Month-end close on autopilot: what an agent can actually do

Month-end close automation means running the checks a controller would run at period end every night instead of once a month, against live data, so that on day one of the close most exceptions have already been found and cleared. A person still signs the close. What changes is how much of it is discovery.

That distinction is the whole design. A close is not one task. It is a stack of tie-outs, and each one fails in its own way: a bank account that does not agree with the ledger, a payout that arrived net of fees, an intercompany balance where one side booked and the other did not, a period someone reopened and edited after sign-off. Software that claims to “do the close” without naming which of those it checks is selling a calendar, not a control.

We run this on our own books across three operating companies. The system is a set of tomtes, our word for one production agent with one defined job, and this group’s job is to reach period end with nothing unexplained.

What month-end close automation actually covers

Six jobs. Get the source documents in. Reconcile each bank and card account to the ledger. Break settlements from payment channels and marketplaces back into gross revenue, fees, and tax. Tie subledgers to the general ledger. Run the exception checks: intercompany, cheques in transit, aged receivables and payables, documentary trail, period locks. Then produce the pack a reviewer signs.

Most tools sold as close automation do the first job and the last one. The middle four are where a close goes wrong, and they are the ones that need to run continuously rather than in a scramble on day three.

Nobody can tell you how long your close should take

Ask how many days a good close takes and you get 6.4 days, quoted on page after page as a current benchmark. Trace it. The APQC figure behind that number appears in a CFO.com article from March 2018. The data is real. It is not 2026 data, and almost nobody republishing it checked.

The most recent defensible trend line is worse than the marketing suggests. Ventana Research, now part of ISG, found that 58% of companies completed the monthly close within six business days in 2023, down from 60% in 2019. Quarterly close performance fell over the same period, from 49% to 44%.

So treat any article offering close duration by industry to one decimal place as invented. If a vendor cannot say where a benchmark was measured and in what year, your own close time last quarter is the only baseline you have. Measure that first. It is also the only number your board will accept afterwards.

What the vendors shipped, and what they are claiming

The category is real and moving. Gartner predicted in February 2026 that finance teams using cloud ERP applications with embedded AI assistants could see a 30% faster financial close by 2028, and that 62% of cloud ERP spending will be AI-enabled by 2027, up from 14% in 2024.

Adoption itself has flattened. A Gartner survey published in November 2025, fielded across 183 CFOs and senior finance leaders, found 59% of finance functions using AI in 2025 against 58% in 2024, after a jump from 37% in 2023.

Two named products landed in the close itself. SAP announced an Accounting Accruals Agent in January 2026 that generates journal entries from historical data and accounting policy, claiming up to 80% less manual calculation effort and up to 50% less review and posting effort. Workday announced a Financial Close Agent in September 2025 alongside a Financial Audit Agent credited with saving early-access customers up to 900 hours a year.

Read those two claims as what they are. Vendor-reported, no published methodology, no sample size. They tell you where the market is going, not what your close will do.

The pressure is real even where the benchmarks are not

Intent is running well ahead of capability. Deloitte’s Q4 2025 CFO Signals survey, fielded to 200 CFOs at North American companies with at least $1 billion in revenue and published in January 2026, found 87% expect AI to be extremely or very important to finance operations in 2026, and 54% naming the integration of AI agents into the finance department as a transformation priority.

Then the gap. AICPA and CIMA surveyed 1,446 senior finance and accounting leaders in December 2025: 88% called AI the most transformative trend of the next 12 to 24 months, while only 8% felt their organisation was very well prepared to manage it. Half named a shortage of skills and talent as the biggest barrier.

That shortage is measurable. The AICPA’s 2025 trends data, reported in October 2025, counted new CPA exam candidates falling from 42,626 in 2023 to 28,082 in 2024. The U.S. Bureau of Labor Statistics projects about 115,300 openings for accountants and auditors each year from 2025 to 2035. Fewer people, same deadline. Something has to absorb the routine part.

How ours runs, between the closes

The order matters more than the model.

1. Documents land in a monitored finance intake, tagged to a company and an account. Statements that arrive as photographs or screenshots are read, not rejected.
2. Every extracted line is checked for continuity. The running balance has to chain from the previous line. A break holds the whole document rather than loading the readable part.
3. A content fingerprint is computed per line, so the same statement uploaded twice cannot record anything twice.
4. Accepted lines load into a standard bank book per account and into the ledger, through write endpoints that are idempotent. A rerun cannot double-post.
5. Anything the agent cannot place goes to a review queue with a stated reason: unknown account, continuity break, or low confidence. It is never written to the books.
6. Storefront and marketplace payouts are classified by the payment channel string the platform actually emits, then matched to the deposits that landed, with fees split out.
7. On a schedule, the exception battery runs against the ledger: trial balance, aged receivables and payables, intercompany, cheques in transit, documentary trail, and period lock violations.
8. The output is a period-end pack. The statements, the checks that passed, and every break with the evidence attached.

The loop runs several times a day, not once a month. By period end the queue is short because it has been worked down daily, which is the actual mechanism behind a faster close. Nothing was compressed. It was just done earlier.

Spreadsheet checklist, ERP close module, or an agent

	Spreadsheet close checklist	ERP or accounting-software close module	Purpose-built agent
When the checks run	Once, during the close	Once, during the close	Continuously, between closes
Reads a scanned or photographed statement	By retyping	No	Yes
Catches a re-uploaded statement	If someone notices	Sometimes	Yes, by content fingerprint
Both sides of an intercompany balance	Manually	Within one entity	Across entities
Behaviour when unsure	Person guesses	Line falls to suspense	Holds, states the reason, writes nothing
Explains a gap that looks like a bug	No	No	Yes, that is the output
Who owns it	You	The vendor	You, on your own infrastructure

The refusals are the product

An example from our own books. One payment channel showed zero transactions in the reconciled data, which looked exactly like a classification failure. It was not. That channel’s last order predated the window the ledger sync covers, and everything before it had been posted by hand under the old process. Zero was the correct answer.

A weaker system would have widened the window to make the gap disappear, dragging thousands of pre-cutover orders into the posting queue. The agent’s job was to explain the gap and stop. Changing a posting scope is a finance decision, not an automation one.

The same principle runs through our ad-spend truth engine, where a platform-claimed 4.87x return came out at 0.33 to 0.50 against the actual ledger, figures as of July 2026. A number nobody reconciled is not a number.

What stays with a person

Judgement, and it is a shorter list than people expect. Accrual and provision amounts where policy needs applying to an unusual case. Anything routed to manual allocation because it did not sum exactly. Internal transfers between company accounts, always, because guessing there invents intercompany balances that then have to be unwound. Reopening a locked period. And the sign-off itself.

Everything above that line is retrieval, matching, arithmetic, and checking, which is where the hours go. This is the same split we described for bank statement processing, and the same one behind automated business reporting: the machine assembles the evidence, the person makes the call.

What it costs to install

A working version on your accounts, your document sources, and your accounting system is an AI Workforce Sprint: from $9,500, four to six weeks, built on your infrastructure, documented, handed to your team at the end. Fixed price and fixed scope. If you want the checks and the current break list mapped before committing to a build, the $1,900 audit does that first. Both prices are public, as of July 2026.

You own the result. The code sits in your repository, the ledger data in your database, and if we disappeared tomorrow your controller still has the keys.

Send us the list of accounts you reconcile at period end, the accounting system they land in, and the format your statements arrive in. You get a written fixed-price scope, a delivery window, and the exact list of conditions under which the agent will hold rather than post, within one business day. No call, ever. Start async.