AI return on claims sits in coverage, not speed

On a claims quality desk there is one number that never appears in the report: how many files this month can be sampled at all. At one European insurer, manual review covers roughly five percent of claims. Inside that five percent, reviewers have found reinsurance recoveries that had never been claimed. The other ninety-five percent is not clean. It is unseen.

The same company projects a five to six times return once review coverage moves from five percent to fifty. That number deserves a pause. It does not come from reviewing faster. It comes from the forty-five percent that nobody was reading.

Who sets the sampling rate

In most processes that carry an inspection, audit or review step, the sampling rate is set by headcount rather than by risk. Given twenty thousand items a month and a quality team of a certain size, whether you sample five or eight percent depends on how many reviewers exist and how many items each clears in a day. Risk decides which items get sampled. It does not decide how many.

Once set, the boundary hardens. The rate enters job descriptions, monthly reports and reviewer targets, and after a while it reads like an industry convention instead of a budget decision. One question exposes it: if the team had three more people tomorrow, what would the rate be? If the answer is a different number, the rate was never a risk standard.

Look inside the sampled five percent and a harder property appears: misses are not evenly distributed. Large claims carry more clauses, more cross-system checks and more exceptions that need a human reading, and those are exactly the files random sampling is least likely to catch and most valuable to catch. Random sampling behaves well on an even distribution and does close to nothing on this tail. A problem rate measured on five percent therefore reads systematically below the real one.

Why coverage pays more at the margin

Time savings are linear. Coverage is not. Misses concentrate in complex files, and complex files carry larger amounts. The sixth through tenth percent of coverage finds files worth more on average than the first five, because the first five were picked where reading is easiest and the later ones run into the genuinely hard calls.

This explains a counterintuitive result. Plenty of review programs move coverage from five to six or seven percent, see no visible return, and conclude the approach does not work. The real reason is that five to seven is too shallow to enter the tail. The coverage curve is flat early and steep later, and adding two or three points is close to doing nothing.

Crossing that threshold cannot be done with headcount, because headcount scales linearly and reviewers are hard to hire. The one thing that bends the curve is taking preparation away from the person. A reviewer should not have to open three systems to align clauses; the machine assembles the material and flags where the documents disagree, and the person makes the final call. That does not change judgment quality. It changes how many judgments fit in an hour.

Working the arithmetic

Take a set of parameters of my own, chosen to show the arithmetic rather than to describe any company. Ten thousand claims a month, twelve minutes per manual review, a fully loaded 45 dollars an hour. At five percent: 500 files times twelve minutes is 100 hours, 4,500 dollars.

Now lift coverage to fifty percent in the same month. Five thousand files go through model triage at five cents each, 250 dollars. Triage flags fifteen percent for a human, 750 files at six minutes each, because the material is already assembled, which is 75 hours and 3,375 dollars. Total 3,625 dollars, below the original 4,500, at ten times the coverage.

The reason it is cheaper is simple. The machine does the searching and aligning, the person does the judging, and searching and aligning were most of the original twelve minutes. This is the cost that gets overlooked, because it has always been booked inside reviewer hours and never carried its own line.

Now the return. Of the extra 4,500 files read, assume two percent carry an issue, which is 90 files at an average recovery of 800 dollars, or 72,000 dollars. Both the hit rate and the per-file value are my parameters. Any company can pull its own from historical review records, and should, because they decide whether the project is worth doing. Put that against 3,625 dollars of running cost and the order of magnitude is clear: costs in the thousands, returns in the tens of thousands.

One line has been quietly dropped: false positives. If only 90 of the 750 flagged files actually carry an issue, the other 660 are wasted passes, and wasted passes are not free. They consume 660 times six minutes, 66 hours, 2,970 dollars, roughly ninety percent of total review effort. This is why coverage programs need tiering. Triage should band by confidence, humans read the high band, the low band gets sampled for validation. Without banding, reviewers drown in false positives, the project is declared a failure, and what actually failed was the routing design.

Two kinds of process, two approval tests

Amazon’s finance team offers the widely cited example: generative AI tracking VAT regulatory updates across markets, cutting 26 minutes per update to two, a 92 percent reduction, with eighty percent of AI summaries accepted by experts without edits. That is company self-disclosure and should be read as such, but the structure is worth seeing clearly.

There the value really is time, because the same team can cover more jurisdictions. The process is throughput-bound: work arrives as a fixed queue and output is set by units handled per hour. A sampling-bound process is the opposite. Work arrives far faster than it can be handled, so the bottleneck is not speed but who gets seen. Pricing the first by hours is sound. Pricing the second by hours systematically understates it, because the formula silently assumes that what is being handled today is all that should be handled.

The test between them is concrete. Ask whether ten times the throughput would mean handling the same volume or ten times the volume. Same volume means throughput-bound. Ten times means sampling-bound. The answer is not a technology assessment. It is how the business owner describes their own work.

The classification then sets what the approval test should say. For throughput-bound work, the test is unit cost and cycle time, and acceptance reads time and price. For sampling-bound work, the test should be coverage from one rate to another, the hit rate, and the expected value per finding, because only the product of the three means anything. Getting the test wrong has a specific cost: a coverage project judged on hours scores zero, since per-file handling time often rises while coverage climbs.

More coverage means more problems reported

This is the hardest stretch of such a program. As coverage moves from five to fifty percent, reported problems rise. The rise is not new problems. They were always there, unseen. For the quality function this means the same team must handle ten times the workload and explain why problems suddenly increased.

A predictable organizational response follows: loosen the standard so the count returns to its old level. That move rarely appears as a decision. It shows up as a definition adjustment, a unified judgment scale, a separate track for difficult files. Six months later coverage is high, the problem rate is back where it started, and the project is summarized as a model that added nothing.

There is one way to prevent it. Write the expected problem count into the approval document: at fifty percent coverage, how many more issues are expected, how many enter remediation, how many are logged only. The business owner should sign that number, because it is their workload for the next year and their defense when questioned. Without a signature, the day coverage rises is the day the project starts being challenged.

The economics are almost never the constraint; triage is

Run the same parameters as an inequality. Reviewing one more case costs under a dollar: across five thousand cases the screening model costs 250 dollars and human judgement costs 75 hours at 45 dollars, which is 0.72 dollars a case. A hit is worth two percent of 800 dollars, or 16 dollars a case. Coverage only starts losing money once the hit rate drops below about one in a thousand, or roughly nine in ten thousand.

That inequality holds easily in most quality, reconciliation and alerting processes, which is why whether to widen coverage is never the real question. The real question is triage: which band closes automatically, which gets sampled, which a person must read. Get triage wrong and the team drowns in the first week, after which the project is closed as too many false positives.

Thresholds should not be set by feel. Take three hundred to five hundred already-decided historical cases, sort them by model confidence, and find two cut points. The bottom band may close automatically only if the problems a human audit finds there are rare enough to accept. The middle band is sampled at five to ten percent. The top band is read in full. The band sizes are not fixed and can loosen as judgement quality improves, but every adjustment needs a freshly labelled sample.

Watch the right metric. Overall accuracy gets flattered by the high-confidence band. The number that matters is the miss rate of the auto-closed band, because it is the only band nobody rechecks. Once that number slips, every coverage gain before it was borrowed.

One business case usually holds two processes

Few processes are purely sampling-bound. In most, the first half is preparation: pulling three systems, aligning clauses, assembling the file. That half is throughput-bound. Only the second half, the judgement, is sampling-bound. Merge both halves into one hours figure and the return comes out as zero.

Split the business case into two lines instead. Preparation is accepted on unit hours and cycle time. Judgement is accepted on coverage, hit rate and expected value per finding. Splitting usually shows that the first half saves a definite amount while the second multiplies it, and that the two halves earn their money in different ways. Merged, neither can be explained.

The same inequality says what not to do. In steps where value per finding and hit rate are both low, such as compliance sweeps over low-value content, the expected gain from wider coverage falls below the opportunity cost of the reviewer hours. The right move is to cap coverage there and spend the freed judgement on high-value steps. Two steps in one company sharing a single coverage target is usually this error.

Our reading

In these processes the return comes from coverage, not speed. That claim predicts the next case of the same kind: wherever inbound volume vastly exceeds handling capacity, including contract review, reconciliation, content moderation, security alerts and invoice audit, the first year of value should be estimated as a coverage increment rather than hours saved. Projects priced in hours will reliably compute a return of zero, and will not be renewed in year two.

Two predictions follow. First, within twelve months a set of AI contracts will use coverage as the acceptance unit, phrased as a minimum sampled share, and vendors that move first will set the price. Second, headcount in quality, audit and risk functions will rise before it falls: up while coverage climbs and problems multiply, down once tiering and automatic disposition mature. Whoever budgets that middle stretch will not be forced to lower the standard in month three.

Three questions for anyone running such a program. What is the sampling rate today, and when was it last argued from first principles? If compute were free, would you want to see everything or a portion? When coverage doubles, who remediates the extra findings, and where is that budget? If the third has no answer, keep coverage out of the approval document, because it will be delivered, and nobody will have claimed the cost of delivering it.

Sources and parameters

Manual review covering roughly five percent of claims, unpublished reinsurance recoveries surfacing inside that sample, coverage near fifty percent and a projected five to six times return: interview-based write-up, secondary, and the return is an expectation rather than a realized result. fundaai.substack.com/p/deepllm-enterprise-ai-application-2e9

VAT update handling cut from 26 minutes to two, a 92 percent reduction, with eighty percent of AI summaries accepted without edits: Amazon finance technology case as cited by Bain, company self-disclosure, not independently verified. bain.com/insights/your-ai-budget-is-growing-your-returns-arent-heres-why

The split between coverage metrics and hour metrics in procurement, and the understatement that follows from approving on hours: this piece’s framing, inferred from the acceptance criteria used for comparable projects in the two sources above. It is an independent reading, not a conclusion stated in either report.

Parameters in the worked example: ten thousand claims a month, twelve minutes per manual review, 45 dollars fully loaded hourly cost, five cents per triage pass, fifteen percent flag rate, six minutes per flagged review, two percent hit rate, 800 dollars per finding. All are mine, set to show the arithmetic, and none is an industry average. Every multiplication and percentage above can be recomputed from them.