Who Signs Off When AI Prepares the Numbers?

Take a company that will never appear in an AI lab’s case study. It runs on a ledger that has been patched more often than anyone remembers. Its auditor arrives every spring, and its lender wants a signed covenant certificate each quarter. This year, an AI agent drafted the bank reconciliations, the accrual schedules and the first pass of the close, and the only AI audit trail is whatever the vendor decided to keep.

The signature did not move. The same people sign the same documents on the same dates, with the same exposure if a figure is wrong.

In my view, few finance AI products are designed around that fact. They are built for the person who uses them. Whether they are kept or cut is decided by the person who signs.

Who still signs

It helps to be precise about which signatures.

• The officers of a US-listed company. An SEC rule from 2002 requires the principal executive and financial officers to certify every quarterly and annual report. They may not have someone sign for them under a power of attorney.
• Management, wherever the books are audited. Under the international standard, ISA 580, management confirms in writing to the auditor that it has fulfilled its responsibility for preparing and presenting the financial statements.
• The borrower and the taxpayer. The compliance certificate to the lender and the tax return carry a name too.

The auditor has not been let off either. When the UK’s Financial Reporting Council issued guidance on generative and agentic AI in audit this March, it stated plainly that “the human auditor is always accountable.”

So I think of the signer as the real buyer of finance AI. The analyst who saves an afternoon is the user. The controller or CFO whose name goes at the bottom of the page decides whether the tool stays.

A number nobody can trace counts as wrong

After insightsoftware acquired CXO Software in 2018, I stayed on and later ran an integration team of about 40 people across finance, product and sales. Deferred revenue came across from a billing system the acquiring finance team had never used. The figure was right. It could not be traced. So it was rebuilt, contract by contract, over two weeks before the CFO signed.

In twenty-five years on the software side of finance, I have watched one rule hold, and AI has not changed it. A signer does not ask whether a number is probably right. A signer asks whether they can defend it. A number they cannot trace gets treated as wrong, however right it turns out to be.

Finance is strict about this for a reason. One error in a polished set of statements makes every other figure suspect, because the reviewer no longer knows which ones to trust. Fixing the error is cheap. Re-checking everything around it is what costs the days.

That, I think, is why first experiences with finance AI so often disappointed. Prep time did not go away. It moved into verification time: hours of checking, done by the most expensive person in the chain, closer to the deadline.

A number the signer doubts gets made twice

Distrust has a cost no demo shows. When the signer does not trust the output, the team does the rational thing and keeps the old process running beside the new one. The AI drafts, people redo the work, and someone compares the two. Cost goes up, because the team pays for both. Reported impact is zero, because nothing the AI produced was allowed near the work that matters.

A product that fails this way fails quietly. Usage looks healthy and the users are happy, yet no hour has left the close. Sooner or later that spend meets the payback question, and a tool stuck in a parallel run has nothing to show.

For investors, I think this changes the measure that counts. Time to an answer is easy to demo and tells you little. Watch time to a trusted number instead: how long it takes to get from raw data to a figure someone will sign. A product that answers in seconds and still needs a day of checking is slower than it looks.

Evidence earns the signature

In my experience, earlier finance AI tools failed the close in two ways. Some summarized: they read the files and wrote commentary that did not tie back to the numbers. Others fabricated: they explained variances with a confidence they had not earned. Both produced fluent output a controller could not use.

What earns a signature is evidence that travels with the number. Glenn Hopper, a former CFO who writes and speaks about AI in finance, reduces it to one rule: no number without a source. The reason is economic. A reviewer can check a cited figure in seconds, and an uncited one only by redoing it.

The same logic makes an honest gap worth more than a clean answer. An unreconciled difference shown as unreconciled is information. Buried in a plug, it becomes one of the most dangerous things in a close: a visible problem made invisible. Tracing a gap is a different job from absorbing it. Double, which builds AI into accounting firms’ close, traces the cause of a reconciliation difference and points to the transactions that need attention.

For a CFO, controls are the floor everything stands on. They cannot wait for a later release.

The market has not caught up. Avalara surveyed 1,505 senior finance leaders, CFOs among them, and published the results in July:

• 92% felt career pressure to prove their AI agents deliver a return.
• 44% were only somewhat confident they could explain an agent’s actions to an auditor or regulator.
• 23% said accountability for a significant agent error would be unclear or sit with no one.
• Only 28% required documented audit logs showing how a vendor’s agent reaches its decisions.

The pressure to show a return is nearly universal. The evidence a signer would need is not.

The audit trail outlasts the model

It is tempting to think the next model release fixes this. I don’t think it will. A 2025 paper on why language models hallucinate argues that they do so because training and evaluation reward guessing over acknowledging uncertainty. A better model guesses less often. It does not tell the signer which answers were guesses.

An AI audit trail belongs to the system around the model: which ledger it read, which version of which file, which rule produced the figure, and who approved it. That record does not get smarter when the model does, and it does not need to. It holds through every upgrade.

I expect that is where durable value in finance software will sit. Models will keep getting better and cheaper for everyone, so a product whose only edge is the model it calls will be matched by the next release. A product that carries the evidence from source to sign-off holds what the signer cannot do without.

Autonomy will go as far as a signer can defend it

I expect finance leaders to hand AI more of the close and the filings at the pace their signers can defend the result, not the pace of the demos. A CFO who cannot explain an agent’s work to the auditor will keep a person in front of it, and should. That person has to know what a wrong number looks like, and AI is taking over the work where finance used to learn it.

Builders who design for the reviewer and the auditor from the first release should find audit, procurement and renewal easier. Those who treat trust as a later feature will keep landing customers they cannot keep.

AI does the work. Finance owns the result.

Wouter Born is the founder and General Partner of Born Capital. He writes AI CFO Office, a newsletter on AI for CFOs and finance leaders with more than 50,000 subscribers.