·6 min read·ai-verification · vertical-ai · ai-market

Accuracy is the wrong bar for AI. Verification is the product.

Most AI vendors chase a higher accuracy number. It is the score you get after the game is already over. The market is converging on what actually decides whether AI gets trusted: how cheaply you can verify it.

Contents

At EP, we don't demo. We prove.

While most AI vendors chase a higher accuracy number, we spend the budget on the check, because the trustworthy check, not the generator, is the asset. This week a tax-AI team with more than $17M in the bank arrived at the same conclusion from a completely different direction. They are not the first, and they will not be the last. The whole market is converging on it.

#The number everyone optimizes is the one that matters least

Filed builds AI data entry for tax firms. They pushed their accuracy past 80%, well above the industry baseline they cite, and many of their customers still complained. Same model, same stack, unhappy users. On the AI Engineer World's Fair stage this month, their CTO named why, and it is the cleanest version of the thing we have been saying since day one:

A higher number did not buy trust, because accuracy is measured after the output is produced. The buyer's real cost lands before that, when they still have to decide whether to believe it.

#The work didn't get removed. It changed shape.

Here is the mechanism every AI buyer feels and few vendors name. Automation takes away the easy part, producing a first draft, and hands back the hard part, checking it. If you cannot tell in advance which outputs are wrong, you have to verify nearly all of them. So raising accuracy from 90% to 97% cuts the number of errors, not the number of checks. The verification burden barely moves.

Chat interfaces and citation trails feel like the fix. They are not. A citation is a pointer to work the reader now has to do: open the source, find the passage, confirm it says what the model claims. That is often as much effort as answering the question yourself. The "fix" hands the burden back with extra steps.

This is not a new discovery. It is a 40-year-old result being rediscovered in every vertical at once. In 1983, Lisanne Bainbridge described the "ironies of automation": automating the easy parts leaves the operator the harder residual work, plus the new job of monitoring a system they can no longer outperform, while remaining accountable for its mistakes. Swap "operator" for "your team" and you have described 2026.

#The market is converging on this

The signal is not one company or one field. It is showing up everywhere the same week.

MIT Sloan researchers put it almost verbatim in their 2026 work on the "verification gap": "AI makes it cheap to produce work, but not to judge whether that work is any good." Verification is the scarce capacity, and it does not scale with production.

The evidence is measurable, not just rhetorical. In a controlled 2025 study, METR found experienced developers were about 19% slower when using AI tooling, while believing they were faster. That gap between felt speed and real throughput is the verification tax, paid quietly.

19%
slower: experienced developers using AI tooling in METR's 2025 study, while believing they were faster

The coding world already lived this and moved past it. The fix for early AI that dumped 200 lines to review was not a smarter model. It was better product design: completion inside the editor instead of a separate tab, a plan you approve before code is written, and reusable skills and memory that compound with every use. And the investors funding the current wave of vertical AI, from a16z to Bessemer, keep landing on the same idea: the model is the commodity, and the defensible work is the system built around it.

#What we build instead

We start from a different question. Not "how do we make the model more accurate?" but "how do we make the buyer's check cheap and trustworthy?" Those are not the same project, and the second one is where the value actually lives.

Three things follow from it.

We spend on the verifier, not just the generator. An independent, ground-checking verification step, one that tests the real result rather than asking the model to grade its own homework, is what turns "it looks right" into "it is right." Producers do not get to be their own judges.

We deliver receipts, not demonstrations. A demo proves the tool works on the vendor's chosen example. A receipt proves it worked on your real task, and shows the check that confirms it. One is theater. The other is evidence.

We size the work so you can verify fast. Presenting a result at the right altitude, the summary and the exceptions first, then the detail on demand, is what lets a reviewer stop auditing every line and start deciding. Simon Willison's line captures the stakes: "A computer can never be held accountable." The accountability stays with the human, so our job is to make the check small.

#What to ask your next AI vendor

If you are evaluating AI for real work, the accuracy percentage on the slide is the least useful number in the room. Ask these instead.

How do I verify a given output, and how long does that take? If the honest answer is "read everything it produced," the tool has moved your work, not removed it.

Who owns the check, you or the vendor? If verification is handed back to you through a chat window and a citation trail, the burden did not move, it just grew steps.

Show me the receipts on my tasks, not a demo on yours. The only accuracy number that means anything is the one measured on your work, with the check attached.

The model was the easy part. The verification is the product. That is the bar we build to, and it is the one the market is finally agreeing on.

#Sources

Explore how we prove and improve agent quality.

Keep reading