Which tasks an AI agent finance close may touch in a small team

Share Now

A brass teller's cage holding draft pages, lit by a blue lamp, with a hand turning the key in the locked drawer below, for an AI agent finance close

Which tasks an AI agent finance close may touch in a small team

Month-end still ends with a person who can defend every number. An AI agent finance close changes who drafts the pack, not who owns the risk. The useful question is which tasks the agent may touch, and which ones must stay with a named human who can stop the process.

Professional bodies already talk about AI in finance work in these terms. In its guide to using generative AI, ICAEW is blunt: “Professional scepticism applies equally to the output from AI, as it does to any manually produced information.” NIST’s AI Risk Management Framework 1.0 makes the same demand in governance language: “GOVERN is a cross-cutting function that is infused throughout AI risk management and enables the other functions of the process.” For a small finance team, that means naming who may start an agent, what the agent may see, and who signs the close.

What an AI agent finance close may draft

Give the agent work that produces a draft a person will still read line by line. That includes pulling a trial balance into a standard pack, listing unpaid invoices older than a set age, drafting variance notes against last month’s signed pack, and preparing a checklist of missing schedules. None of that moves cash. All of it leaves a paper trail the controller can reject.

Keep the agent’s scope on files the firm already controls. The earlier Current Ventures piece on when client data should not go to someone else’s model still applies: if the close pack includes customer names, payroll lines or bank detail, say whether those fields may leave the building. An agent that only reads a redacted extract is safer than one that scrapes the live ledger.

ICAEW’s guide to using generative AI is blunt about why a draft still needs a reader. “Professional scepticism applies equally to the output from AI, as it does to any manually produced information.” It also warns that outputs specific to accounting, audit, tax, standards, laws and regulations may be tailored to specific jurisdictions without making that clear. A Mauritian close pack that quietly borrows a UK accruals story, or a Kenyan VAT treatment, is not a small typo. It is a wrong number with a confident tone.

A worked month-end, in plain sight

Imagine a 6-person professional-services firm in Port Louis. The finance lead also prepares a pack for a Nairobi subsidiary. The usual month-end file has 8 schedules: a trial balance, aged receivables, accrued bonuses, prepaid expenses, an intercompany recharge, a payroll summary, a bank reconciliation and a one-page commentary for the directors.

A safe first use of the agent is narrow. It is allowed to pull the trial balance into the pack template, list receivables older than 60 days, and draft commentary only where this month’s signed-off figure differs from last month’s signed pack by more than a threshold the firm has already written down. It is not allowed to post journals, accept a near-match on the bank rec, or invent a reason for the bonus accrual.

In this picture, the useful failure is easy to spot. The agent drafts a neat bonus-accrual note that “follows last month,” but last month included a one-off award that should not repeat. A person who still reads the note catches it. A person who only checks that a note exists does not. The point of the pilot is to find out which of those 2 reviews you actually have.

The payroll schedule makes the data question concrete. If that schedule carries staff names, national IDs or salary lines, the agent should not see the live file. Feed it a redacted extract, or keep payroll as a human-only schedule. The pack may combine Mauritian and Kenyan personal data in one folder, and the earlier rule on client data leaving the building applies to staff data as well.

What must stay with a person

Keep these with a named human, even when the agent is excellent at drafting.

  • Any action that moves money: payments, refunds, write-offs, or changing a supplier bank account.
  • Bank matching when the match is not exact. Near-matches are where fraud and typo errors hide.
  • Judgement calls on revenue recognition, provisions, related-party items and anything that exists only as an estimate.
  • The final sign-off that the pack is ready for the board or the auditors.

ICAEW’s note on managing risks with robust controls describes the same split as a design choice, not a slogan. Controls can be hard limits, or they can be process design, “such as including a human in the process for any significant decisions.” Many teams, it says, keep a human in the loop by using AI “only to make recommendations for a human operator to choose from.” That is the right shape for a close. The agent proposes. A person still chooses.

The same note argues for kill switches: manual or automatic triggers that can rapidly suspend the system. In a small finance team that can be as simple as a written rule that the agent is turned off mid-close if the bank feed breaks, if a draft invents a figure, or if the named reviewer is away. A better prompt is not a substitute for a stop rule.

A decision framework for each close task

Before the next close, put every recurring task on one page and mark it as draft, review, or human only. The test is not “could the agent try this?” The test is what happens if the output is wrong, and whether a person can still see the source.

How to classify close tasks before an agent sees the ledger
Task Agent role Why
Pull the trial balance into the pack templateMay draftNo cash moves. A person still checks completeness against the source report.
List unpaid invoices older than a set ageMay draftAssembly work. The credit-control judgement stays with a person.
Draft variance notes against last month’s signed packMay draft, must be read line by lineUseful first pass. A confident wrong story is worse than a blank.
Flag missing schedules against a fixed checklistMay draftThe checklist is the firm’s, not the model’s.
Accept a near-match on the bank reconciliationHuman onlyNear-matches hide fraud and typo errors. This is a significant decision.
Post journals, payments, refunds or write-offsHuman onlyThese move the ledger or move money. The agent does not get a login that can do either.
Judgement on revenue, provisions or related partiesHuman onlyProfessional scepticism still applies. A person still makes the judgement.
Final sign-off of the packHuman onlySomeone has to be able to defend every number.
This is a classification aid, not a legal standard. The “human only” rows follow ICAEW’s guidance that significant decisions stay with a person, and NIST’s emphasis on roles under the GOVERN function. Sources: ICAEW, managing risks with robust controls, NIST AI RMF 1.0.

If a task cannot be classified in one sitting, leave it human only until it can. Ambiguity is a reason to wait, not a reason to let the agent try.

A one-page brief for the team

Write 4 lines and stick them next to the close checklist. This is governance in the NIST sense: roles, scope and a stop rule, written down before the agent sees the ledger.

  1. Purpose: the agent drafts the pack. It does not post journals or pay anyone.
  2. Data: list the folders and exports the agent may open. Everything else is out of bounds. If a schedule contains staff or customer personal data, say whether a redacted extract is required.
  3. Human review: name the person who must open every draft before it enters the signed pack, and name a deputy for leave weeks.
  4. Stop rule: if the agent cannot find a source, it leaves a blank and a note. It never invents a figure. If it does, the agent is turned off for the rest of that close.

Pair the brief with the practical boundary in what should stay in the spreadsheet. If the only record of a balance is an untested workbook, fix the record before you automate the draft. An agent that reads a messy spreadsheet will draft a messy pack with more confidence than the spreadsheet deserves.

ICAEW’s generative AI guide also asks firms to be clear on “data that will or will not be shared with third parties, where it is stored, who owns it and how it is controlled,” and to define what is allowed before tools are easy to reach. A 4-line close brief is that policy at the size of a small team.

How to pilot without blowing up month-end

Run the agent on last month’s already-signed pack first. Compare its draft to the signed one. Measure 3 things: missing schedules, wrong variance stories, and any invented number. Record the misses in a short log, not as a vibe. Only then try the live close in parallel, with the usual process still in charge. If the parallel pack disagrees, the live process wins until you understand why.

Do not measure success as “the agent finished faster.” Measure it as “the controller spent less time on assembly and the same or more time on judgement.” Speed without review is just a faster way to sign the wrong pack.

ICAEW’s controls guidance also argues for a regular review after go-live, not just a launch test: a look at trends in the system’s high-level outputs, plus a spot review of a sample of particular decisions. For a monthly close, that can be a 30-minute review after the pack is signed, covering the invented-number count and the times the stop rule was used. If either rises, widen the human-only column before you widen the agent’s scope.

For a longer view of where AI already sits in finance operations, see AI in the finance function that mid-sized firms actually use.

When the agent should stay off

Leave the agent off when the chart of accounts is mid-rebuild, when a bank feed is broken, or when half the team is new and still learning which schedules matter. Automation amplifies the process you already have. A messy close with an agent is a messier close with a confident draft.

Also leave it off when personal data in the pack has no agreed retention or access rule. If you cannot say who may see payroll lines inside the agent’s context window, the answer is not a cleverer prompt. It is a human-only payroll schedule until the rule exists.

The same caution applies when the pack crosses borders. A Mauritius parent that consolidates a Nairobi or Lagos entity should not drop every subsidiary export into one workspace just because it is convenient. Until you can name the system, the country and the redaction steps, those exports stay out of the agent’s reach.

Where this sits next to the working agreement

If the firm already has an AI working agreement, put the close brief under it rather than inventing a second policy. The working agreement covers who may use which tools. The close brief covers which money and people decisions stay human. Keep both short enough that the team will actually open them in the last week of the month.

NIST’s GOVERN function is designed to run through the rest of the work, not to sit in a binder. “Attention to governance is a continual and intrinsic requirement for effective AI risk management over an AI system’s lifespan and the organization’s hierarchy.” For a 6-person team, continual governance looks like a one-page brief, a named reviewer, a stop rule and a short review after each close.

The close will still end the way it does now: with a person who can defend every number. The agent is worth using if that person spends more of the last week on judgement, and less of it on assembly. It is not worth using if the only change is a faster draft that nobody really reads.

Key takeaways

  • An AI agent may draft the close pack. It should not move money, accept near-matches or sign off.
  • Classify every recurring task as draft, review, or human only before the agent sees the ledger.
  • Write a 4-line brief: purpose, data, human review, stop rule.
  • Pilot on a signed historical pack, then run in parallel. Count missing schedules, wrong stories and invented numbers.
  • ICAEW’s guidance on scepticism and human review, and NIST’s GOVERN function, both point to a named owner, not a better prompt.

FAQ

Can an AI agent post journals in the close?

No. Keep posting, payments and write-offs with a named person. The agent drafts. A person posts.

Does professional guidance ban AI in a finance close?

No. ICAEW says professional scepticism applies to AI output as it does to any manually produced information. The close still needs a named person who can defend the numbers.

What is the first pilot that is actually safe?

Rebuild last month’s already-signed pack. Compare drafts on missing schedules, wrong variance stories and invented numbers. Only then run in parallel on the live close.

Where do payroll and client files fit?

Say whether staff, customer or bank fields may leave the building. Prefer a redacted extract the firm controls, especially when a Mauritius pack includes another country’s data.

How does this relate to NIST GOVERN?

NIST AI RMF 1.0 treats GOVERN as a cross-cutting function. For a small team that means a written brief, named roles and a clear stop rule before the agent sees the ledger.

What should the stop rule contain?

If the agent cannot find a source, it leaves a blank. If it invents a figure, or if a bank feed breaks, the agent is turned off for the rest of that close.

Sources

  • National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework (AI RMF 1.0)”, 26 January 2023. nvlpubs.nist.gov
  • ICAEW, “Using generative AI”. icaew.com
  • ICAEW, “Managing risks with robust controls”. icaew.com

Related Posts