A human review step is only a control if there is an explicit standard it checks against. Without one, review is either ceremony or unaffordable.
Earlier this year an accounting firm ran a test designed to fail.
Propeller Industries took a real client's month-end reconciliation and rebuilt the books as they had looked before the close: every bill entered, all the operational activity present, nothing tidied up in advance. As the answer key they used the close their own controllers had actually booked. Then they let an AI agent work the pre-close data and graded what came back against what the humans had really done. The person who ran it, Mabela Zenullari, was explicit that the setup was not built to show the agents could work. It was built to find out how they would fail.
The output was not bad. That was not the finding.
The finding was that checking the agent's work took about as long as doing the close by hand.
Why the checking cost so much
It would be easy to read that as verification simply being expensive, and to conclude that supervision has a price which sometimes exceeds the benefit. That is not what Propeller found, and the real reason is more useful.
The agents improvised. Each reconciliation was worked out afresh, the agent finding its own route through, usually a reasonable route but its own. So every result had to be validated from scratch against a person's judgment rather than against a standard. Zenullari's description of the outcome is exact: not automation, a second set of books to review.
Then the part that matters most. When the output was off, they could not tell whether the agent had failed or whether it had correctly executed a workflow that had simply never been written down. Without a written standard, every error is ambiguous. You cannot debug the tool and you cannot debug the process, because you cannot separate them.
An undocumented process does not only make review expensive. It makes failure uninterpretable, which means the organisation cannot even learn from it.
Propeller's response was not to abandon the automation or to accept the cost. It was to go and write the standard. Controllers were each assigned a set of general ledger accounts and given the job of documenting the exact workflow end to end for each one: what inputs are required, what steps are taken, in what order. Not a redesign of the reconciliation, and not a new template. A written version of what was already being done in people's heads. Their claim about the payoff is specific: once a reconciliation exists as a written standard, an agent's work can be checked against that standard in minutes rather than against someone's judgment in hours.
The second term
Review is a comparison. Something is checked against something else. In most organisations only the first term is named, and the second lives in the reviewer's head, which is fine for as long as the reviewer has time to consult it.
Name the second term or you do not have a control. You have a person looking at output.
That is the variable running through everything below. Three organisations, three very different outcomes, and in each case the thing that determined which outcome they got was whether an explicit standard existed for the reviewer to check against.
When the standard makes the reviewer unnecessary
Glide, a software company, published an account in August of rebuilding its engineering process around agents. The part relevant here is what happened to code review.
Pull request review had been one of the more valuable parts of an engineer's day. As agent-generated code increased in volume it became a queue to clear, and then to clear again the next day. The reason it collapsed is not the one you would expect. The code was good. It was good enough, and consistent enough, that review stopped being a judgment and became an approval stamp. Engineers asked to reduce their review commitments, on the grounds that the code needed less scrutiny rather than more.
Glide did not respond by asking people to read faster, and it did not respond by leaving the step in place. It spent months tightening linters, prompts, constraints, automated checks and evals. It replaced a polyrepo with a monorepo so an agent could see the whole system, moved infrastructure so agents could provision it, made feature flags the default so that shipping and exposing became two separate decisions, and moved to a merge queue. Their own framing of that work is the line to hold onto: none of those changes was the system. They were the foundations that had to be true before the system could exist.
Only then did humans come out of code review. Ninety per cent of changes are now agent-reviewed with no human involved.
Two judgments were deliberately kept. At the gate after a feature is built, a person reviews the feature and asks whether it does what was asked. After an automated repair is proposed, a person reviews the code and asks whether the fix is safe to ship. Neither is line-by-line reading. The system holds the code-quality bar; the humans hold intent and consequence.
That is not a human removed from the loop. It is a human moved up a level, made possible because something else now checks the thing they used to check.
It is worth being precise about what this evidence is. Glide concluded that the code-level review step had stopped carrying judgment, and acted on that conclusion. They publish the ninety per cent, the deployment frequency and the cycle times. They do not publish a defect rate, an escaped-bug count, or incident data for the changes that shipped without a human reading them. The judgment that the control had become unnecessary is theirs, made confidently, and unmeasured.
Directional fallback
One detail in Glide's design deserves separating out, because it is the most transferable idea in this article.
Rolling a change forward always ends with a person. Rolling one back does not. The rollout watcher that monitors every deployment alerts someone when it detects degradation, and then reverts automatically if there is no response.
So the same supervisor, watching the same system, has opposite defaults depending on which way the change is moving. On advancing state, silence blocks: no answer means the feature waits. On reversing state, silence permits: no answer means it rolls back. The organisation has decided that being wrong in one direction is cheaper than being wrong in the other, and has set the fallback accordingly.
That is a genuinely different way to think about a supervisory boundary. Most designs ask what a human must approve. This one asks what should happen when the human does not answer, and gives a different answer for each direction.
The limit is equally important. Silence can only be allowed to authorise reversal where reversal is genuinely the safer state. A feature flag flips back. A payroll disbursement does not, a closed account does not, a discharged patient does not, and deleted data does not. A control you can skip was about a safeguard that had no opinion about the situation it was in. Directional fallback is the opposite move, and it works precisely because someone worked out what reversal costs in this specific domain. Lifting the pattern into a domain where undo is not cheap would turn a good control into a bad one.
When the signature is the only evidence
Now the case where the second term was never named.
Cigna's PXDX system flagged claims for denial in batches. Physicians signed them off. An investigation by ProPublica and The Capitol Forum, working from internal spreadsheets that tracked how fast the doctors worked, found that in two months Cigna doctors refused payment on 300,000 claims through the system at an average of 1.2 seconds each.
The control was formally intact. Every denial required a named physician's sign-off, and every denial got one.
What the sign-off did not do was evidence that an independent standard had actually been applied. Nothing in the design obliged the physician to check anything, and nothing recorded whether they had. No minimum review time. No sampling audit. No friction of any kind between the algorithm's output and the signature. A human step existed, and the system captured its occurrence rather than its substance.
Cigna disputes the characterisation. It says the denials were refusals to pay rather than denials of care, and that the system was built to process claims quickly rather than to reject them. The 1.2-second figure comes from investigative reporting, not from the company.
A ceremony produces no signal
Here is the structural problem with a review step that has stopped working: it does not report its own condition. It produces approvals, exactly as it did when it was working. From inside the system there is nothing to see.
So the evidence came from outside, and only from the fraction of cases where somebody pushed back. In May 2023 the chair of the House Energy and Commerce Committee wrote to Cigna citing figures from a different population altogether: prior-authorisation denials under the company's Medicare Advantage plans, not the claims processed through PXDX. About one in five of those denials are appealed, and roughly eighty per cent of the appealed ones are overturned. She was careful about the inference, framing it conditionally and asking whether the automated review process was implicated rather than asserting that it was.
Read the other way round, four in five of those denials were never appealed at all. Whatever the review was worth in the cases nobody contested, nothing tested it. The committee's point was that the overturn rate was worth asking about — and it could only ask because a minority of people had appealed.
Compare how the other two organisations found out. Glide's engineers said it out loud, asking for less review because the step had stopped meaning anything. Propeller ran a test specifically designed to locate the failure, and published it. At Cigna it took reporters working from internal spreadsheets to make it visible at all, and a congressional committee only began asking questions once they had published.
Detection is not incidental to this. A control that has quietly become ceremonial is invisible to the organisation running it, which means somebody has to go looking, on purpose, before an outsider does.
What the three cases share
None of these is an independently audited account. Two are companies publishing about themselves in venues that also sell their services. The third is investigative reporting and a congressional letter, contested by the company involved. That is the evidence available, and it is worth saying so plainly rather than treating self-description as fact.
What they share is the variable. Propeller had no written standard, so review cost as much as doing the work and failures could not be diagnosed. At Cigna the sign-off recorded that a physician had signed, and nothing more, so the step could not show whether any standard had been applied. Glide built a standard first — checks, evals, tests, and an architecture that made them enforceable — and only then decided which parts of the review still needed a person.
The question is not whether a human is in the loop. It is whether the human step still carries judgment. If it does not, an organisation has three honest options: remove it and say so, build the standard that would make it meaningful again, or stop delegating until it can.
Leaving it in place is not one of them, though it is the most common choice, because a control that has quietly stopped working looks exactly like one that works.
A review with nothing to check against is not a control. It is a signature.
Sources
- Propeller Industries, Can an AI Agent Run Your Month-end Accounting Close?, August 2026
- Glide, Creating Slipstream: How we rebuilt our software engineering process for AI agents, Simon Fletcher, 6 August 2026
- ProPublica and The Capitol Forum, reporting on Cigna's PXDX review process, March and May 2023
- US House Committee on Energy and Commerce, letter to The Cigna Group, 16 May 2023
The Propeller and Glide accounts are published by the organisations that ran the systems described, in both cases alongside commercial material; no independent verification is claimed for either. The Cigna figures come from investigative reporting and from a congressional letter; Cigna disputes the characterisation and no finding has been made against it.