Rahul Bhatia SAP S/4HANA Cloud Solution Architect, HCL Tech.

getty

The finance technology industry has spent 20 years selling automation and shipping workflows. We digitized the approval. We routed the invoice. We added a rules engine to the ledger. The work still got done, but a person stood at every junction—checking, clicking and reconciling. AI agents are now removing those junctions. Most CFOs haven't yet priced in what that does to their function.

Here's the distinction that matters. Automation runs a script. An agent pursues a goal. Tell a script the screen changed, and it breaks. Tell an agent to close the books, and it pulls the subledgers, chases the variance, drafts the entry, escalates what it can't resolve and writes down why it did what it did. When an account doesn't tie out, it works the problem instead of stopping. That isn't a faster bot. It's a different kind of thing, and it breaks the assumptions our finance stacks were built on.

Where This Is Going

I keep seeing the same three shifts across the organizations moving early.

Automation is moving from the task to the process. Nobody serious is still aiming to "automate the three-way match." They're chasing "run accounts payable" as something an agent owns end to end, with people supervising the exceptions. Order-to-cash, procure-to-pay and record-to-report—these are being rebuilt as agent-run processes with a human exception queue, not human-run processes with a bit of automation sprinkled on.

The architecture is going multi-agent. It isn't one giant brain but a set of narrow ones—a tax agent, a treasury agent and an intercompany agent—with a layer above them that coordinates the work. That looks a lot like how a finance team is already organized, which is exactly why the people running these functions will find it intuitive.

And the system of record is collapsing into the system of action. The ledger used to record what happened after a person decided it. Agents propose the move, make it and book it, and the audit trail falls out as a byproduct instead of getting reconstructed three weeks later.

What It Asks Of Your Architecture

This is the part finance leaders underweight. Agents aren't a feature you switch on inside your ERP. They make demands your stack was never designed to meet.

An agent can only reason over data it can find and understand. That means master data, data quality and consistent definitions stop being back-office hygiene and become the thing that decides whether any of this works. The companies that treated data cleanup as a chore to defer are about to find out it was the whole game.

Controls have to live in the architecture, not in a policy memo. When software moves money on its own, a procedures document is worthless. Spending limits, segregation of duties, approval thresholds and mandatory human checkpoints—these have to be built into the system so the agent can't route around them. That pulls an internal audit into architecture decisions far earlier than anyone is used to.

And you have to be able to explain what the thing did. A CFO signing financial statements has to answer how a number was produced. So, every agent action needs to be logged, explainable and reversible. Explainability is the precondition for you being willing to put your name on the result.

What Happens Over The Next Three Years

The close goes first. It's repetitive, rule-bound and miserable to do by hand, which makes it the obvious target. The continuous close people have talked about for 15 years starts to look real once agents are reconciling around the clock instead of in a quarterly sprint.

The headcount conversation changes from cutting to moving. The jobs that grow are the ones that supervise, design and audit the agents—people who can say what good output looks like catch the agent being confidently wrong and own the judgment calls that never go away. The transactional roles shrink. The supervisory ones expand. Treat this as pure cost-out and you'll lose the exact people you need to make it work.

Governance becomes an advantage instead of a tax. The organizations that built real controls and explainable systems will move faster because they can trust their own automation. The ones bolting agents onto brittle foundations are going to hit a bad filing or a misrouted payment, and it will set them back further than if they had just gone slower.

The gap I'd watch most is between the companies treating this as a tooling upgrade and the ones treating it as an architecture problem. The first group gets incremental gains and a lot of frustration. The second group rebuilds how the work flows and pulls away.

What I'd Do Now

Don't wait for the finished agentic finance platform. It isn't going to show up clean, and the years you spend waiting are years you spend piling up the data and control debt that makes adoption hurt later.

Start on the foundation. Fix the data. Write the controls into the system. Pilot agents on a bounded process where you understand the work cold and the cost of a mistake is small. AI agents aren't going to replace finance. They're going to reorganize it—around oversight instead of execution and around architecture instead of activity. The CFOs who see that now will design what their function becomes. Everyone else will inherit a design somebody made for them.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?