AI in Finance Operations: What the Agent Executes and What the Controller Keeps

The finance department is the best starting ground for AI agents, and not because it is fashionable: because of its structure. The rules are written, the volume is high, and the cost of an error is known with precision. When a process meets those three conditions, an agent can execute it reliably and leave a trace of every decision. This article walks through the five areas where I see traction earliest and, in each, separates what the agent executes from what the controller keeps.

The underlying distinction is the same one I apply to any function: I am talking about agents that operate — they touch the ERP, post entries, move receivables — not an assistant that suggests. That difference and its requirements are covered in the hub essay, AI agents in operations. For the full catalog by function, see enterprise AI agents use cases.

Why finance is the ideal ground

Three conditions make it the cleanest candidate.

Explicit rules. Much of accounting is already codified: the chart of accounts, expense policy, agreed payment terms. An agent cannot apply a rule nobody has written, and in finance almost all of them are written.

Volume. An agent’s return lives on repetition. Reconciling hundreds of lines or chasing dozens of due dates every month is exactly the kind of patterned work where a person adds little judgment and a lot of time.

Known cost of error. In finance you know what a mistake costs: an imbalance can be located, a misapplied payment can be traced. That lets you set the blast radius before switching anything on, and widen permissions only where the data supports it.

Reconciliation

The most mature case. The agent matches invoices against orders and delivery notes, and bank movements against the receivables and payables ledger. It pairs what matches unambiguously, flags what does not, and prepares the discrepancy with its context for a person to resolve.

• The agent executes: bulk matching, detecting unmatched items, proposing the entry, documenting each match.
• The controller keeps: discrepancies that need judgment — a difference justified by a pending credit note, a partial payment agreed by phone — and the call on doubtful items.

The invoice-by-invoice detail, with two-way and three-way matching, is developed in AI invoice processing.

Financial close

The close is a calendar-driven process with many repeatable steps and a few that require judgment. The agent takes the former: it balances accounts, checks that figures make sense against the prior period, flags deviations, and documents the reason for each, so the team reaches review with the mechanical work already done.

• The agent executes: balancing, consistency checks, flagging deviations, preparing supporting documentation.
• The controller keeps: estimates and provisions that depend on judgment, the interpretation of a real deviation, and signing off the close.

A faster close is not the goal in itself. The goal is for people to spend their time on what requires judgment rather than on balancing sheets.

Collections

Chasing due dates is work of persistence, not talent, which is why it gets done late and poorly when the team is stretched. The agent tracks the overdue book, sends staged reminders according to each customer’s policy, and prioritizes where to spend human effort.

• The agent executes: tracking due dates, staged reminders, prioritizing the book by amount and risk, preparing the context for the hard call.
• The controller keeps: the conversation with the important customer, negotiating an extension, and deciding when to escalate to legal.

Here a typical business rule — do not chase a certain strategic customer until after a given date — has to be written down and connected to the agent’s context. If it lives only in the CFO’s head, the agent will not apply it.

Expense control

The agent checks each expense against current policy, catches what falls outside the norm, and routes exceptions to whoever should approve. The compliant bulk passes without friction. The doubtful gets stopped.

• The agent executes: verification against policy, detecting duplicates and out-of-threshold amounts, preparing the approval.
• The controller keeps: approving the justified exception and adjusting the policy itself when reality outgrows it.

Reporting

The agent gathers data from the systems, reconciles the figures, assembles the recurring report, and flags what deviates from expectations. It prepares the material. It does not interpret the business.

• The agent executes: collection, reconciliation, assembling the periodic report, alerting on deviations.
• The controller keeps: reading the numbers, the narrative for the board, and the decision made from the report.

The line between the agent and the controller

If I had to sum up where the boundary falls across the whole function, I would put it this way: the agent executes what is ruled, verifiable and repeatable; the controller decides what requires judgment, context or accountability. The agent does not replace the finance team. It removes the mechanical work that today stops them from doing the work that adds value.

For that boundary to hold in production you need three layers. Permissions bound what the agent may read and write in the ERP and up to what amount it acts without human approval, and I detail this in AI agent permissions and controls. Escalation decides which cases go up to a person and with what context, which in finance is a first-order risk decision. And audit: every entry, every collection, every approval is traced, because without a trace no serious board should approve an agent touching the books. How that gets audited is in auditing AI agent decisions.

Before an agent gets real work, you measure how it performs: cases with known answers, quality thresholds, failure review, a method I describe in AI agent evaluation. And when the proposal reaches the board, what sustains the yes is the return, and I cover it in AI operations ROI.

Frequently Asked Questions

Which finance processes should be automated with AI first?

Reconciliation and invoice processing are usually the best first candidates: written rules, high volume, verifiable results. Expense control and collections follow easily. The close and reporting benefit greatly, but it is better to enter through their most mechanical stretches before the parts that require judgment.

Does AI replace the finance department?

No. It executes the ruled, repeatable work — balancing, matching, chasing — and leaves the team what requires judgment: interpreting a deviation, negotiating with a customer, signing the close. The controller moves from doing mechanical work to supervising a system that does it, with more time for what adds value.

How is the risk of an agent touching the books controlled?

With three layers: permissions that bound what it can write and up to what amount, escalation to a person for cases that exceed its mandate, and an auditable trace of every action. The risk is decided before switching anything on. An agent with read-and-propose permission cannot unbalance anything. Write permissions widen only with evidence.

Do I need a modern ERP to use agents in finance?

Not necessarily. The agent works against whatever systems you have, through integrations with concrete permissions, and in many companies that includes old accounting software. What you do need is for the process rules to be written and the inputs and outputs to be verifiable. That weighs more than the age of the ERP.