Gartner research based on interviews with more than 150 corporate controllers and chief accounting officers found that robotic process automation can spare a finance department with 40 accountants 25,000 hours of avoidable rework a year. Similarly, McKinsey estimates that currently demonstrated technologies can fully automate 42 percent of finance activities and mostly automate another 19 percent.

That automation has been enormously useful. Rules inside ERPs, close platforms, and other accounting systems now handle work that used to consume hours every month: recurring entries, standard allocations, cash coding, matching, and other predictable tasks.

But there is a reason so much of finance is only mostly automated. A long tail of accounting work does not behave predictably enough for fixed rules. It contains exceptions, changing inputs, incomplete evidence, and situations where the right answer depends on context.

Take an accrual rule that posts $40,000 for a contractor every month as long as the purchase order remains open. This month, the project was paused two weeks ago and the vendor already told your AP team. The PO is still open, so the accrual posts anyway. Expense is overstated, and someone has to clean it up. You could add another rule to check project status. Then another to account for vendor communications. The problem is that someone has to anticipate each circumstance and tell the system what to do before it happens.

AI agents are built to handle intricate situations like these. They can learn, adapt, and make context-aware decisions, as opposed to blindly following “if-this-then-that” rules. They’re more akin to seasoned accounting professionals who can understand the subtleties of each situation and respond appropriately. 

Where rules fall short

Rules remain the best way to automate routine, high-volume accounting work. They are fast, inexpensive to run, predictable, and easy for an auditor to test. We build them ourselves and use them wherever the workflow allows. The problems start when the work stops arriving exactly as expected.

Rules struggle with exceptions

Say a payroll report lists Erica B. while the HRIS census lists Erica Bennett. The department, salary, and start date all match. An accountant looks at the surrounding evidence and resolves the discrepancy in seconds. An exact-match rule sees two different strings. Unless someone anticipated that particular discrepancy and configured a way to handle it, the record ends up in an exception queue or falls out of the calculation.

Accounting is full of these situations:

• Vendors change names

• Files arrive with different columns

• Descriptions vary slightly

• Supporting documents come in different formats

What looks like an edge case in a software specification is ordinary work during a close.

Exceptions bring the work back to accountants

When a rule cannot resolve something, the work usually does not disappear. It goes back to an accountant.

Someone has to notice the exception, find the relevant source data, understand what happened, decide how to resolve it, and finish the work manually. And because many of these exceptions surface during close, they arrive precisely when the team has the least time to investigate them.

Rules can dramatically reduce transaction volume while leaving a disproportionate amount of the difficult work behind. That is how a workflow can look automated on a dashboard while the accounting team is still working late to close the books.

Change requires reconfiguration

Rules also depend on the world continuing to look roughly like it did when they were configured. A vendor changes its invoice format. A payment processor rearranges its export columns. A bank statement arrives with a different structure. Suddenly a workflow that ran perfectly last month needs attention.

AI is making it much easier to build and modify these rules. Some systems now let an accountant describe a workflow in plain English and generate the underlying configuration or script automatically. We use similar techniques in our own deterministic engine. While that improves setup considerably, it still does not change how the automation executes. If AI generates a rule and a fixed script runs afterward, the system still depends on logic defined in advance.

AI agents: the next evolution of automation

AI agents are the next evolution of AI-powered automation, providing a more intelligent solution that understands the context behind each situation. It works from close procedures written in English, much like the instructions you would give a staff accountant, and works through the evidence it encounters as it performs the procedure.

Agents can use the surrounding evidence

While preparing a multi-currency cash workbook, Max encountered 5,707 rows where the currency field contained both “USD” and “USD” with a trailing space. An exact-match rule sees two values. Max recognized that they represented the same currency, normalized the data, continued the workflow, and recorded what it changed.

The payroll example works the same way. Erica B. and Erica Bennett do not need a bespoke matching rule if the surrounding evidence is strong enough to establish that they are the same employee.

Agents can work through variation

Ordinary changes in the inputs also do not have to trigger a rebuild. Service periods can be extracted across different invoice layouts. Bank statements do not have to arrive in precisely the same structure every month. If the accounting procedure changes, the instructions can be updated in English and the next run follows the revised procedure.

The workflow can absorb more of the variation that accounting teams normally handle themselves.

Agents can ask when the evidence isn’t enough

Sometimes there is no responsible way to resolve an exception automatically. Suppose Max, our purpose-built accounting agent is preparing the August accrual and finds a September invoice that appears to cover services delivered partly in August. The evidence suggests an answer but does not support one confidently.

Max will ask the preparer. The accountant receives the invoice, the relevant source rows, and the specific question that needs answering. Their response becomes part of the record, and the workflow continues.

The exception does not disappear. What changes is how much work the accountant has to do to resolve it. And instead of starting an investigation from scratch, they provide the judgment the workflow actually needs. Those answers and reviewer overrides can also feed back into the procedure so the system does not have to ask the same question every month.

The accountant stays in control

Giving agents more responsibility for preparing accounting work does not change the controls around it. The accountant responsible for a balance still needs to understand how the number was produced. A reviewer needs to be able to re-perform the work, changes to the underlying logic need to be visible, and the evidence supporting a conclusion needs to remain attached to it. These are ordinary accounting controls, regardless of whether the preparer is a person or agents.

Any accounting agent, including ours, should therefore be explainable to the accountant who owns the workflow, accurate on the company’s actual data, and auditable down to the evidence and logic behind its work.

We built Max around those requirements. Accountants review every agent-prepared output. A drafted journal entry remains proposed, balanced, and validation-checked until a named reviewer approves it for posting. Preparer and reviewer remain separate, and the audit trail runs from transaction-level source data through the work performed to the ledger.

There are also accounting judgments that should remain entirely human. Goodwill impairment, M&A accounting, and other conclusions a company has to defend to auditors and investors are not unfinished automation problems. They belong with the people accountable for making them. In workflows that depend on those judgments, Max brings the accountant into the process rather than attempting to make the conclusion itself.

A more automated close

Rules are not going away. They shouldn’t.

For predictable accounting work, deterministic automation remains faster, cheaper, and easier to control. If a workflow is stable and the answer can be reliably expressed as a formula or condition, there is little value in asking a model to spend minutes reasoning through a calculation a deterministic system can perform correctly in seconds.

Agents become useful when the work contains enough variation that every situation cannot reasonably be specified beforehand. Their opportunity lies in what rules leave behind: almost-matches, malformed inputs, changing source data, ambiguous evidence, and questions nobody anticipated when the workflow was configured. These exceptions may represent a minority of transactions, but they consume a disproportionate amount of the human effort left in an automated close.

Forward-thinking accounting teams like Rippling, Zendesk, Cognition and Miro are already moving beyond the bounds of rule-based automation. Let’s chat about how Maxima can help you join them.