A claim is a decision, not a task

The case is strong and worth stating plainly before taking it apart. Of all the workflows an insurance carrier could automate, claims is the one with the most leverage, because it sits on both sides of the combined ratio at once — the losses paid and the expense of paying them. The numbers being quoted are real: straight-through processing of 70 to 90 percent on low-complexity claims, handling costs down 30 to 40 percent, routine claims settled in minutes rather than weeks. Intelligent document processing reads the medical bills and the police reports; agents set reserves, calculate payouts, and flag anomalies without waiting on a human. For the high-volume, low-stakes majority of claims, this is not hype, and a carrier that refuses it will simply lose to one that didn’t.

But notice the word the whole framing rests on. Claims is described as an administrative hurdle — a manual burden to be removed. That description is worth challenging, because a claim is not a task. It is a decision: about money, made under uncertainty, that the carrier is accountable for and the policyholder is depending on. Automate it under the task framing and you have not just removed a cost; you have delegated a decision. So the rest of this piece is mostly questions — the ones we’d want a carrier to be able to answer before it scales that delegation, not after.

Which way is your automation wrong?

Start with the structure of the decision itself. A claim can be gotten wrong in two opposite directions: you can pay what you shouldn’t — leakage, overpayment, the fraud that slips through — or you can deny or underpay what you should have honored. Both are failures. The question worth sitting with is which of the two your scoreboard can actually see.

Overpayment is incurred loss; it lands on the combined ratio more or less immediately, and the whole apparatus is built to catch it. But when a claim is wrongly denied and closed cheaply, what in your current measurement flags it as an error at all? Closed, fast, inexpensive — by the numbers on the dashboard, doesn’t that read as a win? If it is in fact a mistake, where would its cost show up, and how long after the quarter that booked the win? If the honest answer is some version of “later, somewhere else, hard to attribute,” then one of your two error types sits on the ledger now and the other sits off it, on a lag. And if that’s true, here is the uncomfortable one: would a pipeline you tuned to the ledger you can see do anything but get better at producing the error you can’t?

Straight-through to where?

An STP target of 70 to 90 percent and a cost-reduction goal of 30 to 40 percent are pressure in a single direction — close claims fast and cheap. Fair enough. But what is the cheapest way to hit those targets, and is it reliably the right decision on the claim, or merely the fastest close? If conservative reserving and quick denial move the visible number the right way today, what would tell you whether they moved the real outcome the wrong way? And if that answer only arrives after the reporting period, in a different system, would you trace it back to the automation that caused it — or write it off as bad luck in litigation? The worry isn’t that the metric is lying. It’s the possibility that the metric is improving for precisely the reason you should be most concerned.

What does a fraud flag actually tell you?

The fraud case deserves the same scrutiny, because it’s where the questions get sharpest. Graph analytics and anomaly detection do appear to catch organized fraud earlier and more accurately than older methods — a real gain, worth having. But it’s worth pressing on the number. Is “higher accuracy at detecting fraud rings” the same measurement as “low rate of false accusation against legitimate claimants”? If a model raises the first, does anything guarantee it didn’t also raise the second — and would you see it if it had? Who tends to absorb a false positive: the sophisticated claimant who knows how to escalate, or the one who doesn’t? A flag is a signal; acting on it — slowing, investigating, denying — is a decision made against a real person. So the question to answer honestly is whether your process treats the flag as an input a human weighs, or quietly lets a detection score become an automated denial. Which is it, in practice, at volume, on an ordinary Tuesday?

What is the combined ratio not showing you?

Then a quieter question about the scoreboard itself. Is loss spread evenly across claims, or does it concentrate? If most claims are small and routine while the money and the danger live in a minority — the large, the complex, the contested, the borderline — then what is a blended combined ratio really telling you? Could a carrier post an excellent one by becoming superb at the routine 85 percent while systematically mishandling the consequential 15 percent that drives the litigation and the reserve surprises? If that’s even possible, then the aggregate isn’t summarizing the operation; it’s averaging away the part that matters most. Where, in the numbers you actually look at, would you go to find out whether you’ve become most efficient exactly where being wrong costs the most?

The questions worth being able to answer

None of this argues for slowing down on the routine claims. Automate them aggressively; the downside is bounded and the win is clean. The discipline is reserved for the consequential minority, and mostly it amounts to insisting on questions the dashboard doesn’t ask itself.

Are you measuring the decision, or only the throughput? STP rate and cost per claim tell you the pipeline is fast and cheap; they say nothing about whether the calls were right. The measures that would — wrongful-denial and reversal-on-appeal rates, downstream complaint and litigation rates, reserve accuracy against ultimate paid — are the ones the combined ratio lags. If you’re not tracking the errors that land off-ledger, are you measuring claims, or only the half of them that arrives on time?

Which claims has the system actually earned the right to decide? The defensible posture is to keep the consequential ones routed to accountable human judgment, with the AI as instrument rather than adjudicator, until you can show the automated decision is not merely as fast but as good. And when the industry tells you — correctly — to measure these deployments against P&L rather than treat them as IT experiments, the question to put back is: against which P&L? The quarter’s expense line, or the whole of it, lagging liabilities included?

The promise coming due

The combined ratio will improve. It is built to be moved by exactly what AI does cheaply — fewer hands, faster closes, tighter reserves — and it will move. What the scoreboard cannot tell you is whether the claims you closed were closed correctly, and you will not discover the answer from the number that’s going the right way. A claim is the moment the policy stops being a marketing promise and becomes a real one, coming due. Automate the handling of it; the open question — the one worth building the measurement to answer before you scale — is whether you can still tell, afterward, that the promise was kept. The most efficient claims operation in the market and the one quietly breaking faith with its policyholders can look identical on the dashboard for several quarters. What would let you tell them apart?