Ask whether an AI agent can run the month-end close and you get two kinds of answer. Vendors say yes. Controllers who have cleaned up after a bad automation say no. Both are answering the wrong question, because the close is a dozen steps with different owners, and an agent belongs in some of them and nowhere near others.
The split I use is simple. An agent gathers, matches, drafts and chases. A person approves anything that changes the ledger or puts a number in front of management. Hold that line and the preparation work can move to the agent while the sign-off stays exactly where your auditor expects to find it.
Here is how that plays out, step by step.
Bank reconciliation
This is the usual first step for an agent. The bank statement arrives as a file, the ledger holds the open items, and the rule for a clean match is clear.
The agent pulls the statement, matches each line against open items on amount, date, reference and counterparty, and posts nothing. It hands over a list of proposed matches with the evidence for each, plus the lines it could not place. A person reviews the doubtful ones and approves the batch before anything is booked.
The hard part is the residue. Near-identical amounts, partial payments with a deduction, FX rounding and one transfer covering six invoices all produce matches that balance and are wrong. I have written up the patterns where automated matching goes wrong separately. Expect to meet each of them in your first few bank recs.
Intercompany reconciliation
If you run more than one entity, intercompany is often the step that holds the close open. Entity A booked a recharge on the 28th, entity B booked it on the 3rd, and the balances disagree by an amount nobody recognises.
An agent is good at the legwork. It pairs the transactions from both sides and labels each difference by cause. A timing gap, a different FX rate, a missing counter-booking, a booking on the wrong account. Then it drafts the correcting entry for the side that is wrong, and stops there. The person who owns that entity reads the explanation, agrees or disagrees, and posts.
The value is in the explanation. A reviewer who reads "difference of 4,312.50, invoice IC-0917 booked in entity B on 3 October at the closing rate instead of the transaction rate" can decide in a minute. A reviewer who reads "difference of 4,312.50" starts digging.
Invoice matching
Three-way matching of purchase order, goods receipt and supplier invoice is repetitive and rule-bound, which makes it a good agent job. The agent reads the invoice, finds the order and the receipt, checks quantity and price against your tolerance, and either queues the invoice for approval or sends it to the buyer with the specific mismatch named.
Payment approval stays with a person. So does any change of bank details and any invoice above your approval limit. New suppliers go to a person as well. Those limits do not belong in the agent's instructions. A model told never to accept a changed IBAN will follow that most of the time, and the exception will be the run with the forged letterhead. Put the limit in the connector code, where the model cannot reach it.
Accruals
Accruals mix arithmetic with judgement, and the agent should take only the arithmetic. It can list goods received without an invoice and recurring costs whose invoice tends to arrive after month end. It also picks up last month's accruals that need reversing. For each one it proposes an amount and shows where the number came from.
It does not decide that a disputed contract needs a provision, and it does not estimate a bonus accrual from a conversation it never heard. Those are calls a named person makes. The agent drafts the journal. A person reviews it and posts it.
Variance commentary
Management wants to know why travel costs are over budget this month. The answer sits in forty ledger lines. Pulling them together and writing a first explanation is tedious work, and agents do it well.
The draft stays a draft. The agent sees the bookings but does not know the trade fair moved from November to October, or that a supplier contract was renegotiated. The controller adds what the ledger cannot show and owns every sentence that goes into the pack. Let the agent's text go out unread and sooner or later a variance gets explained with the wrong reason, and the CFO repeats it in a board meeting.
The close checklist
Every close runs on a checklist, usually a spreadsheet with a few dozen tasks and a column for initials. An agent can keep it current. It checks which reconciliations are done and which inputs are still missing, and it messages the owners before the deadline instead of after.
It never ticks its own work as reviewed. The agent fills the preparer column and a person fills the reviewer column, every time. That keeps the segregation of duties your auditor tests, and every item the agent touched has a human name next to it.
What stays with people
Pulled together, the line looks like this:
- Posting journals, or at minimum approving each batch before it posts
- Judgement calls on provisions, write-offs and disputed items
- Sign-off on every reconciliation the agent prepared
- Releasing payments and changing supplier master data
- Locking the period
- Anything that goes to management, the auditor or the tax office
None of that has to move for the close to get shorter. The hours go into preparing and chasing, and that is the part the agent takes.
Start with one step in shadow mode
Pick one step and run it alongside your team for full close cycles before it writes anything.
Bank reconciliation for a single account is a sensible first step, because the result can be checked line by line and a wrong match surfaces within days. The agent does the whole job on live data with read-only access while your team reconciles the usual way. Then you compare. Shadow mode only proves something if the pass mark is set before the run, so agree up front which matches and which exception list you would accept.
Plan the review side at the same time. Every item the agent is unsure about lands with a person. If that list is longer than the person can clear by the second working day, the close gets slower. Size the review queue for flagged items before go-live, and make sure each item arrives with the agent's reasoning attached.
Run two closes in shadow mode, not one. The first month always has something odd in it, and you want to watch the agent handle an ordinary month too. Then promote the step, keep sampling by hand, and only then pick the second one.
Bring your close checklist to a scoping call
Bring the close checklist you actually use, initials column and all, to a thirty-minute scoping call. We go through it with you, mark which steps an agent could prepare and which stay with your team, and tell you what building the first one would cost. If the answer is that none of it is worth automating yet, you hear that too.