AI Agents in FP&A: Where Should Human Approval Begin?

The AI-agent question I keep coming back to is not whether the agent can do the work.

It is whether I would let it do the work while everyone in Finance is at lunch.

That changes the conversation pretty quickly.

I would happily let an agent monitor actuals, compare them with forecast, notice an expense that looks strange and put it in front of me. I am considerably less relaxed about the same agent deciding the forecast should change, changing it, and then explaining the new number to management as though the matter has been settled.

Those are both “AI in FP&A.” They are not remotely the same risk.

As a mother, this distinction feels familiar. My kids being capable of operating the air fryer and my wanting them to redesign dinner while I am gone are two separate governance questions.

I draw the line at consequence

I do not think Finance needs one grand rule saying agents are autonomous or supervised. I would look at the consequence of the action.

If the agent is gathering information, checking a threshold, finding an anomaly or preparing a first pass, I am comfortable giving it room once the workflow has been tested.

If it is recommending something that changes an official number, I want review. If it can move money, post something material, communicate externally or make a decision that is difficult to reverse, I want a person before execution.

That is a much more useful boundary than arguing about whether the underlying model is “smart enough.” Smart enough for what?

A forecast is where this gets interesting

Suppose an agent notices that enterprise pipeline has weakened and recommends taking $2 million out of next quarter’s revenue forecast.

I want that recommendation. I want the supporting deals, historical conversion, stage movement and whatever else caused the alert. I may even want the agent to rebuild the scenario.

What I do not want is for the system to quietly turn a recommendation into the company’s new expected outcome.

There is business context hiding behind that number. Maybe one large deal moved. Maybe Sales knows something that has not reached the CRM. Maybe the pipeline really is deteriorating and everybody would prefer not to say it yet.

That last one is not exactly rare.

The hard part of FP&A is often deciding which explanation deserves belief.

I am much less protective of the mechanical work

Finance still spends an astonishing amount of time carrying information from one place to another.

Refreshing reports. Checking whether headcount changed. Comparing periods. Finding accounts over threshold. Pulling pipeline. Updating schedules.

I have no sentimental attachment to any of it.

If an agent can do those things reliably, let it. We survived dial-up internet. I do not need to preserve slow manual work for the nostalgia.

The point of automation should be to move human attention toward the exceptions: the customer behaving strangely, the hiring plan that no longer matches revenue, the margin improvement nobody can explain, the cash forecast that suddenly stopped agreeing with the P&L.

The control I would insist on is boring

For material workflows, I want to know what the agent did, what information it used, what it recommended, what the human changed and who approved the final action.

That is not the sexy part of agentic AI.

Finance should be fine with that. We have built an entire profession around being interested in things nobody puts in the movie trailer.

I also want the approval line written before the workflow goes live. Give the agent a narrow lane, watch how it behaves, and expand the lane when the evidence earns it.

The question I would put in the implementation meeting

For every agent action, I would ask: If this happens incorrectly at 2:00 a.m. and nobody notices until morning, how much do I care?

A weird variance flag? Fine. A draft scenario? Fine. A changed management forecast, posted journal entry or payment? I care.

AI agents can take a lot of work out of FP&A. I want that. But autonomy is not the prize. Better finance is.

Sometimes better finance means the agent works alone. Sometimes it means the agent gets everything ready and then waits for an adult.

Sarah’s broader AI agents in Finance article covers the wider autonomy question. I also wrote separately about trusting AI-generated forecasts.