Where AI Actually Works in Insurance

Insurance has been promised AI transformation for about a decade and has mostly received chatbots. That history has made the industry appropriately sceptical, which is fine — but two things did change substantially, and they are worth separating from the noise.

One: reading documents

Insurance runs on unstructured paperwork. Loss reports, medical records, engineering surveys, policy wordings, broker submissions, photographs of damage.

Getting structured information out of these was previously done in one of two ways. By hand, which is slow and expensive. Or by template-based systems, which required configuring for each specific form and broke whenever a broker changed their layout.

Modern models handle heterogeneous documents without per-template configuration, at accuracy levels that make a human-in-the-loop workflow economical rather than theoretical. That single capability changes the cost structure of submission intake, claims first-notice handling and medical review more than anything in the last twenty years.

It is not a marginal efficiency gain. It removes a category of labour that was previously irreducible.

The pattern that works is not full automation. It is extraction with confidence scoring routed into a review queue:

• Extract the fields, attach a confidence score to each
• Auto-accept above a threshold you have validated against a labelled sample
• Route the rest to a human who sees the source document beside the proposed values

The reviewer confirms or corrects in seconds, where transcribing took minutes. You are not removing the human — you are removing the typing.

Two implementation details separate deployments that hold from ones that quietly degrade. Keep the source document linked to every extracted value permanently, because disputes arrive years later and you need to show what the document said and what the system read. And log the model version alongside every extraction, because you will eventually need to answer which decisions came from which version, and reconstructing that afterwards is not possible.

The most common failure is setting the auto-accept threshold from intuition — someone picks 0.9 because it sounds high — then discovering the aggregate error rate at quarter end. Measure it against your own documents, per field type, because the cost of an error varies enormously between a sum insured and a free-text description.

Two: sorting the work

Deciding which claims can be straight-through processed, which submissions merit an underwriter's attention, and which cases warrant investigation.

This is a classification problem, and it is one where models genuinely outperform static rules — because the useful signal is distributed across many weak indicators rather than concentrated in a few thresholds a person can write down.

Rules-based triage has always struggled here for a structural reason: the rule that catches the fraud you already know about does not catch the next one, and adding rules indefinitely produces a system nobody can reason about.

What did not change: the actuarial core

Pricing risk remains a statistical discipline with regulatory scrutiny, explainability requirements and a body of professional practice behind it.

Teams that try to replace their rating methodology with an opaque model discover the regulatory objection quickly and the reputational one eventually. Use models to enrich the inputs to pricing — better data extracted from submissions, better exposure identification — rather than to replace the pricing itself.

The line on underwriting

Worth being precise, because this is where liability accrues.

A submission arriving with forty pages of attachments contains perhaps fifteen facts an underwriter actually needs. Surfacing those with citations back to source, flagging inconsistencies between the application and supporting documents, and highlighting exposures the submission did not mention but the address or industry code implies — genuinely valuable, and no regulatory problem, because a qualified person decides on evidence they can inspect.

The problem arises when a model makes or effectively makes the decision. Effectively is the operative word. If a model produces a score and the workflow makes overriding it costly or unusual, the model is deciding regardless of what the governance document says. Regulators have become sophisticated about this distinction, and so have plaintiffs' lawyers.

Build the audit trail from day one: for every automated contribution, record the inputs, the model version, the output and the human action taken. That is not only compliance overhead — it is the dataset that tells you whether the assistance is improving decisions, which most deployments cannot currently answer.

Where this fits in a wider programme

The intelligence layer is the highest-return first project for most insurers, and it is routinely skipped because it is less strategically dramatic than a core system replacement.

It returns in months rather than years. It requires no change to your core platform. And it generates operational data that makes every subsequent decision — including the core replacement decision — considerably better informed.

The sensible order is: intelligence layer first, then fix the data foundation, then the worst operational workflow, then pricing tooling if your book justifies it, and the core only if it is genuinely blocking.

That is deliberately the reverse of how these programmes usually get pitched.

Full guide — build versus buy across the six functional areas, bitemporal data modelling, the integration surface and core replacement mechanics: Insurance Software Solutions. If you are weighing a programme and want a straight read, talk to us.

Frequently Asked Questions

Where does AI genuinely help in insurance?

Two places: document intelligence, extracting structured data from heterogeneous unstructured documents without per-template configuration, and triage, classifying what can be straight-through processed versus what needs human attention.

What is the right pattern for document extraction?

Extraction with per-field confidence scoring, auto-accepting above a validated threshold and routing the rest to a reviewer who sees the source document beside the proposed values. Reviewing takes seconds where transcribing took minutes.

How should auto-accept thresholds be chosen?

Measured against a labelled sample of your own documents, per field type, because error costs vary enormously between a sum insured and a free-text description. Re-validate quarterly, since document mix drifts.

Can AI price risk?

No — the actuarial core is unchanged and carries explainability and regulatory obligations. Use models to enrich the inputs to pricing rather than to replace the methodology, which invites both regulatory and reputational objections.

When does underwriting assistance become a liability?

When a model effectively decides — including when a workflow makes overriding its score costly or unusual, whatever the governance documentation states. Regulators assess practical effect rather than stated policy.

What should an insurer do first?

The intelligence layer over existing systems. It returns in months, requires no core change, and produces the operational data that makes larger subsequent decisions better informed. It is skipped mainly because it is less dramatic than a core replacement.