Ask almost any claims executive how to use AI responsibly, and you will hear the same four words: human in the loop.
Let the model review the file, route complex cases to a person, and let that person decide. It is a sensible design, but it rests on an assumption the industry rarely tests. We put a lot of effort into asking who checks the machine, but we put almost none into asking: who checks the human?
The Noise Problem: When Experts Disagree
In Noise, Daniel Kahneman, Olivier Sibony and Cass Sunstein asked the question. They describe a large insurance company that gave identical case files to experienced professionals and compared their figures. Executives expected two experts to differ by about 10%, but the real gap was 43% for claims adjusters valuing the same claims and 55% for underwriters pricing the same risks.
The authors call this discrepancy "noise": unwanted variability in the judgments of skilled, well-trained professionals. Much of this noise came from stable differences between individuals, each sure their own view was right. Crucially, noise can be measured without knowing the correct answer. You need only two people and the same file, which matters in a business where the right outcome is rarely confirmed.
Binary Decisions in Trade Credit Insurance
No comparable figure has been published for trade credit insurance. Once a claim is accepted, the payout largely follows the policy formula. Whether anything is paid often turns on binary, "yes or no" calls: Did the insured take reasonable steps to limit the loss? Was a late overdue notification material, or a technicality? Should a limit breach be forgiven for a long-standing client? Two handlers who disagree here don't differ by 43%; they differ absolutely. One pays, and the other declines.
Our usual controls miss this: four-eyes reviews, claims committees, and audits test a file against the rules; none shows whether a colleague would have decided it the same way, because each file is judged once. A well-argued rationale proves a decision is defensible, not that it is consistent.
Why This Matters for AI
This is where AI comes in, because this matters for AI. Insurers often ask whether a model can capture decades of claims expertise. In many teams, that expertise exists as several private versions, each coherent, and seldom compared. Train a model on all of them, and it will faithfully learn the noise.
The pattern differs by organization. Private insurers face noise at scale: high volumes, delegated authority, and teams in several countries reading the same wording. A breach forgiven in one office and penalized in another is invisible to any single file review. Export credit agencies handle fewer, larger, more complex claims, where every case feels unique, and noise hides easily. Yet they act on a government's behalf, so treating like cases alike is a matter of fairness, not just efficiency.
Before AI touches a claim, insurers rightly ask whether it can explain its reasoning, treat similar cases alike, and pass tests on known outcomes. The current process deserves the same questions. AI needs scrutiny too, since a flawed instruction repeats on every file. But an agent can be retested on the same files, and a flaw fixed once is fixed everywhere.
A Framework for Progress: Measure, Agree and Check
The teams that progress fastest do three things: measure, agree and check.
Measure: Several handlers, working alone, assess the same handful of anonymized closed files, and the outcomes and reasons are compared. The aim is to understand the process, not to grade people.
Agree: They write down how contested decisions should be made, with real examples showing where the line sits. They ask committee members to record a view before discussion, so the most senior voice doesn't decide for everyone.
Check: Each decision is compared with the agreed approach before it goes out. AI can do this on every file, flagging departures for a person to judge. A deliberate exception is judgment. An exception nobody noticed is noise.
Where AI Belongs in the Process
The same test also shows where AI belongs. Decisions on which handlers already agree are good candidates for AI support: for private insurers, often high-volume routine calls; for export credit agencies, preparing complex files and checking them against precedent.
Where handlers disagree: agree on an approach first, or AI will simply learn the disagreement.
Decisions that require genuine commercial judgment should stay with professionals, with their reasoning recorded. None of this indicates that claims teams are too "noisy" for AI. In fact, it demonstrates the opposite. AI in claims management works well when it is built on decisions the team has measured and agreed. Then AI makes it possible to apply that standard to every file.
Checking the Human in the Loop
Human in the loop remains the best principle. It works best when the loop includes a check on the human, too. When we built that workflow in London, one participant suggested adding a "human in the loop" to the "human in the loop" step. It got a laugh… but it might have been the best idea in the room.
