Claims leakage, the gap between what an insurer or takaful operator pays and what it should have paid, rarely comes from fraud alone. Much of it comes from ordinary mistakes made under time pressure: a limit not applied, a duplicate bill paid twice, an exclusion missed in a long policy wording.
That is exactly the kind of work AI agents do well. It is also where the temptation to let AI decide is strongest, and most risky.
Where AI genuinely helps
Reading the whole file. An agent can read every invoice, discharge summary and repair estimate, and match each line to the benefit schedule or policy wording in force on the event date. Adjusters often don't have time to do this for every claim.
Applying limits consistently. Room-and-board limits, co-payments, sub-limits and waiting periods are easy to miss and easy for an agent to check.
Spotting duplicates and patterns. The same invoice submitted twice, the same repairer and towing operator on a cluster of new policies, a bill dated before the admission: these are patterns that only become visible across many claims.
Explaining each issue. A good agent doesn't just score a claim. It says what it found and which clause or document line it relates to.
Where AI shouldn't decide
Interpreting ambiguous wording. Policy and certificate wording is sometimes ambiguous, and fair treatment of customers means ambiguity is resolved carefully. That is a judgement for an experienced adjuster, often with reference to past decisions and regulatory expectations.
Medical necessity and reasonableness. Whether a treatment or a repair was reasonable is a professional judgement. An agent can gather the evidence; it should not make the call.
Declining or reducing a claim. Decisions that affect a customer's payment must be made and explained by a person who is accountable for them.
Suspecting fraud. A flag is a reason to look, not a conclusion. Treating flags as accusations is unfair to customers and damaging to trust.
Designing the split
In practice, the best results come from a clear division of labour:
- The agent prepares a summary of each claim: eligible items, items above limits, possible exclusions and missing documents, each with its source.
- The adjuster reviews the summary, applies judgement and decides, recording reasons for any changes.
- A separate weekly review of patterns goes to investigators, who decide which deserve a case.
Fairness and evidence
This split also produces better evidence for complaints and audits. Every decision shows what the agent found, what the adjuster decided and why. When a customer asks why their claim was reduced, the answer is ready and consistent.
Getting started
Start with one product line where volumes are high and rules are clear, such as hospitalisation or motor own-damage. Run the agent alongside your adjusters on recent claims and compare its findings with their decisions. Where they differ, find out why. That conversation is where the value starts.



