A claims adjuster denies a claim. The policy does not formally cover the case—that much is clear. Nevertheless, he picks up the phone before the letter goes out—because he knows the customer has been insured with this company for twenty years, because he knows the story behind it, and because a “no” at this point costs more than the claim amount itself. This is precisely where the debate about AI in insurance gets interesting—not in terms of whether a machine can classify the case correctly, but in terms of who makes the call.
Things get tricky when a decision not only produces a result but also expresses a stance. An AI can calculate that a customer poses an increased risk after three claims in two years. Whether to cancel their policy as a result is not a mathematical problem—it’s a decision about what kind of insurer you want to be. This is precisely where the difference lies between a task that can be delegated and a responsibility that must be retained.
A second blind spot in many automation projects: they underestimate how much tacit knowledge is embedded in “simple” decisions. An experienced underwriter rarely rejects a policy based solely on the numbers. They notice patterns that don’t appear in any dataset—the broker representative who has come close to making a mistake twice in the past, the type of business that looks harmless on paper but, in practice, leads to disputes more often than average. If this knowledge isn’t consciously fed into the systems, it quietly disappears a little more with every step toward automation. The problem doesn’t become apparent right away. It will become evident in five years, when the generation that possessed this intuition has retired and there’s no one left who can recognize the exception before it becomes a costly precedent.
The honest answer to the original question is therefore: AI may prepare nearly every decision in insurance, but it must not be the sole decision-maker. The difference is not a technical limitation, but an organizational one. Where regulations, liability, or the trust of a long-standing customer are at stake, we need someone who takes responsibility not just in the sense of signing a document, but in the sense of genuine deliberation. Everything else—the legwork, the pattern recognition. the sifting through thousands of policies to find a specific clause—can and should be handled by the machine. The more time this frees up, the more room there is for precisely those cases where a human should pick up the phone instead of sending a form letter.