Sarbanes-Oxley did not reduce the number of signatures in a company. It increased them. Sections 302 and 404 took an assertion that had been institutional and attached it to two named individuals with personal consequences.
That is the pattern every wave of automation has followed. The clerks go. The signatories multiply. Nobody has yet built the technology that removes the requirement for a person to stand behind a claim, and it is worth understanding why, because the current wave is larger than the ones before it and the arithmetic is about to get uncomfortable.
The stack
A finance function is not one job. It is five, stapled together by an accident of history. The person who could reconcile the ledger became the person who signed for it, and the title covers both.
Five jobs, one title. Four of them compress.
Bar length is what remains of each job once agentic systems have taken what they can.
Four of those compress. One does not, and the reason is not technical.
The floor
An agent cannot be punished.
That is the whole constraint, and it is worth stating plainly, not dressed in governance language. Attestation is not a computation. It is somebody standing behind a claim with their name on it, their position at stake, and in some jurisdictions their liberty. The value of a signature comes entirely from what happens to the signatory when it turns out to be wrong.
A system with no exposure cannot supply that, however accurate it is. Accuracy and accountability are different goods, and only one of them is available for purchase from a model provider.
So the CFO is not going anywhere. The department around the CFO is a different question entirely.
The concentration
Here is the part that does not get said in the vendor deck.
The analyst layer, the controller layer, the reconciliation layer, all of the people who currently sit between the signature and the number, are the compressible part. They are also the people who historically noticed when something was wrong. Not through governance. Through familiarity. Somebody had seen last month’s file and this one looked odd.
Remove them and the signature does not get smaller. It gets larger, faster and lonelier.
That is a worse job than the one that exists today, and the people being handed it can feel it coming even where they cannot articulate it. Treating this as a headcount story misses what is actually happening. It is a redistribution of exposure onto a smaller number of named individuals.
The reconstruction problem
There is a second-order effect that has barely been named.
If a period’s numbers were produced by an agent stack, and that stack is retrained, replaced or reconfigured, can the period be reproduced? Can anybody say which version of which model, working from which data definitions, generated the assertion the officer signed?
Ask a company to re-derive one prior close from source, on today’s semantics. Most cannot. That gap is currently invisible because nobody asks, and it is a live liability sitting under a signature that has already been given.
Finance is only where it is legible first. Any agentic system producing assertions that a human signs for has the same exposure, and industrial systems produce a great many of them.
The precedent
None of this is new. It is only new to finance.
Manufacturing has been signing for machine-produced output since the war. A calibration certificate is a named person asserting that an instrument reads true, traceable through an unbroken chain to a national standard, and it expires. First article inspection under AS9102 is a named engineer asserting that a process, not a part, has been proven. Production part approval in automotive is the same discipline in a different dialect. Nadcap accreditation puts a body behind a process the customer will never watch running.
Every one of those has the four properties that finance is about to need. A named signatory. Evidence rather than confidence. Traceability back to something outside the organisation. An expiry date, because capability decays.
The instinct in the current debate is to reach for explainability and model cards. Manufacturing’s answer was cheaper and better. Do not try to understand the machine. Qualify the process, sample the output, hold a named person accountable, and make the certificate expire.
The absent discipline
The organisations that will hold together over the next decade are the ones that treat attestation as an engineered capability rather than a compliance afterthought. Right now almost nobody owns it. Audit tests it after the fact. Legal papers it. Technology functions build the thing that produces the claims and are not asked to prove the claims are reproducible.
Manufacturing engineers qualify processes rather than inspecting outputs. They build traceability chains that survive the departure of the people who set them up. They design gates that can be held or lowered but never quietly raised. They put an expiry date on a capability claim and make somebody re-earn it.
Kaipability works at this interface. As agentic systems take over the production of assertions, the scarce good becomes the evidence that lets a named person sign for them, and that is an engineering problem long before it is a governance one.
If somebody in the organisation is about to sign for output they cannot reconstruct, that is the conversation.
Start a conversation What we do about it →Questions this dispatch answers.
Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.
