Future CFO talks series: When AI does the work, who signs the books?

At Payhawk’s Future CFO Talks in Munich, a finance director described an invoice process that was already close to running end to end. Invoices were captured at source, coding was suggested automatically, approvals followed policy, and approved data moved into the ERP without someone manually carrying it from one system to another.

One human checkpoint remained. Before release, a bookkeeper still reviewed the postings and tax codes.

Then came the question from the stage: how much longer would that final review survive?

The answer, in effect, was: for now.

That phrase captures the uncertainty inside almost every finance function experimenting with AI. The human checkpoint is moving, and most teams can see that it will continue to move, but far fewer have made a deliberate decision about where it should eventually stop.

Most of the debate still focuses on capability. Can AI extract the data, code the invoice, identify an anomaly or recommend a decision accurately enough? Those questions matter, but they address only half the problem. The harder question is one of authority: what should an AI system be allowed to do without asking a person each time, and when it does act, whose name stands behind the decision?

The line is moving before anyone has drawn it

Finance teams are not removing human involvement in one dramatic leap. It is happening through hundreds of smaller choices.

First, people stop entering the data. Then they stop matching it. Straightforward transactions no longer need manual coding, and low-value approvals begin to disappear. Each decision appears reasonable in isolation because the machine is faster, the process is cleaner, and the audit trail may even be better.

The combined effect, however, is more consequential than any single decision. The point at which a human must intervene keeps retreating, often without anyone stepping back to decide where the final boundary should sit.

That is how accountability gets redesigned by accident.

Ask a finance leader where AI already touches the organisation’s workflows and the answer will often be incomplete, not because anyone has been careless, but because no individual deployment ever seemed large enough to demand a complete map. What emerges is an accountability perimeter drawn by accumulation rather than intent.

“For now” is what that looks like when spoken aloud.

Three convenient answers, none of them sufficient

When organisations do try to decide where human involvement should remain, they usually fall back on some combination of three tests: accuracy, compliance and instinct.

The accuracy test says that a person should remain wherever AI is not reliable enough. This sounds sensible, but it confuses performance with responsibility. Accuracy can indicate whether a system is capable of doing the work; it cannot determine whether the organisation should delegate the decision.

A model may be correct 99.9% of the time, yet the remaining fraction still matters if the failures involve duplicate payments, incorrect tax treatment or transactions that cannot easily be reversed. More importantly, a signature has never meant that the signer is the most accurate component in the process. It means someone is answerable for the result.

The compliance test says that a human should remain wherever an auditor or regulator requires one. Regulation matters, but it establishes the minimum boundary rather than designing the entire operating model. A required checkpoint can become the weakest control in the process when one person is expected to review hundreds of machine-generated decisions without the time, information or practical authority to challenge them.

That is not meaningful oversight. It is review theatre, because the presence of a person creates the appearance of control without providing much of the substance.

The instinct test is simpler: keep the human wherever the decision feels too important to automate. It is probably the most common approach in practice, and also the least defensible, because “feels important” usually tracks familiarity rather than exposure. Teams preserve human involvement in work they understand and automate the work they find repetitive or tedious, even though some of the largest finance risks are buried inside ordinary decisions repeated at scale.

The boundary should therefore be drawn by materiality, reversibility and the consequences of failure, not by habit or discomfort.

Finance already has the beginnings of an answer

The useful reframe is that this is not really a debate about tasks. It is a debate about delegated authority, and finance has dealt with that problem for decades.

Every serious finance function has some form of delegation-of-authority matrix. It defines who may approve what, up to which value, under which conditions and with which exceptions. Nobody expects the CFO to approve every transaction personally; accountability scales because authority is bounded, documented and escalated.

AI does not need an entirely new philosophy. It needs a place inside the one finance already uses.

That does not mean giving software legal responsibility or pretending that a machine can own a decision. It means defining what the system may do, under which conditions it may do it, and the point at which the decision must return to a person.

The human signature does not disappear. It moves.

Instead of signing each individual output, people increasingly sign the operating rules: which decisions the system can make, which thresholds apply, which exceptions must escalate, what evidence must be retained, and who has the authority to intervene.

This is a more scalable form of control, but it is also a heavier one. Signing off one payment means being accountable for one payment. Signing off the rules under which a system releases thousands of payments means being accountable for the perimeter around all of them.

The number of manual signatures may fall, while the weight behind the remaining ones rises.

The Munich panel was already living this model

The most interesting part of the discussion was that the practitioners on stage had already started drawing these boundaries, even if they did not describe them in those terms.

A treasury specialist at a global manufacturer explained that AI could flag anomalies, such as an unusually large invoice, inconsistent master data or another pattern that did not fit expectations. Yet no material payment left the company without human release.

The boundary was not based on what the technology could technically do. It was based on exposure.

The finance director described the other side of the same model. Her company had allowed broader experimentation with AI tools, but use inside the finance function was held back until legal guardrails and acceptable hosting arrangements were in place.

That distinction matters because experimentation and delegation are not the same thing. Companies need room to test what AI is good for before every idea is buried under months of process, but once a system begins touching sensitive data, influencing financial decisions or taking action at scale, the standard must change.

The principle is neither “govern everything before you begin” nor “move fast and add controls later.” It is more disciplined than either extreme:

Experiment broadly. Delegate selectively. Govern according to exposure.

“Human in the loop” is the wrong end goal

The phrase “human in the loop” has become a comfort blanket in enterprise AI because it reassures boards and regulators that a person is still involved somewhere. Yet involvement is not the same as control, and control is not the same as accountability.

A person can review an output without having the authority to reject it. They can approve an action without understanding how the recommendation was produced. They can nominally own the process while technology, the vendor and the business each assume that someone else is responsible when it fails.

That gap is the real risk.

Finance may assume that technology owns the model, while technology assumes that finance owns the process. The business may assume that the vendor owns the output, while the vendor points to its terms and conditions. Everyone participated, but nobody is clearly answerable.

The better question is not whether a human remains in every loop. It is whether a named person can explain why the system was allowed to act, under which conditions, how its decisions can be reconstructed, and what happens when it gets something wrong.

That is what control looks like when humans stop touching every transaction.

So where does the last signature sit?

The bookkeeper’s final review may eventually disappear in its current form. Coding and tax checks will improve, exceptions will become easier to identify, and routine postings will require less manual attention.

The books do not become ownerless when the task disappears.

A mature finance function will not keep a person on every transaction, but it will keep a person answerable for every system. The last signature therefore does not sit permanently on the invoice, payment or journal entry. It moves one level higher, onto the operating model itself.

What may the AI do? Where must it stop? Which decisions are too material or irreversible to delegate? What evidence must it retain? Who can override it? And whose name stands behind those choices?
AI can take the work and it can be given authority, but it cannot take accountability.

That part still needs a name.