Skip to content
Home » AI in fraud detection is working. What’s Next?

AI in fraud detection is working. What’s Next?

  • 5 min read

The Association of British Insurers reports that £1.16 billion of fraudulent general insurance claims were identified in 2024, a 2% rise on 2023, with the volume of detected cases up 12% to 98,400. UK insurers have spent the last decade pulling in cleaner internal data, enriching it through industry sources, and tuning models that have brought both false positive and false negative rates down. That is concrete progress, and the people running fraud operations are the first to say so.

But two things are also true. Data and orchestration constraints inside any one insurer set a ceiling on how much further the detection layer alone can push. And the volume of referrals coming out of detection has outgrown the capacity of most SIUs to work them.

Both problems are best solved by extending AI through the next stage of the workflow, not by stacking another detection layer on top.

What detection has fixed, and where the ceiling is

Predictive models have been one of the more positive forces in the fraud function. Rates of false referral are coming down. Heads of Fraud increasingly talk about prioritisation rather than exclusion: the value of a good detection model now is that it points the team at the cases worth working first.

The constraint is not the model. It is the data feeding it and the orchestration around it. Claims systems have changed three times. Free-text fields have been used inconsistently for fifteen years. Industry feeds, internal watchlists and intel records sit in different shapes. Insurers have done serious work to close the gaps with enrichment and integration, but cleaner data is unevenly available, and that unevenness sets a ceiling on how much sharper a single predictive layer can get on its own.

The arms race has moved outside the predictive layer

Generative AI has changed the economics of attack. Allianz reported a 300% rise between 2022 and 2023 in cases where apps were used to distort real-life images, videos and documents. Aviva’s head of claims counter fraud, Peter Ward, has flagged a steady rise in manipulated images and documents supporting opportunistic claims. Mark Allen, the ABI’s Head of Fraud and Financial Crime, put it plainly in the November 2025 release: “Fraudsters are increasingly taking more sophisticated, agile approaches, aided by AI.”

A faster predictive model on its own does not catch a manipulated invoice or a synthetic medical report. That is now a generative-AI problem, and it is solved by a layer of AI that reads the documents in the case, not the model that scored the claim.

The next gains sit in a layered model

Three families of AI are starting to combine across the fraud funnel.

Predictive models continue to do the scoring and prioritisation at volume. Generative AI reads the documents, extracts the facts, summarises the case and surfaces the contradictions. Agent-style automation runs the lookups across LexisNexis, Experian, DVLA, Companies House, IFB, IFR and the internal claims system. The investigator works the cases that need judgement and clears the ones that do not. Each layer makes the next one more effective.

The result is a workflow where the priority cases get worked first, lower-value referrals clear faster, and manipulated documents get caught at evidence review rather than slipping through to settlement. False positive rates come down further because more context arrives at the case earlier. Cycle times tighten in both directions.

Under the FCA’s Consumer Duty, in force since 31 July 2023, fair claims handling is a regulatory expectation, not just a service one. A layered model makes that expectation easier to meet. Every check is logged, every AI action is visible and reviewable, every decision has a trail back to the investigator who made it.

What this looks like in practice

Detection points the team at the right cases. Generative AI extracts and summarises. Agents run the lookups. Investigators decide. The human stays in the loop where it matters.

This is the operating model FraudOps has built for insurers and TPAs in the UK. We sit downstream of detection and alongside the data vendors, the IFB and IFR feeds, and the claims systems already in place. The aim is not to replace any of those layers. It is to make sure the value the detection layer creates does not leak out between the flag and the decision.

 

Detection is doing its job. Generative AI is starting to do its part. The next move is to connect them through the investigation workflow that turns one into outcomes.

Sources