AI adoption in insurance is accelerating. But much of the investment is still focused on efficiency: a four-month transformation to turn five clicks into three, automate a manual step, or shave a few cents from the cost of processing a claim.
Meanwhile, dollar outcomes remain largely the same.
This is because insurers are asking the wrong questions by fixating on low-impact, efficiency-centric objectives. Whereas, the highest-impact applications of AI come not from productivity uplifts, but from an entirely different question altogether: What can we do now that was not practically possible before AI?
Claims quality assurance is one of the clearest examples. Traditionally, carriers have audited a small sample of claims around 5%, and sometimes considerably less. The findings arrive weeks or months after the relevant decisions were made. By then, the opportunity to change the claim outcome has passed and QA becomes another rearview mirror: useful for training, compliance, and understanding what went wrong, but limited in its ability to affect the claim itself. That is, dollar outcomes remain, again, unchanged.
With AI, however, this needs no longer be the case. The opportunity being missed is that AI has enabled for reviewing every claim, continuously, and identify issues while there is still time to act. That is, moving claims QA from retrospective auditing to a real-time operating system for better claim outcomes.
The problem with sampling
I recently met with a top-10 carrier whose team described the scale of the problem well. Whilst they were supposed to audit roughly 1% of claims. Even that proved difficult at scale, so the team narrowed its review to California and Florida. Yet, still, in those two states alone they identified roughly $22 million in recoverable opportunity.
It raises an obvious question: what remains undiscovered across the rest of the book?
Traditional QA has always faced a structural constraint. Claims organizations simply cannot have people manually review every file, every note, every document, every call, and every decision. Leading to sampling.
The problem is not only that issues outside the sample remain invisible. Small samples also make it harder to distinguish systematic problems from noise. Teams can spend meetings debating whether the findings are representative rather than deciding what to do about them. And it goes without saying teams that audit themselves has never been the most robust way to independently evaluate quality.
AI removes much of that constraint. Instead of auditing 1% or 5% of claims, carriers can evaluate close to 100%. But coverage alone is still not enough.
Going from 5% to 100% is only the first step
Reviewing every claim sounds transformative. And it is. But, that output remains a quarterly report followed by a training session and the implication being that only the measurement system has improved, whilst operations remained unchanged. Again, dollar outcomes remain unchanged.
A Chief Claims Officer does not need another report telling them that adjusters should contact claimants faster. The organization already knows that. Adjusters know it too. The challenge is making sure the right action happens on the individual claim, at the right moment.
That is why the most powerful version of AI-driven QA is real-time QA.
Instead of identifying problems retrospectively, AI continuously evaluates the claim as it develops. When something meaningful appears, such as a reserve that no longer reflects the exposure, a missed subrogation opportunity, an incomplete investigation, an unsupported liability decision, the system can surface it directly to the people who can act.
The supervisor does not discover three months later that an intervention should have happened. The adjuster gets the signal while the outcome can still be changed, creating a safety net claims organizations have always tried to build, but historically could not operate at scale.
AI should change the outcome, not just the workflow
There will continue to be worthwhile AI projects focused on efficiency. Insurance contains plenty of repetitive work that should be automated. But efficiency should not define the ambition for AI in claims.
Turning five clicks into three is useful. Identifying a reserve problem before it becomes leakage is different. Finding a missed recovery opportunity before it disappears is different. Helping an adjuster identify the three claims that genuinely require attention today is different.
For decades, claims organizations have known that quality matters, but the economics of manual review forced QA to operate retrospectively and on a sample. AI removes that constraint. The opportunity is not simply to audit more claims. It is to make quality assurance part of the claim itself: continuously evaluating what is happening, identifying the highest-value interventions, and giving claims professionals the opportunity to act while the outcome can still be changed.
.png)
.png)
.png)
.svg)
.png)
