Financial controllers are the control because the books close on their signature, not the AI agent's. An agent can match transactions, prepare reconciliations and draft accruals and journals, but the controller decides what is posted, what is investigated and what the numbers mean. Their training has to match that: less preparing, much more reviewing.
What does a finance agent prepare during the close?
A good deal of the mechanical work. At depth 1 the agent reads the general ledger, bank feeds, sub-ledgers and supporting documents. At depth 2 it drafts: bank reconciliations with unmatched items listed, accrual journals based on open purchase orders, prepayment releases, intercompany matching and variance commentary for the management pack. A person approves and posts.
Some routine postings may later be certified at depth 3, executed within bounds the business sets, such as standard monthly prepayment releases under a fixed value. Anything that moves money, such as releasing a payment run, is depth 4, and most finance teams keep it held for a named approver.
Why does segregation of duties need rethinking?
Finance controls rest on separating the person who prepares a journal from the person who approves it. When an agent prepares the journal, who is the preparer? If the answer is "nobody", the controller who approves it is now both preparer and approver in substance, because nobody else has looked at it.
A trained controller makes this explicit. Either a named accountant owns the agent's drafts as preparer and reviews them before they reach the controller, or the controller's review is designed to be a full preparer-level check, with a second approver above certain values. The record shows which model was used for each journal. What matters is that someone can say who checked what, and the record lets them prove it.
Where should a financial controller look hardest?
Agent-prepared close work is usually accurate on the routine and weakest at the edges. The controller's attention belongs at those edges:
| Area | What the agent does well | Where the controller looks |
|---|---|---|
| Bank reconciliation | Matches high-volume items quickly | Old unmatched items cleared by small adjustments |
| Accruals | Accrues against open orders | Services received with no order raised |
| Supplier invoices | Codes to the usual account | Invoices arriving by email, which are external claims |
| Variance commentary | Explains movements from the data | Explanations that sound right but are not true |
| Intercompany | Finds matching pairs | Differences parked in suspense across months |
The supplier invoice row deserves emphasis. Under the Pryme Context Protocol an invoice that arrived by email is external, and an external claim cannot authorise a payment. A trained controller knows that an agent-coded invoice is still a claim from outside until it is matched to an order and a receipt.
What should a financial controller never hand over?
Judgement on estimates, materiality and anything that will be read by auditors or a board. Provisions, impairments, revenue cut-off on unusual contracts and the final view on whether the numbers are right all stay with the controller. So does any decision that cannot be reversed cleanly once posted and reported. The list is set out in what financial controllers should never let an AI agent do.
There is also a quieter risk. Variance commentary drafted by an agent reads well, and controllers under time pressure may adopt it wholesale. If the commentary is wrong, the controller has put their name to an explanation they did not test. Training should treat narrative as seriously as numbers.
The close timetable makes this harder. Agents compress the early days of the close, which tempts finance teams to shorten the timetable. If review time is squeezed in the process, the controller ends up approving more agent work in less time, on the days when errors are most expensive. A trained controller protects review time first and shortens the close only once the record shows reviews are catching what they should.
What does a financial controller need to learn?
Depth and the minimum chain come first: why the agent may draft a journal but not post it, and how the lowest of five bounds (role scope, certificate, sources, the business's own limits, platform guardrails) decides each action. Then verdicts: what held, refused and withheld mean, and that withheld actions are a platform floor no setting can change. Then certificates: they lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day, which during close week means a sudden rise in drafts to approve.
Finally, controllers need to read the record as an auditor would. Every verdict and every approval is filed in an append-only record, and the external auditor will sample it. Compare the neighbouring roles in credit controllers as the real control and the role map in AI agent skills by role. The longer-term question is covered in will AI agents replace financial controllers, and the brief on financial controllers in logistics in India shows the role in a high-volume, multi-entity setting.
Where the Academy fits
Pryme Intelligence Academy's Operator track suits controllers who review agent work daily: it covers human review, reading the workspace, when something goes wrong, and usage and billing. Certified Pryme Operator needs a written exam and a practicum of real review decisions, signed off by a Pryme assessor. A finance function can make journal approval rights depend on a current certificate.
Questions people ask
Can an agent post journals during the close?
It can if that capability is certified at depth 3 for those journal types and the business's own limits allow it. Many controllers start with drafts only and allow execution for a narrow set of recurring journals once the evidence supports it.
Will auditors accept agent-prepared reconciliations?
They will look at who reviewed them and what that review covered. A reconciliation prepared by an agent and approved by a named person, with the record intact, is auditable; one approved without review is not.
How does a controller know the agent used the right data?
Every fact the agent used carries its source. The controller can see whether a figure came from the ledger, a bank feed or an emailed document and decide how far to trust it.