looked at data from more than 218,000 claims lodged between January and March 2022, from six insurers. Its message on communication was blunt: "Insurers must prioritise proactive and clear communication with their customers throughout the claim's life cycle." Status updates are frequent, repetitive and easy to check, which makes them a natural first workflow.

Flagging is the other quick win. ASIC called for better recognition and management of "expressions of dissatisfaction and complaints". So AI can read every incoming message and flag the likely complaints for the complaints team. However, it should never be the thing that decides a message is not a complaint.

What do APRA and ASIC expect from AI governance?

Neither has written AI-specific rules for insurers, and ASIC says existing obligations already apply. APRA's operational risk standard covers the systems and service providers an insurer relies on, and ASIC has warned that governance must keep up with licensees' AI use. Treat an AI tool like any other critical system.

According to APRA, CPS 230 sets "minimum standards for managing operational risk", including business continuity and service provider management. It has applied to all APRA-regulated entities since 1 July 2025, and targeted amendments commence on 1 July 2026. An AI vendor that touches claims data may need assessing as a service provider like any other.

ASIC's 2024 review of 23 licensees, including general and life insurers, analysed 624 AI use cases. Its chair, Joe Longo, noted that AI use "has to date focussed predominantly on supporting human decisions". He was also clear that "Existing consumer protection provisions, director duties and licensee obligations put the onus on institutions".

Privacy adds a further layer. Claims files often hold health and financial details, and the OAIC recommends that organisations do not enter personal information, particularly sensitive information, into publicly available generative AI tools.

How should a claims team start?

Start with customer status updates or document sorting, measured against how the team works today. Move staff onto business AI plans and connect AI to the claims system read only. Run the workflow for a month with every output checked. Track time per claim and complaint numbers before widening it.

Read only matters here. A connection that can write to the claims system could change a reserve or close a file. So start with reading, and grant write access one tool and one action at a time, once each has earned it. The guide to AI agent permissions covers how to do that.

Why does a claims team need one source of truth?

Because claims decisions depend on the exact policy wording, the current claims guideline and the approved way of explaining a decision. If each officer's AI works from old documents or memory, the answers drift. One current, owned source keeps every draft consistent with what the insurer has actually agreed to.

That source holds policy wordings by version and date, claims guidelines, approved customer language and the reasons behind past decisions. Each is owned by a named person. When a new document contradicts an old one, it goes to that person rather than being silently merged. A company brain works this way, holding what is true in sets that are each sourced and owned.

Versioning matters more here than in most businesses, because a claim is assessed against the wording of the policy the customer actually holds. APRA's data risk guidance makes a related point, listing accuracy, completeness, consistency, timeliness, availability and fitness for use as core dimensions of data quality.

Skein builds a Company OS for insurance and claims businesses: what is true about the business, held in sets that are each canonical, sourced and owned, with every connection starting read only and nothing sent without a named person approving it; the Company OS page shows how the layers fit together. That suits an insurance or claims business of 10 to 200 people whose staff already use AI. A large insurer with its own AI platform team, or anyone wanting AI to decide claims, is not a fit.

Frequently asked questions

How is AI used in insurance claims?

Mostly for the paperwork around each decision: summarising first notice of loss, sorting documents, listing what is missing, drafting customer updates, writing call notes and flagging possible complaints. A claims officer checks each output. Decisions to pay, decline or settle a claim stay with a person who holds the authority.

Can AI decide insurance claims?

It can technically be built to, but it should not. Accepting, declining and settling claims carry the insurer's duty to act efficiently, honestly and fairly, along with judgements about vulnerability and non-disclosure. AI should gather evidence and draft the reasoning, and an authorised person should make and sign the decision.

Does APRA CPS 230 apply to AI tools?

CPS 230 is not an AI rule, but it sets minimum standards for managing operational risk, including business continuity and service provider management, for all APRA-regulated entities. An AI tool or vendor that supports claims handling should be assessed under those requirements like any other system the insurer relies on.

Can claims staff put customer details into ChatGPT?

Not into public AI tools. The OAIC recommends that organisations avoid entering personal information, and particularly sensitive information, into publicly available generative AI tools. Claims files often contain health and financial details, so staff need business AI plans with appropriate terms and connections that start read only.

Where should an insurance business start with AI?

Start with a frequent, checkable job such as customer status updates or document sorting. Measure how long it takes today, move staff onto business AI plans, connect AI to the claims system read only, and review every output for a month before measuring the change and choosing the next workflow.

Related reading