AI Agents for Month-End Close: Controls the Controller Will Sign

An AI agent can speed up month-end close if it acts as a preparer and a human stays the reviewer.

That one rule keeps your controls intact. The agent drafts, matches, and chases. A person reviews and signs. The agent never approves its own work.

Here is which close tasks suit an agent, how to keep segregation of duties, what evidence the auditor will want, and a safe first task.

Where close time goes

Close is slow for boring reasons. The work is a long chain of small tasks with handoffs between people:

• pulling balances and subledger reports
• matching bank and intercompany items
• preparing reconciliations
• chasing owners for accruals and support
• writing variance explanations for the flux review
• posting and reversing journal entries
• checking that everything ties out
• getting reviews and sign-offs

Most of the elapsed time is waiting. Waiting for a report, waiting for an owner to reply, waiting for a reviewer who is busy with three other things.

You will see claims online that AI cut close from 10 days to 3. I could not find a primary source for any of them. Treat those numbers as marketing until you can see the method, the company size, and what was measured. I also do not have a trustworthy public benchmark to give you for monthly close days, so measure your own baseline.

My view: the goal is not a faster close at any cost. The goal is a faster close that your controller, your auditor, and your CFO would all sign. If you are building the broader case, the CFO guide to AI agent investment covers how to frame the money side.

Suitable tasks

I rank close tasks by risk and by how well they fit what an agent does well: reading, matching, summarizing, and drafting with a source.

Good fits:

1. Checklist chasing. The agent tracks the close calendar, sees which tasks are late, and sends reminders and requests for support. Low risk, high time saved. It touches no ledger.
2. Reconciliation prep. The agent pulls the GL balance and the subledger or bank balance, matches items on rules, and lists unmatched items with possible explanations. A person reviews the exceptions.
3. Flux commentary drafts. The agent compares the period to the prior period and budget, finds the movements above your threshold, pulls the drivers from transaction detail, and drafts a one-line explanation with links. The preparer edits and owns it.
4. Support gathering. The agent collects invoices, statements, and contracts for accruals and attaches them to the workpaper.
5. Tie-out checks. The agent re-computes totals and checks that schedules agree to the trial balance.
6. Recurring entry proposals. The agent drafts a standard accrual or reversal from a template and prior-period logic. It does not post it.

The PCAOB’s staff reported in July 2024 on outreach to audit firms and public company preparers. Some preparers said they use generative AI to assist with less complex, repetitive work such as preparing account reconciliations or identifying reconciling items. They also stressed that human involvement in supervising the output remains necessary. The same document notes that generative AI output may not always be reliable. It is a staff update, not a rule. But it matches how I would draw the line.

Poor fits for a first deployment:

• anything that posts to the ledger without a person
• judgmental estimates and reserves
• revenue recognition conclusions
• anything that changes master data or approval limits

Preparer and reviewer separation

Segregation of duties is the control that auditors check first. The GAO’s Green Book, the federal internal control standard, defines it as separating the responsibilities for authorizing transactions, processing and recording them, reviewing them, and handling related assets, so that no one individual controls all key aspects of a transaction. It is a federal standard (2014), but the principle is the same one your auditors apply.

With an agent, the question becomes: which role does the agent hold?

My recommendation is simple. The agent is a preparer. Always.

Role	Who or what holds it	Rules
Preparer	Agent, or a human using the agent	Drafts the rec, the commentary, the proposed entry
Reviewer	A named human, not the person who prepared	Reviews the output and the evidence, signs
Approver (entries)	A human with posting authority	Approves and posts
Agent permissions	Read access to GL, subledgers, bank files. Write access to a draft workspace only	No posting rights. No approval rights
Control owner	Controller or assistant controller	Owns the agent’s scope and changes to it

Three design rules that follow:

1. The agent never reviews its own work. If an agent drafts a reconciliation, a second agent does not count as the reviewer. A person does.
2. Reviewers must see evidence, not just conclusions. A reviewer who clicks “approve” on a summary has not performed a review. The workpaper must show the source rows and the matching logic.
3. Least privilege. OWASP’s guidance on excessive agency recommends limiting an LLM application’s tools and permissions, using narrow functions, and requiring human approval of high-impact actions. It also says authorization should be enforced in the downstream systems, not left to the model. For close, that means the ERP enforces what the agent can do, not the prompt.

If a preparer uses the agent, the preparer still owns the output. “The AI did it” is not a review note.

Evidence and audit trail

Public companies must assess internal control over financial reporting each year. Under the SEC’s rule implementing Section 404, management reports on its assessment, and the auditor attests to it. Management cannot conclude controls are effective if a material weakness exists. Private companies are not bound by that rule. Their lenders, investors, and auditors still care.

So if the agent is part of a control, you need to be able to show how it works. What I would capture for every agent-prepared item:

• the task, period, and entity
• the inputs: report names, run times, and the file or query used
• the matching or calculation logic and its version
• the output, with links to the source rows
• the model and prompt version in effect
• who reviewed it, when, and what they changed
• any items the agent could not match or flagged as uncertain

Keep these in the same place your auditors already look, with the same retention as the rest of the workpaper. The full list of what to log is in the AI agent audit trail guide. For the wider context of finance teams adopting agents, see AI agents in financial services.

Ask your auditor before the pilot, not after. Show them the scope, the roles table, and a sample workpaper. Ask what they would need to rely on the output.

Number tie-out checks

Language models are not calculators, and they can be inconsistent. The τ-bench paper tested agents on repeated runs of the same tasks and found agents often failed to succeed consistently across trials. That was 2024 research on customer service tasks, so do not overread it. But finance has a simple fix: do not let the model do the math.

My rule: the model reads and writes. Code adds and compares.

Build these checks as plain code that runs after the agent finishes:

• Footing and cross-footing. Every schedule adds up in both directions.
• Tie to source. Each balance in the rec equals the GL or subledger report, exactly.
• Tie to trial balance. Totals in the flux table agree to the TB.
• Sign and unit checks. No flipped signs, no thousands-versus-units slips.
• Completeness. Every account on the close list got a rec, or a documented reason it did not.
• Reasonableness. Flag any drafted explanation that cites a number not found in the source data.
• Repeat run. Run the same task twice. If the output differs, hold it for review.

If a check fails, the item goes to a human with the failure shown. It does not get retried silently until it passes.

This is also where you set a stopping rule. Anthropic’s guidance on agents suggests stopping conditions and pausing for human feedback at checkpoints. In close, a failed tie-out is a checkpoint.

What not to automate

Some close work should stay with people, at least at first:

• Posting journal entries without human approval.
• Judgment calls: estimates, reserves, impairments, revenue recognition positions.
• Management review controls over significant estimates. The reviewer must apply judgment, not read an AI summary.
• Changing the chart of accounts, mappings, or approval limits.
• Communications with the auditor on contested items.
• Fixing a failed control by changing the agent’s prompt mid-close.

Also be careful with the idea that AI will “reconcile everything.” Reconciliations have two jobs: prove the balance and surface problems. An agent that auto-clears items to reach a zero difference helps with neither. Set rules for what the agent may clear (exact matches within a stated tolerance) and send everything else to a person.

Many close problems start upstream. See AI agents for accounts payable exceptions.

Pilot scope

Keep the first scope small and measurable.

Here is the pilot I would run.

A safe first task: flux commentary drafts for one P&L area. Or: reconciliation prep for a small set of bank accounts. Both are read-heavy and reviewed by a person. Neither posts anything.

Say a mid-size company picks operating expense flux for one business unit, with a threshold of 10% and $5,000 (these are made-up numbers; set yours with your controller).

Step list:

1. Document the current process: who prepares, who reviews, how long each step takes. Time two or three closes to get a baseline.
2. Write the agent’s scope on one page: inputs, outputs, permissions (read-only plus a draft workspace), and the reviewer.
3. Build the tie-out checks as code before you build the agent.
4. Run in parallel for one close. The preparer does the work as usual. The agent produces its draft separately. Compare them. Nothing from the agent enters the real workpaper.
5. Run assisted for the next close or two. The preparer starts from the agent’s draft, edits it, and signs as preparer. The reviewer reviews as always.
6. Measure: preparer hours, reviewer rework (how many commentary lines were rewritten), tie-out failures, and days from period end to review complete.
7. Review with the auditor before expanding.

Hero metric: preparer hours per flux package, with reviewer rework rate as the guardrail. If hours drop but reviewers rewrite most lines, you saved nothing.

Stop rules: stop and review if any drafted number fails tie-out and reached a reviewer, if the agent touched anything outside its scope, or if a reviewer approved without opening evidence.

Checklist before you go live:

• Agent holds the preparer role only, with no posting rights
• A named human reviewer, not the preparer
• Permissions enforced in the ERP, not in the prompt
• Tie-out checks run as code after every output
• Logs captured and stored with the workpaper
• Auditor briefed on scope
• Rollback plan: the team can close by hand without the agent
• Prompt and model changes frozen during close week

Bottom Line

The agent is a preparer. A human is the reviewer. The ERP enforces permissions.

Start with something that posts nothing: flux commentary drafts or reconciliation prep. Do the math in code, keep evidence with the workpaper, and brief your auditor before you start.

Run one parallel close, then an assisted one. Measure preparer hours and reviewer rework.

FAQ

How can AI speed up month-end close? Mostly by cutting waiting and drafting time. An agent can chase close tasks, gather support, prepare reconciliations, and draft flux commentary. A person still reviews and signs. Be wary of headline claims like “10 days to 3” that come with no method.

Can AI do reconciliations? It can prepare them: pull balances, match items on rules, and list exceptions with likely causes. PCAOB staff noted in 2024 that some preparers use generative AI for tasks like this, with human supervision. A person should review the exceptions and sign the rec. Do the arithmetic in code, not in the model.

How do I keep segregation of duties? Give the agent a preparer role only, with read access and a draft workspace. A different named person reviews. Posting and approval stay with humans, enforced by the ERP’s permissions. The agent never reviews or approves its own output.

What does the auditor need? Evidence that the control works: inputs, logic and its version, outputs with links to source rows, model and prompt versions, and who reviewed and what they changed. Show your auditor the scope and a sample workpaper before the pilot.

What is a safe first task? Flux commentary drafts for one area, or reconciliation prep for a few accounts. Both are reviewed by a person and post nothing. Run one close in parallel, then one or two assisted, before you widen scope.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect