AI in insurance: where it delivers value today

Most insurers now have an AI strategy. Far fewer have AI in production doing work that customers, regulators or the finance team would notice. The difference is rarely the model itself. It usually comes down to choosing use cases that suit AI and building a process that lets people check and trust its output.

This article looks at where AI is delivering value in insurance today and what the successful projects have in common.

What AI does well

Current AI systems, and large language models in particular, are strongest at three kinds of work:

• Reading unstructured material such as documents, emails, PDFs, filings and broker submissions.
• Producing structured drafts by turning that material into data, configuration or a proposed decision.
• Explaining what changed, why a result came out as it did, or what information is missing.

They are weaker when asked to make a final, regulated decision without explanation. Successful projects use AI for the first three tasks and keep people responsible for the decision.

Pricing and product

Pricing is a strong candidate because much of the delay is translation: from a rating guide to a working model, and from an agreed rate change to a deployed model. AI can draft models, extract rate tables and run regression tests, which reduces implementation from months to days. The models need to stay readable and covered by tests so they can be approved. We cover this in more detail in AI in insurance pricing.

Underwriting

In commercial and specialty lines, underwriters spend a lot of time reading submissions, schedules, loss runs and surveys. AI can extract and standardise that information, flag missing data and fill in rating inputs. The underwriter still makes the decision, but spends less time reading.

Distribution

AI is changing how insurance is sold, from quoting in chat to AI agents that request prices through an API on a customer's behalf. This depends on a pricing engine that can be called reliably and that explains its results. AI for insurance distribution covers this in more detail.

Claims

First notification of loss, document intake and triage suit AI well. It can classify a claim, extract details from photos and documents, and route the claim to the right team. Fully automated settlement is a narrower opportunity, usually limited to simple, low-value claims with clear rules.

Operations and compliance

Some of the quickest returns come from routine work: summarising policy wordings, comparing document versions, drafting replies to broker queries and preparing evidence for audits and regulators.

What successful projects have in common

Projects that reach production tend to share five features:

1. A named owner for every decision. AI makes a proposal and a person or team approves it.
2. Explanations designed in from the start. If an output cannot be traced back to its inputs and rules, it cannot be explained to a customer or a regulator.
3. Tests. A use case is ready when it passes a set of known cases every time.
4. Structured output. AI that produces data or configuration fits into existing systems more easily than free text that has to be re-entered.
5. A small first scope. One product, one line of business or one process, expanded once it works.

Where to be careful

Treat claims of wholesale replacement with caution. Fully automated underwriting of complex risks, products priced by models nobody can explain, and claims decisions that cannot be explained to the policyholder all conflict with regulatory expectations on fairness and transparency, as well as with what customers expect.

Summary

AI works best in insurance when it speeds up reading and drafting, shows how it reached its output and leaves the final decision with a person. Use cases that fit this pattern are the ones most likely to reach production.