What an Optical Character Recognition (OCR) pass found across 2 commission books, and the arithmetic that explains why nobody had looked before.
In short. Across a complete enrollment cycle, every policy in 2 books totalling 29,435 was verified individually against its actual status at the source. 10,291 of them had charged back, about 35% of the total, against advances that had already been paid. The 2 books belong to affiliated agencies running different payment models, 1 on fixed advances and 1 as earned, and the same pass read both. The verification had never been done at this agency at any price, and the reason is arithmetic rather than negligence. At 1 minute a record it is 490 hours of work with no guaranteed payoff. The interesting design decision was not the reading, it was choosing to abstain rather than guess, because the cost of a wrong flag and the cost of a missed one are nowhere near equal.
The work nobody could authorise
A producer is paid before the money is confirmed. A policy is written, the advance goes out that week, and whether the policy survives is settled months later by somebody else. Almost nobody checks, and the usual explanation is that firms are careless. That is not it.
Checking means opening records one at a time. 29,435 of them. At 1 minute a record, generous for a task that involves finding an account, locating the right row, and reading a status field, that is 490 hours. Call it a third of a full time person doing nothing else for 6 months, spent on a task whose payoff is unknown until it is finished.
No operator authorises that. Not because they doubt there is money in it, but because the request is to spend a known cost against an unknown return, and the person asking cannot tell you the return without first spending the cost. The task is not hard. It is simply priced out of existence.
That is the shape of a lot of genuinely valuable work. It does not require insight, it requires a volume of attention nobody can afford to give.
What the pass found
The window was a complete enrollment cycle, 15 October 2025 through 10 July 2026. 268 days. Roughly 110 records a day, every day, including the ones nobody was working.
29,435 policies verified against their actual status at the source, then matched back to the advance already paid on them. 2 books, 2 payment models, 1 instrument reading both. 10,291 had charged back. That is 34.96% of the book, which rounds to the number worth remembering, about 1 in 3.
Per record, a producer advance in this book typically runs 60 to 100 dollars. Some are materially higher. Some are as low as 20, depending on where the producer sits on the ladder.
You can multiply those together. We do not publish the product, and the reason matters more than the number would.
The number we do not publish
2 things stop a headline figure from being honest here.
The first is that some share of those 10,291 would have surfaced on their own. Chargebacks are not invisible forever. Some get caught at the next reconciliation, some get caught when a producer disputes a balance, some get caught by accident. A recovery claim assumes none of them would have been found without us, and that assumption is false. We cannot say by how much, so we do not build a number on top of it.
The second is that the per record value is a distribution, not a constant. A range of 20 to 100 with an unstated shape is not something to take an average of and present as a total. Ladder position drives it, and the ladder is not uniform across a book.
So the claim is exposure identified, not money saved. 10,291 advances that should not have stood, found and flagged, across 29,435 verified. That is a statement that survives someone checking it, which is the only kind worth making to a counterparty who would rather you were wrong.
Precision over recall, on purpose
Here is the part that actually shaped the build.
The output of this pass is not a report. It is a claim against somebody. That changes which errors you can afford.
A missed chargeback costs 1 record. Call it 60 to 100 dollars, and it stays findable later. A wrong chargeback costs something that does not have a dollar figure, which is the credibility of the entire list. An instrument that produces false claims gets ignored after the 2nd one. Once the counterparty has found 2 things on your list that are not true, every other line on it becomes negotiable, including the true ones.
The errors are not symmetric, so the system should not treat them as though they are. It is tuned to abstain.
That is why the reading works in small confirmed steps rather than 1 pass over the whole screen. Find the record. Crop to the table. Isolate the row. Then read each field. Every step confirms what the last one found before the next one runs. When a step cannot confirm, the record is flagged for human review with the reason attached, and it is never guessed at. Every verdict carries its own trail back to what was read and where.
The effect is that uncertainty leaves the system as a question rather than as a wrong answer. Records that could not be confirmed become a review queue with a stated cause, which is a cost, and an honest one. A system that guessed at those would have produced a longer list and a weaker one.
This is the same instinct as declining to publish a recovery total. In both cases the temptation is a bigger number, and in both cases the bigger number is the one that loses the argument in the room where it matters.
2 failures, 2 instruments
Worth separating, because the figures get read as one thing and they are not.
This pass finds money that went out and should not have. An advance was paid, the policy died, the advance stood anyway. The direction is outward.
The commission ledger behind the same book finds the opposite failure, money that should have come in and never did. A policy survived, the payment was owed, and no statement ever mentioned it. The direction is inward, and it is a harder problem, because a reconciliation fed by the counterparty can never detect something the counterparty simply left out.
Different failures, different instruments, different numbers. They are not 2 views of the same money. The mechanics of the reading pass are written up on the OCR Policy Checker entry in the archive, and the omission side has its own instrument and its own post.
What transfers
The pattern is not insurance. It is any place where a machine readable document from a counterparty has to be held against a ledger you keep, one record at a time, at a volume that defeats manual review.
Carrier statements. Remittance advice. Explanations of benefit. Bank confirmations. Credential and license status. Court dockets. Anywhere the answer already exists in a system you are allowed to read, but does not exist in the system you actually keep.
The work was never impossible. It was priced at 490 hours, and nothing that costs 490 hours of undirected attention gets done on the chance it pays.