Why Point AI Tools Can't Show Their Work in the Close

An auditor rarely stops at whether a number is right. The next question is how you got it, and whether you can reproduce it.

That's the test every number in the close has to pass, and it's why controllers hesitate on AI. They aren't against speed. They're against answers they can't trace. In July 2024, staff at the Public Company Accounting Oversight Board (PCAOB) reported on how preparers view generative AI. Preparers cited the black-box nature of some tools and inconsistent output as open questions for auditability.

If an AI tool can't show its data, its steps, and the same result on a second run, it hasn't saved you time. The tool has created a new item to defend. Point AI tools make that harder, because each one sees only part of the close.

A point AI tool does one job in one place. Some sit inside a reconciliation tool or a consolidation tool. Others sit outside the close entirely, as a general AI assistant pointed at exported data.

Every one of them works only with what it can see. That's fine for a contained job, such as drafting variance commentary or flagging anomalies in one account, if the output can be checked. But the close isn't a contained job. The close runs across entities, systems, and people, and every handoff is a place where numbers change. A tool that sees one piece delivers efficiency in that piece, not across the close.

Gartner reports that most chief financial officers (CFOs) are still early in adopting AI in cloud enterprise resource planning (ERP) finance applications. Data quality, integration complexity, skills gaps, and uneven multientity support are holding them back. More AI in more places adds to that integration work instead of removing it.

The first cost of a narrow view is context.

A flag says something looks wrong. Context says why, and what to do about it.

Take an intercompany entry that one side never booked. Caught mid-month, it's a quick correction. Caught at month-end, it's a scramble before sign-off that delays the close. Catching it early takes visibility across entities and the process, not one ledger.

Reconciliations work the same way. Traditionally they wait until the books are closed, so problems surface after the fact. When transaction matching and anomaly detection run during the month, more reconciliations clear on their own at close. People then focus on the exceptions that remain.

A reconciliation tool that can't see consolidation timing, or the upstream entries feeding an account, can raise the flag. A person still has to find the cause.

Narrative reporting has the same limit. AI can draft the explanation of a variance, but it can only describe what it can see. The real story comes from commentary captured during the process, alongside the evidence. Without it, you get a narrative on the numbers, not an explanation of them.

Context helps a person explain a number. Lineage lets an auditor verify it. Lineage is the record of where a number came from and every step it took. In a close that spans several applications, lineage has to cross every handoff.

AI can help assemble evidence from several systems. It can't create the link between them. The trails live in separate tools, so someone stitches them together by hand when the auditor asks.

Reproducibility matters just as much. If the same request returns different answers once numbers are final, the controller can't tell what changed: the data, the request, or the tool. A report that returned different results on each run wouldn't pass review, and AI output shouldn't either. Both sides need documenting: what the tool does, and how it's asked, including the parameters, dates, and accounts.

PCAOB staff also noted that some audit firms stress the auditability of both source data and AI-created content. Some design their tools to document the source data used. A Finance team's AI faces the same expectation.

Controls work only where the work happens. Four conditions keep AI inside your control framework:

• Approval is enforced inside the workflow, with segregation of duties, so an agent can't finish a step without a person.
• AI work stays inside the process. A result built from exported data or a side script sits apart from where human work and approvals happen, so it can't serve as evidence.
• Overrides are logged, and access follows the ERP. A user shouldn't see through AI what the ERP wouldn't let them see.
• Security covers unfinished results. AI with access to in-progress consolidations sees numbers that aren't final.

Security is also a recurring concern. In Deloitte's Finance Trends 2026 survey of 1,326 global finance leaders, 47% named data security as the top barrier to AI in their finance functions.

PCAOB staff also heard from some firms that preparers' use of AI could amplify existing IT risks, such as risks related to segregation of duties.

Treat AI like any other process that's in scope for controls. Document what it does, where its data comes from, which parameters it uses, and who reviews its output. When an auditor asks how the AI works, the answer should already exist on paper.

Accountability doesn't move. The controller or CFO signs the certification, and AI is a tool. AI can do the preparing. The reviewing, and the accountability that comes with it, stays with a person.

These six questions work for any AI feature, from any vendor.

1. Does it run on governed, reconciled data, or on data exported out of the process?
2. Do its actions land in the same audit trail as human actions?
3. Can approval be bypassed?
4. Can you reproduce the output when an auditor asks?
5. Can users edit AI results without a log?
6. Does it follow the same access controls as your ERP?

A "no" or an "I'm not sure" on any of them is a gap you'd have to explain at audit time. Write the answers down. The answers double as the documentation your auditors will ask for.

The pattern behind the questions is one platform, one set of data, and one trail. That's the design principle behind the OneStream platform. The consolidation engine does the calculation, AI prepares the explanation, a person approves, and each step lands in one audit trail.

Unifying consolidations, account reconciliations, transaction matching, and journal entries also removes the stitching described above. Financial data quality keeps the source data governed before AI ever sees it.

You don't need to rebuild the close to start. Pick high-volume, low-judgment work such as reconciliations and transaction matching, and set the guardrails first: security, approvals, and what the AI can touch.

Cox Enterprises took this route. The company built on a unified data platform first. It then started its AI agents in accounting, not financial planning and analysis, where workflows were predictable, high volume, and governed by clear policies. Cox also refined business context, example questions, and security setup with end users before scaling.

Then run the AI in parallel with your current process, the same way you would with a new ERP or consolidation system. Compare results, document every difference, tune the instructions, and keep going until both paths match. Don't start with a broken process, because AI can't repair one.

Talk to your auditors before year end. Ask what they'll test, including the IT side, so you aren't surprised after you've already relied on the tool.

AI can prepare the work, but it can't answer for it. Explore OneStream financial close and consolidation software to see how OneStream brings consolidation, reconciliations, and Finance-focused AI together on one platform.