The signature problem
Something is wrong and it is hard to name.
The pilot worked. Everyone agreed it worked. Then the person who ran it moved to another team, and nobody could answer what it had actually been measured against.
The scene
Output arrives faster than anyone can read it. It is good — usually. Some of it is wrong in ways that only show up weeks later. The review step now costs more attention than the production step, and the review leaves nothing behind: no declared bar, no record of what was checked, no name attached to the acceptance.
So when someone senior asks can we show that this was done properly?, the honest answer is that it was done, and it looked fine, and there is no way to demonstrate either.
The reframe, with the evidence
The bottleneck never was generation. It moved to custody — who declared the standard, who checked independently, and where the record is kept. That is measured, not predicted.
- ·Every third webpage created now has some AI-attributable content.
- ·Press releases — a source the public is meant to trust — reached a ~50% chance of containing AI-generated text in June 2026, doubling within a year.
- ·On Deezer, over half of new uploads were fully AI-generated at peak, while AI tracks took 1–3% of streams. Generation got cheap; unverified output became worthless.
Figures reported by Exponential View, “Damming the slop floods”, 28 Sep 2026 — exponentialview.co. Cited as evidence of the problem. They have not seen this standard.
The bottleneck moves
“Coding is largely a solved problem … Code review becomes a bottleneck.”
Jeetu Patel, Chief Product Officer, Cisco — HBR IdeaCast, 1 Oct 2026. Cited as evidence of the problem; Cisco has no relationship with this standard.
The sequence is the argument. Coding gets solved, so review becomes the constraint. Review gets solved, so deciding what to build and what not to build becomes the constraint. Every fix hands the load to a different person — and none of those gates is real until the bar is declared and written down.
This is why waiting for better models does not help. A better model does not remove the constraint; it moves the constraint onto judgement, which is the expensive part.
Four failures of custody
- ✗The bar was never declared. Something was produced and reviewed, but “good” was never written down, so nothing could be checked against it.
- ✗Independence was asserted. The reviewer was the producer, or a model reading its own output. Both were called independent.
- ✗The record was never kept. The review happened in a thread, a meeting, or a tool that has since been replaced.
- ✗The gate could not fail. Five approvals in thirty days, zero refusals. A gate with no failures has not been tested.
What this costs
| reputation | You cannot defend work you cannot describe. The first failure that reaches a customer arrives with no record behind it. |
| control | Every delegation you cannot inspect is authority you have handed over without a receipt. |
| attention | Reviewing unlabelled output is the most expensive reading there is. In one measured month, an operator received 236 same-shape approval requests — each one individually reasonable, collectively a tax. |
| judgement | When the bottleneck is judgement, the cost lands on your most senior people, and it lands there silently. |
What it does not have to cost
The fix is not better agents. It is a declared bar, an independent reading, and a record that survives the tool.