Guide · November 19, 2026

AI in the Close: What Your Auditor Will Ask, and How to Prove It

Introduction

Why Your Caution About AI Is Sound

An auditor doesn’t just want the numbers. The auditor wants to know how you got it, and you have to prove it. Artificial intelligence (AI) makes getting the number faster. Proving it is still your job.

Controllers ask harder questions about AI than most buyers. They should. You’re asking whether the number can be traced.

Look at what happens when you’re not careful. In October 2025, Deloitte Australia agreed to refund part of the A$440,000 the Australian government paid for a report. According to the Associated Press, the report contained apparent AI-generated errors. Those included a fabricated quote from a federal court judgment and references to research papers that don’t exist. The report went to a government department, and the errors reached the client.

In my prior life, I was in internal audit and also a Controller. My position on AI in the close is simple: AI prepares and flags, controlled systems calculate, and people review, approve, and sign. This guide covers what to document, test, retain, and show an auditor.

The rule this guide follows“AI prepares and flags, controlled systems calculate, and people review, approve, and sign.”
— Tina Huysmans

Chapter 1

What an Auditor Will Ask About an AI-Prepared Number

Picture an auditor pointing at one number in your close package and asking how you got it. Three tests decide whether you can answer: trace it, rerun it, explain it.

Trace: Can You Show the Path?

Show the source, the steps, and the parameters. Did the number come from the general ledger (GL) or from a spreadsheet somebody built? Was it extracted before or after the late entries posted? You already document the parameters you choose when you set up an enterprise resource planning (ERP) system or a close tool. Hold AI to the same standard.

Rerun: Do You Get the Same Answer Twice?

When I was a controller, internal audit asked me to show the parameters I’d used to run a report I relied on for review. Then they asked me to rerun it and confirm the result matched. Nobody enjoys a rerun, but the test was fair.

For a number, I apply that test to AI: same source snapshot, same parameters, same result. Rewording a request counts as changing a parameter, like running a report with a different department list.

Explain: Can You Tell the Story?

An auditor won’t question an AI agent. The auditor will question you. Be ready to say how you got the number, where AI fit in, what you use it for, and what you don’t.

What the Standards Say

For audits under Public Company Accounting Oversight Board (PCAOB) standards, Auditing Standard 1105 covers information produced by the company, which auditors call IPE. That includes the key reports you rely on in a review. Auditors test its accuracy and completeness, or the controls over them. Company-produced information (IPE) is more reliable when those controls are effective. Amendments take effect for audits of fiscal years beginning on or after December 15, 2025. They name information technology (IT) general controls and automated application controls, where applicable.

I’d expect an auditor to treat AI-prepared figures the same way. (That’s my reading, not a PCAOB statement about AI.) Other standards may differ, so ask your auditor.

Not Everything Called “AI” Carries the Same Risk

A vendor’s “AI” label tells you little about control risk. What matters is what the capability is permitted to do and what happens when it’s wrong. Four kinds show up in a close:

Rules-based automation

Applies fixed rules, such as matching on invoice number and amount. Same input, same output. Test the rules and control changes to them.

Analytical AI or machine learning

Scores or ranks items. It can miss real problems and flag fine ones, so measure both and retest when the data or tool changes.

Generative AI

Drafts text. The wording can change between runs, and a draft can state a cause nobody verified.

Agentic AI

Takes multiple steps with access to systems. The risk is less what it says than what it can reach and change.

Set Controls by Activity, Not by Label

Drafting an explanation

The draft is a draft. A person verifies every stated cause against evidence.

Identifying an anomaly

A flag opens a question. Someone records what they did about it.

Matching transactions

Rules or machine learning can propose matches. Review the exceptions, and sample the matches too.

Calculating an authoritative balance

A controlled, deterministic system does this, and you can reproduce the result on demand. A language model doesn’t.

Changing, approving, or posting

A person, or an enforced workflow step a person completes. An agent that can post alone has a different control profile from one that only reads.

The Evidence I’d Retain

Take a variance explanation drafted by generative AI for one entity’s operating expenses. For each one, I’d keep:

  • Source data: Report, entity, period, extract date and time, and a snapshot or version ID. Books change after extraction.
  • Instruction: The exact prompt or template, the parameters, and who ran it.
  • Tool and version: System and version, plus model version if the tool exposes it. If it doesn’t, record the gap and tell your auditor.
  • Raw output: The draft, saved unedited.
  • Human changes: The edited version, with who changed what and when.
  • Support: For each cause stated, the evidence: an invoice, an operations report, a business owner’s email.
  • Exceptions: What the AI flagged or couldn’t explain, and what happened to each.
  • Review and approval: Reviewer, approver, timestamps, decision, and any rejection or escalation rationale.
  • Final commentary: The version that goes in the close file.

Ask your auditor and your records policy how long to keep all this.

When a Rerun Doesn’t Match

When a rerun doesn’t match, save the first output before anything else. Then:

  1. Compare the inputs: same snapshot, same parameters, same tool and model version. Changed inputs are the cheapest explanation.
  2. Classify the difference. A different number from identical inputs means stop. Don’t rely on that output until someone explains it.
  3. Log what differed, the root cause, and who decided what.
  4. Escalate what you can’t explain and fall back to the manual process for that item and anything built on it.

Chapter 2

Why Your Close Breaks Before the Audit Starts

The damage starts in the month, when small errors sit unnoticed until close. Three problems keep coming back.

Why Do Late Entities and Open Reconciling Items Break the Close?

Because each one points to a process problem nobody has looked at yet, and the auditor will look at it for you.

A close confirms everything that belongs in the numbers is in them, they’re accurate, and they reflect the month’s activity. Sort each late entity into a bucket: personnel, setup, or business process. If regional controllers report to you, you need to see what they see, entity by entity and month by month. Only then can you tell which bucket a repeat offender belongs in.

Open reconciling items are the second signal. When one carries over month after month, the auditor will ask why. A pile of them widens the scope of the work and slows consolidation. For each item, I’d want an owner, an age, a reason it’s open, and a clear-by date. AI can build that list. It can’t make anyone clear it.

Why Are Intercompany and Foreign Exchange (FX) Issues a Challenge?

Usually because the error started earlier in the month and nobody found it until the end.

A lot of close delays trace back to bad processes during the month. Transactions get recorded incorrectly. Or the ERP isn’t set up to handle a transaction the way the business needs, and somebody adjusts by hand at close. Intercompany and FX corrections follow the same path.

Take this made-up example: One side of a $1 million intercompany entry never gets booked. Find it on the last day, and it’s a bottleneck. Find it in week two, and it’s a quick fix. Matching intercompany activity in both directions during the month gets you week two.

Customer example: TopBuild

TopBuild’s Director of SEC Reporting and Corporate Accounting says intercompany eliminations used to eat time in gathering, filtering, and formatting data. With transaction matching, the team now has its matches a minute after pushing a button and works only the exceptions. Trial balances that once meant exporting and importing from several ERP systems now load at the click of a button. Consolidated financials reach executives in as little as 15 minutes.

  • 1 min

    After pushing a button, the team has its intercompany matches and works only the exceptions

  • 15 min

    As little as 15 minutes for consolidated financials to reach executives

Why Do Account Reconciliations Still Sit Outside the Close?

Because they need final numbers, and final numbers show up late.

Here’s what the alternative looked like when I first came into my role as a controller. Reconciliations were still done in excel and usually after the month had been closed, leaving issues open for another month. Most transaction matching was done in an excel file with formulas. Too laborious to be done on a daily basis.

Waiting means you learn about problems after the close. No structure and control means you do things more than once. During the month, run matching on high-volume accounts and monitor those with less activity. If an account normally gets one standard journal entry a month and a second one lands, someone should hear about it then. At close, reconcile against final balances with most items already cleared. Moving checks into the month costs reviewer time, so budget for it.

Chapter 3

What Good Looks Like, and How to Shorten the Cycle

AI prepares explanations and surfaces exceptions. A consolidation engine does the calculation. A person approves. One audit trail covers all of it.

FX translation and intercompany eliminations belong in a system built to calculate them; with results you can rerun. ERP integration comes down to lineage. What the AI works on should be governed and reconciled, not exported and moved. Every hop is another place for the trail to break. Auditors test company-produced information, or the controls over it, for accuracy and completeness.

AI can assemble evidence and draft a narrative. It can’t tell you about pending litigation or a customer’s credit situation. It can’t tell you why paving revenue fell in a rainy summer at the building materials company where I once worked. People supply that context, and someone should capture it as commentary during the month. If the explanation lives in an operations report or an email, link it into the file. An explanation nobody can retrieve isn’t support.

Where the Simple Rule Gets Hard

The number is right, for the wrong reason.

A matching tool pairs two items because the amounts tie, though they’re unrelated transactions. The total reconciles, so nobody looks. Test the logic, not just the result: Review a sample of matches by their reason and keep the match criteria with the evidence.

The numbers reproduce, but the explanation changes.

Don’t demand identical prose. The approved explanation and its support are the record. What must not change is the numbers it cites and the causes it claims.

A reviewer edits the draft and adds a cause nobody verified.

“Higher volume” reads well and may be wrong. Treat the edit as a new claim that needs its own support, whoever typed it. Track edits so you can see which claims came from the tool and which from a person.

The source data changes between runs.

Late entries, reclasses, and reloads create a different input, not a failed rerun. Snapshot the data for each run, rerun against the snapshot, and note what changed afterward.

An agent can see in-progress results outside the user’s access.

Stop. Treat the exposure as an access-control failure and report it to IT and internal audit. Don’t resume until the agent carries the user’s permissions, no more.

How Do You Shorten the Close for a Complex Company?

Stop saving the checking for the last days. (I won’t promise you a number of days — I don’t know your close, and a number without context is a sales claim.)

Here’s what I’d do, in order: Fix the causes first, move checks into the month, and start where volume is high and judgment is low. Then prove the new process against the old one before you rely on it.

That last step is a control procedure, so write it down before the first run:

  • Scope: Process, entities, accounts, periods.
  • Baseline: How the current process performs: hours, exceptions found, rework, who reviews.
  • Expected results: What correct looks like, ideally from closed periods where you know the answer.
  • Tolerance: How much difference you’ll accept, tied to materiality and policy. You set it. I won’t.
  • False positives and false negatives: How each is counted and handled. A missed real exception is worse than an extra flag.
  • Reviewer: Who investigates differences, independent of whoever configured the tool.
  • Discrepancies and escalation: Root-cause every difference, log it, and name what stops the pilot.
  • Evidence retained: The list above.
  • Cutover and exit criteria: What must be true to switch, and what ends the pilot.

An illustrative example: Piloting transaction matching

You pilot transaction matching on bank reconciliations for one entity group. The baseline is three closed periods and the exceptions your team documented. Expected results: every documented exception shows up in the tool’s list. Plant a few known errors too, such as a duplicate and a missing deposit, in a test environment or a copy of closed-period data, never the live ledger. Document each one before the run, so you can tell a quiet tool from a clean one.

Set a dollar threshold from your materiality. Any real exception above it that the tool misses stops the pilot until someone explains why. Extra flags below it aren’t failures, but count them and record reviewer time because that’s the real cost. Cut over only after live-close results also land inside tolerance and your reviewer, internal audit, and auditor have seen the evidence.

Chapter 4

How to Choose Close Software and Set the Controls

Judge it on seven questions, and make the vendor show the answers.

Does the AI work on governed, reconciled data?

“Show me where the data behind this output came from. Now change the source data and show me what happens to the output.”

Are AI actions in the same audit trail as human actions?

“Show me one log with an AI action and a human edit to the same number.”

Can anyone bypass approval?

“Attempt to bypass approval, as an administrator and as a regular user.”

Can you reproduce an output?

“Rerun yesterday’s request with the same inputs. Show me what the system tells me when the results differ.”

Can users edit AI results without a log entry?

“Change this output, then show me what the audit log records.”

Can you retrieve the evidence later?

“Show me what an auditor could retrieve 6 months from now, including the exact source information behind this explanation.”

Do the access controls in the solution carry over to agentic AI?

“Sign in as a division controller and ask for another division’s numbers. Now show me what the agent can reach.”

Stricter controls slow things down. For audit prep, I’d take the extra step.

Who Is Accountable When AI Helps?

People, always.

Companies that file periodic reports with the SEC have a certification requirement. The principal executive officer and principal financial officer, or persons performing similar functions, must each certify every quarterly and annual report. Separately, a periodic report containing financial statements must be accompanied by a written statement from the chief executive officer and chief financial officer, or equivalents. Knowingly certifying a report that doesn’t comply carries a fine of up to $1 million, up to 10 years in prison, or both. Willful certification carries up to $5 million, up to 20 years, or both.

  • $1M

    Fine for knowingly certifying a report that doesn’t comply, up to 10 years in prison, or both

  • $5M

    Fine for willful certification, up to 20 years, or both

AI isn’t accountable for anything. People are. Build approval into the workflow, and segregate duties so an agent can’t finish a step alone. Require human sign-off on anything that changes a number or leaves the company. If you work at a private company, these particular rules may not apply to you. The control logic still does.

Conclusion

Your First 30 Days

Consider this plan as a pilot, not a complete rollout. A live parallel run needs a close, so if yours falls outside the 30 days, run the live leg in the next one.

Days 1 to 5: Select and baseline

Pick one process that is high in volume, repeatable, and low in judgment, with closed periods where you know the right answer. Stay away from estimates. Measure the current process before AI touches it: hours, exceptions found and how long they took, rework, and reviewers. Write the acceptance criteria from Chapter 3, including tolerance and materiality, and get your approver to agree in writing.

Days 6 to 12: Configure, with IT and internal audit

Limit access to the pilot scope. Turn on logging and enforced approval. Record tool and model versions. Agree with the vendor how you’ll be told before a model version changes, and get the vendor’s SOC 1 Type II report to IT and internal audit. Decide where evidence lives and who can retrieve it. Have IT and internal audit review access, logging, and retention now. A logging gap is cheaper to find in week one than in the audit.

Days 13 to 25: Run in parallel

Run closed periods first, then the live close, beside the current process. Log every discrepancy with a root cause and a decision. Sample matches, not just exceptions. Brief your external auditor before the live run: scope, evidence plan, and what they would test on the accounting and IT sides.

Days 26 to 30: Decide

Success means:

  • Results sit inside the tolerance you wrote down
  • Reviewers can trace, rerun, and explain every sampled item from retained evidence alone
  • IT, internal audit, and your auditor have no open objections
  • Your team can run the process by hand and get the same result

There’s no universal threshold for any of this. You set each one from your materiality, risk appetite, and policy.

AI prepares. People sign.

Request a custom demo from OneStream → If you’d like to test a vendor against the eight demo tests in Chapter 4.

FAQ

Frequently Asked Questions

What will an auditor ask about an AI-prepared number?

How you got it. Three tests decide whether you can answer: trace it, rerun it, explain it. Show the source, the steps, and the parameters; rerun it with the same source snapshot and the same parameters and get the same result; and be ready to say how you got the number, where AI fit in, what you use it for, and what you don’t. An auditor won’t question an AI agent. The auditor will question you.

How do auditing standards apply to AI-prepared figures?

For audits under Public Company Accounting Oversight Board (PCAOB) standards, Auditing Standard 1105 covers information produced by the company, which auditors call IPE. Auditors test its accuracy and completeness, or the controls over them. Amendments take effect for audits of fiscal years beginning on or after December 15, 2025. The author expects an auditor to treat AI-prepared figures the same way; that’s the author’s reading, not a PCAOB statement about AI. Other standards may differ, so ask your auditor.

What evidence should you keep for an AI-drafted variance explanation?

The source data and its snapshot or version ID, the exact prompt or template and who ran it, the tool and model version, the raw output saved unedited, the human changes, the support for each cause stated, the exceptions and what happened to each, the review and approval, and the final commentary that goes in the close file. Ask your auditor and your records policy how long to keep all this.

How do you shorten the close for a complex company?

Stop saving the checking for the last days. Fix the causes first, move checks into the month, and start where volume is high and judgment is low. Then prove the new process against the old one before you rely on it. Moving checks into the month costs reviewer time, so budget for it.

What should end an evaluation of close software with AI?

Three results end the evaluation: an output change that leaves no log entry, an approval step that can be bypassed, and an output you can’t trace to its source data and parameters. Before the demo, ask for the vendor’s SOC 1 Type II report and its list of complementary user entity controls (CUECs). Your auditor will ask for both, and the CUECs are controls you have to run.

Who is accountable when AI helps in the close?

People, always. AI isn’t accountable for anything. People are. Build approval into the workflow, and segregate duties so an agent can’t finish a step alone. Require human sign-off on anything that changes a number or leaves the company.

Demo Sign Up