News

Pankaj Prasoon is Sr Director Finance Tech & Strategy at LinkedIn, leading finance technology and building AI fluency across finance teams.

gettyFinance has always known how to hand power to actors it can’t watch in real time. It just never wrote that power down in a form software can follow.

At 2:13 a.m., an agent flags a supplier invoice that looks like a duplicate. The payment run leaves at 8. The agent has access to do three things. Message the invoice owner. Place the payment on hold. Open an investigation and pull the PO history itself. Technically, all three are open. Ask around the next morning, and nobody can tell you which one it was allowed to do.

The first two pieces in this series were about economics: what we ask people to give up, and what the benefit is worth. This one is about permission, which is the wall you hit the moment any of it starts working.

Ask the same question about a person, and you’d have an answer in seconds. We have approval matrices, segregation of duties, materiality thresholds and a board-approved delegation of authority that someone updates every year even if nobody enjoys updating it.

That document exists in every company of scale. It’s also typically a PDF, written in prose, for humans, with sentences like “the appropriate level of management.” No agent can “read” it, and no system enforces it against a nonhuman actor. That, not model accuracy, is what’s capping autonomy in finance right now.

Being right isn’t the same as being allowed.Finance has always kept two questions apart that most AI programs merge immediately: “Is this correct?” and “Is this actor permitted to act on it?”

A senior accountant can be completely right about a reclassification and still lack authority to post it. That’s by design, not a challenge on competence. We’ve never treated correctness as permission for people, and I don’t know why we’d start for software. In practice we sort of have, because a confident output feels like a sanctioned one.

An agent that drafts a journal entry is a productivity tool. An agent that posts it is inside your control environment, and your auditor will treat it that way whether or not you’ve decided to. The difference isn’t sophistication. It’s whether power was delegated, by someone with standing to delegate.

Rewrite the delegation so a machine can read it.What finance teams need isn’t another guardrail workshop. It’s the delegation of authority they already have, rewritten into something an agent can parse and a control can test against.

The academic work frames autonomy from the human’s seat: operator, collaborator, approver and observer. Useful instinct, wrong vantage point. The rung that gets encoded and audited is the agent’s: observe, recommend, prepare, approve and execute.

The artifact is dull, and about a page. For every agent you run, you should be able to name:

• Identifier

• Business owner, by name

• Control owner, by name

• Transaction types it may touch

• Autonomy level on each

• Monetary ceiling

• Reversibility class

• What forces escalation

• What evidence it leaves behind

If you can’t produce that page for an agent already running in your environment, you’ve handed something capability without authority, which is close to the only thing finance has learned not to do.

Not all authority should be the same size.Scale it against four things: materiality, reversibility, how well understood the action is and whether anyone outside the company ever sees it.

A small internal reclass that’s reversible, routine and invisible externally can carry a lot of autonomy. An irreversible outbound payment, a disclosure input, anything touching a customer or a regulator gets almost none, no matter how good the model looked in testing. Agents just force us to say out loud what we already do for people, in a format that compiles.

The control has to get there first.Most of our controls are detective. They inspect what happened after it happened, and they work because there’s a human-speed gap between mistake and damage. A misposting sits overnight. A payment file waits for the run.

Agents close that gap. A control that catches the problem on Thursday can be perfectly designed and useless.

So the control moves in front of the action. What data may the agent use? Which version of the policy is in force? What threshold may it cross? When is review mandatory, and when does the workflow stop?

The evidence changes, too, and that’s the part people underestimate. Knowing which user clicked approve no longer tells you much. You need to reconstruct what the agent knew, which rule it applied, what authority it held and why the action sat inside it. A prompt log won’t get you there. That’s a decision record, and someone will eventually ask to sample it.

Which raises a question few have answered. If a model produced the estimate and an agent posted the entry, who’s attesting to the control environment around it? Management still signs, and that signature doesn’t become conditional because the actor was software.

Name two people, not a principle.Governance conversations love to land on “A human remains accountable.” Fine. Which human?

The process owner, the model owner, the controller, the executive who set the threshold or whoever accepted the exception at quarter-end? When everyone is accountable in principle, nobody is accountable in operation. We’ve all watched that movie.

Every material financial agent needs a named business owner for the outcome and a named control owner for the authority framework. Technology builds and runs the platform. It shouldn’t inherit a financial decision because the actor is software.

Here’s the question that comes before the pilot.For years the question was whether a system could automate a control. The better one now is what authority we’re willing to hand software, asked before the demo gets impressive and before a workflow quietly slides from helping into doing.

Defining decision rights isn’t a brake. It’s what lets you scale, because teams know exactly where machines can act and what must exist afterward.

A CFO can’t delegate accountability to an agent. A CFO can decide, with real precision, what the agent gets to decide.

The information provided here is not investment, tax or financial advice. You should consult with a licensed professional for advice concerning your specific situation.​

Forbes Finance Council is an invitation-only organization for executives in successful accounting, financial planning and wealth management firms. Do I qualify?