What should financial controllers never let an AI agent do?

Nothing that changes the books beyond routine or speaks for them. Financial controllers should never let an agent approve journals, post large manual adjustments, reopen a closed period, set estimates or provisions, change the chart of accounts or posting rules, sign off the close, answer auditors, or release payments outside tight limits. Each needs a named person.

Which lines protect the integrity of the ledger?

Approving journals, its own or another agent’s. The preparer must not approve. An agent may draft and, within strict limits, post recurring standard entries. Any held journal is approved by a person. If an agent could approve, segregation of duties would disappear.

Posting manual or adjusting journals above threshold. Manual adjustments are where errors and manipulation hide. The agent can prepare the entry with support attached; the controller or a delegate approves.

Reopening a closed period. Reopening changes numbers that may already have been reported. It needs a person’s decision, a reason and usually a second approval.

Changing the chart of accounts or posting rules. These are the rules every other entry follows, including the agent’s own. An agent that could change them could change its own bounds. Changes go through a governed process.

Which lines protect judgement?

Setting estimates, provisions or impairments. Bad debt reserves, inventory provisions, impairment tests and accruals based on estimates are judgements. The agent can calculate a starting point from policy and data. The controller decides the number and documents why.

Signing off the close or the reporting pack. Sign-off is a statement that the numbers are right, made by someone who can be held to it.

Which lines protect what the business says to others?

Answering auditors or giving representations. An agent can assemble evidence and draft responses to audit requests. What the business tells its auditors, and any representation about the completeness of information, comes from a person.

Releasing payments outside tight limits. Releasing money is depth 4, hard to undo. Even where an agent is certified to pay small, routine amounts to verified payees, payments outside those limits are held. Bank details changed on an emailed request are an external claim, and an external claim cannot authorise a payment.

Why do these hold however good the agent is?

Because the controller answers for the numbers, and these acts define what the numbers are. A governed agent works within five bounds at once: its role scope, its certificate, its sources, the firm’s own limits and the platform guardrails. The red lines live in the firm’s limits, so they are refused or held at any depth. The agent’s depth is also temporary: certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day.

What does a crossed line look like in a close?

Take a financial controller at a logistics company. On day three of the close, a held item proposes a manual journal releasing part of a bad debt provision, with a note that two large customers have paid. The agent drafted it from the sealed cash receipts and the provision policy. It is above the posting threshold, so it is held.

The controller checks. One customer paid; the other’s “payment” is a remittance advice emailed by the customer, external, with no matching bank line. The release would overstate profit. The controller approves only the first part, notes why, and asks the agent owner to make sure provision releases never rest on external evidence.

The agent did not break a rule. It drafted from what it was given. The controller held the line because an estimate is a judgement.

How should a finance team keep these lines?

Write them into the firm’s own limits so the agent refuses or holds each one. Separate the person who sets the agent’s posting limits from those who approve its held journals. Sample the record each close for manual journals and provision movements. Put certificate dates in the close calendar. A structured plan is in the AI agent training plan template for financial controllers.

Related lists cover credit controllers and, for judgement-heavy roles in insurance, underwriters. Market practice differs; compare financial controller training in India with Singapore. The overview is in AI agent skills by role.

Where the Academy fits

Pryme Intelligence Academy’s Operator track certifies the people who approve agent work, with a practicum of real review decisions that must show no missed red line. A finance team can adopt the practice of tying journal approval rights to a current certificate, so authority lapses when competence is not renewed.