Job

Can an AI agent close your books at month-end?

Researched November 2026

Can an agent do this job today?

Supervised

Two unrelated organizations run agents on parts of the close in production, both reported by the vendors involved. In a Microsoft customer story, Pamarco says agents automate roughly 40% of its financial close tasks, with human owners responsible for work that needs judgment or review. In a Nominal case study, Team Car Care says agents handle about 70% of inventory reconciliation, with the other 30% flagged for human review and journal entries submitted to the GL team for posting. Product documentation keeps a person in the loop: Microsoft's Account Reconciliation Agent recommends an action that a user accepts or changes, and Puzzle says its close agents propose changes and ask you to confirm. Adjusting entries change the financial statements, and public issuers' officers must personally certify periodic reports, so the level is supervised, not ready.

Abstract illustration of a ring of green segments that is nearly closed, with one amber segment at the end beside the gap

On this page

Can an AI agent do month-end close?

Part of it, under supervision. An AI agent for month-end close can move tasks along your close checklist, prepare reconciliations, flag mismatches and draft variance commentary. A person still approves adjusting entries, handles accruals and estimates, locks the period and signs off the numbers.

Two organizations report agents on the close in production, and both reports come from the vendors involved. In a Microsoft customer story, Pamarco, a manufacturer with about 225 employees, says agents automate roughly 40% of its financial close tasks. In a Nominal case study, Team Car Care, a Jiffy Lube franchisee with nearly 500 locations, says agents handle about 70% of its inventory reconciliation, with the remaining 30% flagged for people to review. Neither company describes the agent signing off the close.

Trust is lower than adoption. Datarails, a finance software vendor, surveyed 270 CFOs and finance leaders at U.S. organizations with over 1,000 employees and $100 million or more in revenue in July 2026. In that survey, only 4% trusted AI with their month-end close. Treat this as a supervised hire: the agent prepares, a person decides.

This page is not financial, tax or accounting advice.

What the agent does and what stays with you

Agent
- Track the close checklist and move tasks whose dependencies are done
Agent with review
- Reconcile bank, card and investment accounts
- Match subledgers to the general ledger and resolve mismatches
- Work through high-volume reconciliations such as inventory
Human
- Calculate accruals, prepaids and other estimates
Agent with review
- Run variance (flux) analysis and draft commentary
- Review the month for discrepancies before closing
- Assemble recurring management and board reports
Human
- Approve and post adjusting journal entries
- Lock the period and sign off the financial statements

10 tasks: 1 agent, 6 agent with review, 3 human

Task	Who does it	Why
Track the close checklist and move tasks whose dependencies are done	Agent	Pamarco's agents check hourly for due tasks with completed dependencies, do the work, mark the task complete and email the results, per Microsoft's customer story.
Reconcile bank, card and investment accounts	Agent with review	Puzzle's close checklist creates one reconciliation task per such account, and its agents propose changes that you confirm.
Match subledgers to the general ledger and resolve mismatches	Agent with review	Microsoft's agent recommends an action, such as creating a journal entry, for voucher amount mismatches, and a user accepts it or picks another; each reconciliation can be undone.
Work through high-volume reconciliations such as inventory	Agent with review	Team Car Care reports agents handling about 70% of inventory reconciliation, with the rest flagged as exceptions a person reviews.
Calculate accruals, prepaids and other estimates	Human	No source reviewed for this page documents an agent doing this in production.
Run variance (flux) analysis and draft commentary	Agent with review	Puzzle's flux task flags accounts that moved more than 5% or $1,000, lists the driving transactions and a confidence level, and does not change the books without approval.
Review the month for discrepancies before closing	Agent with review	Puzzle's accuracy-review task scans records and generates action items for a person to follow up.
Assemble recurring management and board reports	Agent with review	At Pamarco a board sales presentation arrives populated on the sixth working day, and the sales director reviews and adds context.
Approve and post adjusting journal entries	Human	Entries change the statements; at Team Car Care the person supervising the agent submits entries to the GL team for posting.
Lock the period and sign off the financial statements	Human	Sign-off is accountability; for SEC filers, the principal executive and financial officers must personally sign certifications of periodic reports.

The dividing line is the general ledger. The products cited here give a person review or approval actions over what the agent prepares before anything is posted: reconciliations, proposed entries and variance analysis. Our recommendation is to let the agent prepare those with review. Anything that changes the statements or declares them final stays with a person.

The product documentation cited here draws the same line. Microsoft's Account Reconciliation Agent for Dynamics 365 Finance evaluates voucher amount mismatches and recommends an action, such as creating a journal entry. A user can accept it or choose to reverse, link transactions or accept without change, and each reconciliation can be undone. Puzzle says its close agents "are designed to propose changes and ask you to confirm before changing anything, and every action is logged." At Team Car Care, the person supervising the agent builds journal entries where needed and submits them to the GL team for posting.

For public companies, the law names who signs. Under SEC Rule 13a-14 (official text, readable copy at LII), each principal executive and principal financial officer of an issuer filing Forms 10-Q or 10-K under Section 13(a), or persons performing similar functions, must sign a certification, and the rule says it may not be signed on their behalf under a power of attorney or other form of confirming authority. That rule applies to SEC-reporting issuers. If you are private, your lender or auditor may set their own sign-off terms; ask your auditor or counsel which obligations apply to you.

A worked example: one account

This is an illustration, not a customer story. Assume, for this example, a company with one entity and a 5% or $1,000 variance threshold.

On business day 2, the agent reconciles the operating bank account. The ledger shows $214,300 and the bank statement $212,850. The agent matches all but two items: a $1,200 customer deposit recorded in the ledger on the 31st that the bank shows on the 1st, and a $250 bank fee with no ledger entry. It marks the deposit as a timing item and proposes a $250 entry to bank fees, linking the statement line.

Its variance run then flags bank fees: $250 this month against $40 last month. The commentary names the fee. The controller opens the proposal, sees the fee is a one-off wire charge, approves the entry, and asks the bookkeeper to check whether it should be recharged to a customer. The agent did not post anything and did not close the period.

What you need to provide

The close touches nearly every account, so the agent reads financial data across the ledger, plus personal data that sits in payroll, expense and customer records. Give it read access to the general ledger, subledgers and bank feeds, and a user role that can draft entries but cannot post them or close a period.

Write down the rules your team carries in its head: which accounts are in scope, the materiality threshold per account, how you treat timing differences, and which accounts need supporting schedules. Give it the last few closed months as reference. Keep accruals and estimates with your team and hand the agent the finished schedules.

Check data handling. Microsoft, for example, says its generative AI features may require the "Move data across regions" setting, and that Microsoft might update the underlying AI model automatically. Confirm both against your data-residency rules before you switch an agent on.

Pamarco's approach is a useful pattern: it defines agents for specific jobs with appropriate permissions, rather than giving every user an unrestricted assistant.

Job brief template

Use this as the brief for a vendor, an operator or your own team. Replace the bracketed values with your own.

Copy-ready brief

Goal: By business day [5] after month-end, every balance sheet account in scope is reconciled or has a documented exception, and a draft close package with variance commentary is ready for the controller's review.
Deliverable: A reconciliation for each in-scope account with matched items, open items and supporting transactions; a list of proposed journal entries with reasons; variance commentary for accounts above your threshold; and a close status report sent to the controller.
Inputs: Read access to the general ledger, subledgers (AP, AR, fixed assets, inventory) and bank and card feeds
A user role in the accounting system that can draft journal entries but cannot post them or close the period
The close checklist with owners, due dates and dependencies
Reconciliation rules and materiality thresholds per account, in writing
Last 3–6 closed months: reconciliations, adjusting entries and variance notes as reference
Accrual and estimate schedules prepared by your team
Acceptance criteria: Every in-scope account has a reconciliation or a documented exception by business day [5].
On a sample of 10 reconciliations, the controller finds no unexplained difference above $[500].
Every account that moved more than [5]% or $[1,000] month over month has commentary naming the transactions behind the move.
No journal entry is posted and no period is closed by the agent; every proposed entry links to its supporting transactions.
On a 20-entry sample of proposed adjustments, the controller accepts at least [18] without changes.
Systems: accounting system or ERP (for example QuickBooks Online, Xero, NetSuite or Dynamics 365 Finance)
bank and card feeds
close checklist or task tracker
spreadsheet or reporting tool for the close package
Constraints: The agent reads financial data across the ledger, plus personal data in payroll, expense and customer records; limit access to the accounts in scope.
The agent's account cannot post entries, change closed periods or edit the chart of accounts.
Your provider's data storage and location must meet your retention and data-residency rules; Microsoft, for example, requires cross-region data movement to be enabled for its agent's generative AI features.
Your auditor may ask how agent-prepared reconciliations are reviewed; keep the agent's activity log.
Volume: Example: 30–150 balance sheet accounts, one to three legal entities
Cadence: Example: monthly, with daily or weekly reconciliation runs during the month
Human checkpoints: Proposed adjusting entries: the controller or accountant approves each before posting.
Accruals and estimates: prepared and approved by your team.
Reconciliations with open items above threshold: the account owner reviews.
Period lock and sign-off: the controller or CFO, after reviewing the close package.

The acceptance criteria matter more than any vendor's automation rate. A controller checking 10 reconciliations and 20 proposed entries each month gives you your own accuracy number.

Risks and human checkpoints

The agent proposes a correcting entry that hides a real problem, such as a missing invoice or a cut-off error.
How you notice it
Review every proposed entry above threshold against its supporting transactions, and consider the other actions your tool offers; Microsoft's agent lets the user reverse, link transactions or accept without change instead of creating an entry.
Variance commentary sounds right but is built on the wrong data.
How you notice it
Spot-check flagged accounts each month; in Datarails' survey, 65% of CFOs named confident answers based on the wrong data as their most common AI frustration.
Accounts drop out of scope silently, for example when an exception limit is reached and runs are skipped.
How you notice it
Compare the list of reconciled accounts against the trial balance each close; Microsoft's agent can notify you when a run is skipped due to a limit.
Reviewers rubber-stamp agent output because it is usually right, and errors reach the statements.
How you notice it
Track how often reviewers change the agent's proposals; a rate near zero for months is a reason to re-sample, not to stop reviewing.
The audit trail does not show who approved what, and the auditor cannot rely on the reconciliations.
How you notice it
Before the trial, confirm the tool logs agent actions and human approvals; Microsoft and Puzzle both document an activity log.

Human checkpoints

Proposed adjusting entries: the controller or accountant approves each before posting.
Accruals and estimates: prepared and approved by your team.
Reconciliations with open items above threshold: the account owner reviews.
Period lock and sign-off: the controller or CFO, after reviewing the close package.

The main risk on the close is not a wrong number the agent flags. It is a wrong number that looks reconciled. A proposed entry that forces a balance to agree can hide a missing invoice or a cut-off error. Review proposed entries against their supporting transactions, not against the balance they produce.

Expect review time. In Datarails' survey, finance teams spent 26% of their working week verifying or correcting AI output, and 75% of CFOs named lack of auditability as their main reason for not fully trusting AI with critical finance tasks. Those are large U.S. companies, but the lesson applies to small ones: budget the reviewer's hours and insist on a log.

How to run a trial

Run the agent beside your current close for two or three months before relying on it.

Pick a slice: bank and card accounts plus one subledger, such as AP or inventory.
Give the agent draft-only permissions. Your team closes as usual.
After each close, compare the agent's reconciliations with your team's: matched items, open items and proposed entries.
Count the proposed entries the controller accepted, changed and rejected. Rejections need a reason written down.
Time the review. Hours spent checking the agent are the cost you are trading against preparation time.
Plant a known error, such as a duplicated bill in the test slice, and confirm the agent flags it.

Pick a slice
Bank and card accounts plus one subledger, such as AP or inventory
Give the agent draft-only permissions
Your team closes as usual
After each close, compare the reconciliations
Matched items, open items and proposed entries
Count the proposed entries the controller accepted, changed and rejected
Rejections need a reason written down
Time the review
Hours spent checking the agent are the cost you are trading against preparation time
Plant a known error
Such as a duplicated bill in the test slice; confirm the agent flags it

If the agent meets the brief's criteria for two consecutive closes, let it prepare that slice for real and widen the scope one area at a time. Keep accruals, entry approval and sign-off where they are.

For structuring the test, see how to run a paid trial with an AI agent, and for the brief, how to write a job description for an AI agent. Related finance jobs: bookkeeping and reconciliation, accounts payable, FP&A and forecasting, the bookkeeper role and the finance overview.

Questions buyers ask

Will an agent shorten my close? The sources cited here do not give a days-to-close figure attributable to an agent. Team Car Care describes a different effect: with reconciliation running during the month and exceptions flagged as they arise, month-end becomes a review rather than a recovery. Measure your own days to close before and during the trial.

Can the agent post routine entries on its own? Some companies allow it for narrow cases. Pamarco says it turned on auto-posting for non-purchase-order invoices after its freight coding reached around a 95% success rate. That was invoice coding, not close adjustments, and the company chose it after measuring. Our recommendation: start with draft-only permissions for every close entry.

Do I still need an accountant? Yes. At Team Car Care the person supervising the agent is not a trained accountant, but entries still go to the GL team for posting, and judgment on estimates and sign-off stays with your finance lead.

Sources

Account Reconciliation Agent (production ready preview)
Microsoft LearnPrimaryAccessed Nov 14, 2026
Month-End close with AI Close
Puzzle Help CenterPrimaryAccessed Nov 14, 2026
Flux analysis in Puzzle
Puzzle Help CenterPrimaryAccessed Nov 14, 2026
17 CFR 240.13a-14 - Certification of disclosure in annual and quarterly reports (2025 edition)
U.S. Government Publishing Office (govinfo)PrimaryAccessed Nov 14, 2026
17 CFR § 240.13a-14 - Certification of disclosure in annual and quarterly reports
Legal Information Institute, Cornell Law SchoolPrimaryAccessed Nov 14, 2026
Pamarco accelerates finance operations with autonomous agents and Copilot Cowork
MicrosoftVendorAccessed Nov 14, 2026
How Team Car Care Reached Over 70% Reconciliation Automation With Nominal
NominalVendorAccessed Nov 14, 2026
Finance Teams Now Spend a Quarter of Their Week Fact-Checking AI, New Datarails Survey Finds
DatarailsVendorAccessed Nov 14, 2026

What we could not verify

Both production figures (Pamarco's roughly 40% of close tasks, Team Car Care's roughly 70% of inventory reconciliation) are reported in vendor-published customer stories. No independent study of agent accuracy on the close was found.
No source was found showing an agent calculating accruals, prepaids or other estimates in production, so that task is marked human.
Agent pricing is not shown. Pricing for the close products cited here was not researched to a sourced range, so the page has no cost comparison.
Days-to-close reductions published by close software vendors were not used; the ones found were single-customer claims and not clearly attributable to agents rather than to a new system.
Sign-off obligations for private companies (lender covenants, auditor engagement terms) vary by agreement and were not researched; the page cites only the SEC certification rule for public issuers.

Use this brief

The AI agent job board isn't open yet. Tell us about the job and Nikolas will reply personally.

I build or operate AI agents