AI for Controllership: What to Prepare and What to Approve

AI for Controllership: What to Prepare and What to Approve

How to split controllership work between preparation tasks an AI agent can draft and approvals that stay with the controller, with logs and limits.

•

Want to apply this in your company?

Book a call with Juan →

In controllership, AI helps most when it prepares the work rather than deciding on its own. An agent can collect data, reconcile entries, draft variance analyses and assemble the close package. The actions that change accounting records, estimates, policy exceptions and payment releases should stay with the controller. The practical split is simple: the agent prepares with a trail, and the person approves with judgment.

Key takeaways

  • Preparation is the agent's territory: collection, reconciliation, draft analysis and organizing evidence.
  • Approval is the person's territory: accounting adjustments, estimates that require judgment, exceptions and payments.
  • Every prepared task should come with a source, a cut-off date and a record of what was read and compared.
  • Without restricted permissions and logs, the agent becomes one more risk point inside the close.
  • The safe path starts with a real workflow, a success metric agreed in advance, and a decision date to expand, adjust or stop.

Why the split starts with the type of decision

Much of controllership work is mechanical: matching a payment to an invoice, classifying a bank statement line, checking whether a period has been closed. Another part requires judgment: deciding whether a provision makes sense, whether a variance has a one-off explanation or points to a recurring problem, or whether a delegation limit exception can be accepted. Language models are good at reading, comparing and summarizing large volumes of material. They are weak at taking responsibility for outcomes, and they should not be asked to.

There is also a specific risk. Invoices, supplier emails and attachments can contain text that tries to make the system do something outside its scope. OWASP lists prompt injection among the main risks for applications built on language models, and the practical recommendation is to treat received content as data to analyze, not as instructions to execute. For anyone running a close, this means the agent should never have the ability to post, pay or change a vendor record simply because a document asked it to.

If you want to understand the difference between an agent that runs bounded tasks and the broader idea of superintelligence, read what an agent does today, without the hype. For controllership, the useful point is different: the agent needs clear functions, not autonomy.

What the agent can prepare

The tasks below are good candidates for drafting, provided the output is reviewed before it becomes a record or a report for the board.

  • Classify bank statement entries and match them against the ledger, flagging differences without correcting them.
  • Build the accounts payable and receivable reconciliation, with exceptions listed separately.
  • Draft the budget versus actual variance analysis, presenting likely causes as hypotheses, not conclusions.
  • Organize electronic invoices (NF-e) and receipts by period, supplier and cost center.
  • Prepare the close checklist, with open items, owners and deadlines for each one.
  • Write a first version of the commentary for the board, with each number linked to its source.

The gain from these tasks is not only speed. It is consistency: the agent applies the same comparison rule to every item and records what it did, which is hard for a person to keep up during a tight close.

What stays with the controller

Some actions require accountability that cannot be delegated to a system. They belong to the person who answers for the result.

  • Approving adjustment entries and provisions, especially when they depend on estimates.
  • Deciding whether a variance is explained by a one-off event or signals a problem that needs action.
  • Authorizing exceptions to purchasing, discount or approval limit policies.
  • Releasing payments and transfers, even when the agent has already assembled the queue.
  • Signing the close and answering to the board and, where one exists, to the board of directors.

There is also a personal data point. Payroll, bank details and customer information that appear in statements require care under the LGPD, Brazil's data protection law. Before connecting any source, the company needs to know which fields are necessary, where the material will be processed and who has access to the agent's records.

How to design approvals without blocking the close

A simple rule helps the team decide case by case. Use the list below as a design guide, not as legal advice.

  • If the action posts an entry, changes a balance or moves money, it requires human approval before execution.
  • If the action only reads, compares or summarizes, the agent may run it, as long as it logs what it consulted.
  • If the result depends on an estimate, the agent presents its assumptions and a range of alternatives, never a single isolated number.
  • If the information comes from an external source, such as a supplier email or a file sent by a third party, it is treated as data to verify.
  • If the decision affects a person, such as blocking a supplier or initiating collection, human review and a right to contest must be built in. This topic is covered in more depth in automated decisions and Article 20 of the LGPD.

In practice, the agent's permissions should follow the principle of least privilege. It should have its own user, with read access only to what it needs and no payment credentials. OWASP's authorization guidance helps check that each role has only what its task requires.

Illustrative scenario: month-end close at a distributor

Illustrative scenario, no real data: a mid-sized distributor closes the month with dozens of bank entries that have no match in the system. Today, the analyst reads the statement line by line. With an agent, the routine changes in a bounded way. The agent gets read-only access to the statement and the ledger, matches entries by amount, date and description, and returns a list with three groups: matched, probable and unexplained. For each item in the probable group, it shows the criteria it used.

The controller reviews the exception list and notices that two payments were posted twice because of an import failure. He approves the reversing entry in the system himself, because the correction changes the record. The agent posts nothing. At the end, the log shows which statements were read, which criteria were applied and which items were flagged, so the board can understand the close without relying on memory.

A four-step plan: map, limit, pilot, transfer

Before any automation, the team needs a plan that fits in one meeting. In how we approach this, the path is to map, limit, pilot and transfer, and the decision to expand comes only afterward.

  • Step one, map the current close: list the tasks, who performs them, how much time they take and at which stages errors tend to appear. Without this map, automation tends to target the wrong task.
  • Step two, limit data and actions: define what the agent can read, what it must never access and which actions are blocked. The preparation material also needs a standard for names, dates and categories. The article on data ready for superintelligence explains what to standardize before starting.
  • Step three, pilot a real workflow: pick one account or routine, such as accounts payable reconciliation, and run the agent alongside the manual process for a few close cycles. Agree in advance on a success metric, such as the share of flagged exceptions the team confirms as correct, and record every case where the agent was wrong.
  • Step four, transfer and decide: document the workflow, train the people who approve, and based on the pilot, decide whether to expand, adjust or stop. Stopping is also a valid decision when the gain does not justify the control it requires.

The goal is for the agent to raise the team's analytical capacity without replacing the responsibility of whoever signs. If the metric chosen measures individual productivity, be careful: it can turn into surveillance. This is covered in how to measure AI productivity without watching your team.

Frequently asked questions

Can AI post an accounting adjustment to the system on its own? Technically, a system can be configured to do this, but it is not advisable in controllership. Adjustments change the record and need an identified owner. The safe path is for the agent to draft the suggested entry with its justification and send it to the controller's approval queue.

Which data should a controllership agent not access? In general, the agent does not need payment credentials, full payroll data or personal information that the task does not require. Minimize the fields, log what is consulted and check where processing happens. When data is sent outside the country, the LGPD requires specific attention, covered in international data transfer with AI.

Where should a small finance team start? Start with one repetitive, read-heavy task, such as reconciling a single account, with read-only access and mandatory review. A small pilot builds the habit of logging, reviewing and measuring, which matters more than the technology chosen. Only after that does it make sense to consider other routines.

How do you know if the agent is making mistakes? Do not rely only on the impression that the output looks good. Compare a sample of reconciliations against manual work, write down the errors and classify their causes, whether badly formatted data, an ambiguous rule or a reading failure. That history is what supports the decision to expand, adjust or stop.

Sources

AI for controllershipAI agentsmonth-end closefinancial controlshuman approval