AI Agents in Finance: What Should CFOs Actually Let Them Do?

There is a meaningful difference between AI helping with Finance and AI doing Finance.

We are crossing it.

Until fairly recently, most of my conversations about AI in Finance sounded something like this:

Can it write the variance commentary?

Can it summarize this report?

Can it help with the forecast?

Can it analyze this file?

Useful things.

But essentially, we were giving AI homework.

Now we’re talking about AI agents that can monitor information, make decisions about what needs to happen next, move through multiple steps and actually execute parts of a workflow.

That’s different.

And if you’re a CFO, Controller or FP&A leader, I think you should stop for a minute before getting too excited about the word autonomous.

Because autonomous sounds wonderful right up until you remember this is Finance.

I don’t even give myself unrestricted access to everything without thinking about it.

So What Is an AI Agent in Finance?

Let’s get the definition out of the way without making you sit through a technology conference.

A normal AI assistant generally waits for you.

You ask something.

It responds.

An AI agent can be given an objective and then work through multiple steps toward accomplishing it.

In Finance, that might mean monitoring actual results, identifying a variance, pulling supporting data, determining likely drivers, preparing commentary and routing something for review.

Or monitoring receivables, identifying overdue accounts, prioritizing collections and initiating part of the follow-up process.

Or watching business drivers and updating a forecast when conditions change.

That’s why people are getting excited.

Deloitte reported that 54% of surveyed CFOs said integrating AI agents into Finance was one of their top finance-transformation priorities for 2026.

McKinsey is already describing AI agents as a way to make continuous financial planning practical at scale, allowing companies to identify risks and evaluate trade-offs faster.

So this isn’t theoretical anymore.

The question is becoming much less:

Will Finance use AI agents?

And much more:

What exactly are we comfortable letting them do?

That’s the conversation I’d want to have.

Preferably before we hand one access to the ERP.

I Have a Very Simple Rule

The more consequential the action, the less interested I am in complete autonomy.

I realize this is not a particularly sexy AI strategy.

Nobody is inviting me onstage at a technology conference to announce:

Sarah’s groundbreaking vision for artificial intelligence is apparently “it depends.”

But it does.

Suppose an agent identifies that Marketing is $430,000 over forecast.

Great.

Flag it.

Investigate it.

Pull the transactions.

Compare the spend with the plan.

Look for timing differences.

Draft the explanation.

Send me what you found.

I’m delighted.

You’ve saved someone a lot of tedious work.

Now suppose the agent decides Marketing is spending too much and blocks the next payment.

We have moved into a different relationship.

I would like a human.

Think About AI Agents Like New Employees

This is probably the easiest way I know to think about them.

Imagine you hired a new FP&A analyst Monday.

Smart person.

Very fast.

Apparently never sleeps.

Can read practically everything in the company.

Has an unsettling enthusiasm for repetitive work.

Excellent hire so far.

Would you give that person unlimited authority on Tuesday?

Probably not.

You’d give them access appropriate to the job.

You’d review their work.

You’d see where they’re strong.

You’d figure out where they make mistakes.

You’d increase responsibility as trust develops.

And you certainly wouldn’t say:

You’ve been here 36 hours. Feel free to move cash.

At least I hope not.

I’d approach AI agents similarly.

Capability isn’t the same thing as authority.

That’s an important distinction.

Start With Read, Not Write

If you’re deciding where to deploy AI agents inside Finance, this is where I’d start.

Let them read before you let them write.

Give an agent permission to:

monitor,

analyze,

compare,

identify,

flag,

recommend,

and prepare.

There’s plenty of value there.

For example, imagine your FP&A agent wakes up every morning—although technically it never slept, which already makes it better suited to Finance than I am—and reviews:

actual revenue,

pipeline,

bookings,

headcount,

expenses,

cash,

customer activity,

and whatever operating metrics matter to your business.

It notices something changed.

Then it tells Finance:

Conversion in the enterprise pipeline has fallen for three consecutive weeks and is now below the assumption in the current revenue forecast.

I want that.

Very much.

What I don’t necessarily want is:

I reduced the revenue forecast by $4.2 million and notified the board.

Let’s work up to that.

The Approval Line Matters

Here’s an exercise I’d do with your team.

Pick an AI-agent use case.

Now draw a line through the workflow.

Above the line, the agent can act independently.

Below the line, a human approves.

Where does your line go?

For variance analysis, maybe the agent can do almost everything until commentary is distributed.

For forecasting, maybe it can generate the baseline but a human approves changes to the management forecast.

For accounts payable, maybe it can match invoices and flag exceptions but not release certain payments.

For cash management?

My line is moving considerably higher.

You don’t need one company-wide answer.

That’s the point.

Autonomy should depend on consequence.

A $14 software reimbursement and a $14 million wire transfer do not need identical AI governance.

This concludes today’s advanced Finance lesson.

Not All Mistakes Cost the Same

Finance already understands this concept.

We build controls around risk.

Some transactions receive more scrutiny because the consequences of being wrong are larger.

AI shouldn’t be different.

Ask two questions:

How likely is the agent to be wrong?

And:

What happens if it is?

That gives you a much more useful way to decide how much autonomy to allow.

Suppose an AI agent drafts monthly variance commentary incorrectly.

Annoying.

A human catches it during review.

Fix it.

Now suppose an AI agent incorrectly changes a supplier’s banking information.

Different Tuesday.

That’s why I’m less interested in broad declarations like:

“We’re moving to autonomous Finance.”

Okay.

Autonomous doing what?

That part matters quite a bit.

I’d Put Finance-Agent Work Into Three Buckets

If I were sitting with your Finance team, I’d get a whiteboard and start sorting.

Bucket 1: Please Take This Away From Us

These are repetitive, reversible, lower-consequence tasks.

Data gathering.

Basic reconciliations.

Report preparation.

Variance identification.

Routine validation.

Document classification.

Initial commentary.

Monitoring.

Anything involving copying something from one place and putting it somewhere else immediately has my attention.

Finance has talented people doing astonishing amounts of digital carrying.

If an agent can reliably remove that work, please proceed.

I will personally hold the door.

Bucket 2: Do the Work, Then Show Me

This is probably where a lot of FP&A work belongs.

Generate the baseline forecast.

Analyze the variance.

Build scenarios.

Identify anomalies.

Prepare the management-reporting draft.

Recommend an accrual.

Flag a working-capital issue.

Now bring it to someone.

The agent does 80% of the mechanical and analytical work.

The human handles context, judgment and approval.

I suspect this will become an extremely common operating model.

And frankly, I’m fine with it.

I have no sentimental attachment to manually updating 11 tabs before I’m allowed to have an opinion.

Bucket 3: Absolutely Not Without a Human

Moving meaningful amounts of money.

Changing material accounting treatments.

Submitting external financial information.

Making material forecast commitments.

Changing compensation.

Approving significant expenditures.

Anything involving regulatory reporting.

Anything that could produce the sentence:

“We need to call the auditors.”

I’m not saying AI will never perform more of these activities.

I’m saying the control environment should get substantially more serious as the consequence increases.

Speed is useful.

So is sleeping at night.

I’ve grown increasingly fond of the second one in my 40s.

“Human in the Loop” Can’t Mean Nobody Actually Looks

This phrase is going to become very popular.

Human in the loop.

Sounds reassuring.

But I want you to check what it means inside your company.

Because there’s a version that looks like this:

AI produces something.

Human receives it.

Human is busy.

Human sees that AI is usually right.

Human clicks approve.

Human continues clicking approve for six months.

Congratulations.

You have technically maintained human oversight.

You have also invented a very expensive button.

This is a behavioral problem, not just a technology problem.

The better AI gets, the more humans may trust it.

And the more they trust it, the less carefully they may review it.

That means your control can’t simply be:

A person approves this.

You need to know what they’re expected to review.

Someone Still Needs to Own the Number

This is especially important in FP&A.

Let’s say an AI agent generates the revenue forecast.

Who owns it?

The agent?

Finance?

Sales?

The CFO?

I would decide that before the forecast meeting.

Because I can already see where this goes.

Revenue misses.

CFO asks why.

Sales says Finance built the forecast.

Finance says AI generated the forecast.

AI presumably has no plans to attend the meeting.

Very convenient for AI.

Somebody human needs to own the assumption.

Technology can produce the number.

It cannot absorb accountability for the decision.

The Audit Trail Is Going to Matter

If an AI agent changes something important, I want to know:

What did it do?

When?

Why?

What information did it use?

What rule or instruction governed the action?

What changed afterward?

Who approved it?

Can we reverse it?

This is not bureaucracy for entertainment.

Although Finance does occasionally enjoy that.

It’s because six months later someone may ask why something happened.

And “the agent did something” isn’t going to be enough.

Current finance guidance around AI agents is already emphasizing governance, auditability and human oversight as prerequisites for higher-impact workflows.

That part of the AI conversation is going to get much bigger.

It should.

Don’t Automate a Process Nobody Understands

Here’s another place I’m going to ruin everyone’s fun.

Before you put an AI agent into a Finance process, ask someone to explain the process.

All of it.

Why does this report exist?

Why is this approval here?

Why does this person touch it?

Why do we reconcile these numbers?

Why is that spreadsheet involved?

Why does Linda email this every Thursday?

If the answer to half of those questions is:

“That’s just how we’ve always done it,”

you are not ready to automate.

You are ready to investigate.

Otherwise, you’re going to give AI the ability to execute a bad process faster and around the clock.

That’s not transformation.

That’s shift work.

Your Data Doesn’t Need to Be Perfect

I don’t want to overcorrect here.

Companies sometimes hear conversations about AI governance and decide they need to spend three years fixing every piece of data before doing anything.

You don’t.

Recent FP&A guidance on agentic AI makes a similar distinction: organizations don’t necessarily need perfect data to begin, but they do need trusted data, consistent definitions and well-understood processes.

That’s much more realistic.

Pick a contained workflow.

Understand it.

Know the data.

Define the permissions.

Define the approval point.

Measure what happens.

Then expand.

You can learn without handing over the building.

What Happens to the FP&A Analyst?

This is the question sitting underneath a lot of these conversations.

If agents can monitor the business, update forecasts, analyze variances, build scenarios and draft reporting, what does the analyst do?

Hopefully, something better.

I don’t think the highest use of an FP&A analyst has ever been moving information between spreadsheets.

We simply needed someone to do it.

If agents remove more of that work, analysts should have more time to:

talk to the business,

understand drivers,

challenge assumptions,

investigate anomalies,

evaluate decisions,

and communicate what matters.

That’s the optimistic version.

There is another version.

Companies automate half the work and decide they need half the people.

We should acknowledge that possibility too.

Technology doesn’t automatically make jobs more strategic.

Organizations make that choice.

If you’re leading Finance, this is where you have some responsibility.

Don’t tell your analysts AI will free them to become strategic business partners and then use the productivity calculation exclusively to reduce headcount.

People have memories.

Some of them even work in Finance.

I Think We’re Going to Need an Agent Org Chart

I’m only half joking.

Imagine Finance five years from now.

CFO.

Controller.

Head of FP&A.

Analysts.

And underneath or beside them:

Close Agent.

Forecast Agent.

Cash Agent.

Reporting Agent.

Collections Agent.

Each with permissions.

Owners.

Escalation rules.

Controls.

Performance expectations.

Suddenly the question isn’t only:

Who reports to whom?

It’s:

Which humans own which agents?

Which agents can talk to which systems?

Which agents can initiate actions?

Where does human approval begin?

Who reviews agent performance?

I can already feel HR trying to figure out whether the Forecast Agent needs a performance review.

Please don’t invite me.

Here’s Where I’d Start Monday Morning

If you’re a CFO or Finance leader thinking about AI agents, don’t start by shopping.

Take one Finance process.

Preferably one everyone hates.

Map it.

Then mark every step:

Read

Analyze

Recommend

Act

Approve

Now ask:

Where could an agent operate safely today?

Where would we require human approval?

What’s the consequence if it makes a mistake?

How would we know?

How would we reverse it?

Who owns the outcome?

If you can’t answer those questions, that’s useful.

You’ve just learned something important before spending money.

I consider that a successful Finance meeting.

They’re rare. Enjoy it.

AI Agents Are Going to Become Coworkers

Not metaphorically forever.

Operationally.

They’ll monitor things.

Prepare things.

Recommend things.

Eventually they’ll execute more things.

Finance shouldn’t be afraid of that.

But we also shouldn’t confuse technological capability with permission.

Those are different decisions.

I want AI doing a lot more Finance work.

I want analysts spending less time gathering information and more time understanding it.

I want forecasts updated faster.

I want risks identified earlier.

I want fewer people manually moving numbers around because we’ve somehow accepted that as a profession.

But when the agent wants the keys?

We’re going to have a conversation first.

I’m a mother of three.

“Because I can” has never been a particularly persuasive argument in my house.

I’m not planning to accept it from the software either.