Local CFO Asks AI to Close the Books, AI Asks for Supporting Documentation Like Everyone Else

SATIRE — At 4:47 p.m. on the third business day, when optimism has usually been reclassified below the line, CFO Martin Vellum announced that artificial intelligence would close the books. The company’s finance team, which had spent the previous 47 hours explaining that the books could not be closed until the bank feed stopped describing wire transfers as “miscellaneous cloud,” watched the launch with professional interest. Within nine seconds, the AI returned its first question: “Please provide supporting documentation for the $1.8 million variance.”

“This is exactly what we wanted,” said Vellum, speaking from a conference room called The Visionary Ledger. “A tireless partner that can eliminate manual work.” He then forwarded the question to the controller, who forwarded it to accounting, who forwarded it to AP, who forwarded it to a former employee whose auto-reply said she was hiking somewhere with no cell service.

The automation roadmap encounters source data

The project began after Vellum attended a breakfast panel hosted by the Institute for Frictionless Enterprise, where a consultant demonstrated an AI-generated dashboard showing revenue rising in a smooth diagonal line. No one at the event asked where the numbers came from, which made the demonstration unusually efficient.

Back at Vellum’s company, Northstar Synergies & Light Industrial, the AI was granted access to the ERP, the expense platform, the shared drive, three budget models, and a folder named FINAL_final_USE_THIS_ONE_v6. It reviewed 14,000 transactions, identified 2,903 as requiring judgment, and politely noted that the word “accrual” appeared in email threads with five different meanings.

“I can draft a reconciliation, but I cannot determine whether ‘we’ll sort it next month’ constitutes evidence.”

This was interpreted by leadership as a temporary configuration issue. A cross-functional task force was formed, comprising the CFO, the VP of Technology, a strategic transformation director, and one senior accountant invited to “ground the conversation in the process.” The senior accountant brought a list of 63 open items. The transformation director renamed the list an “exception-management opportunity set.”

AI proves remarkably capable of locating the same bottlenecks

By noon, the system had independently rediscovered several established laws of finance operations. It could not match a payment without a remittance. It could not validate a prepaid balance without a schedule. It could not determine whether a vendor invoice belonged in the current period when the business owner had written “not sure, maybe capitalize?” in the approval field.

The AI also discovered that the company’s fixed-asset register contained an item labeled “Big Printer,” acquired in 2014, with a net book value of $17.42 and an assigned useful life of “still working somehow.” When asked to close the asset subledger, the system requested an inventory count, disposal documentation, and confirmation that Big Printer remained in service. The facilities team replied with a photo of a printer covered by a jacket.

“It’s asking a lot of questions,” said Vellum. “Can we train it to be more strategic?”

The company’s finance recruiter, called in to advise on the future staffing model, explained that this was already the strategic part. “The system is not refusing to close,” she said. “It is identifying the work that must happen before someone can responsibly say the numbers are closed.” The room fell silent, except for the AI, which had opened a ticket requesting the signed lease amendment.

A new role emerges: Director of Explaining Context to Software

The company then posted a role for a Director of Autonomous Finance Enablement. The job required 12 years of progressive accounting experience, deep ERP expertise, advanced data engineering, prompt architecture, procurement negotiation, change management, and the ability to explain to a language model why the CEO’s personal conference hotel charge should not be recorded to “customer success.” Compensation was described as competitive, subject to the successful candidate demonstrating an ability to build the compensation model.

Interview candidates were asked to complete a short take-home exercise: reconcile 18 bank accounts, assess revenue-recognition implications for seven contracts, summarize the close calendar, and propose an AI governance framework using a provided dataset. The dataset was a screenshot of a pivot table. Candidates who asked clarifying questions were told they might not thrive in ambiguity.

• The first interview assessed whether the candidate had “an AI mindset.”
• The second asked for a walkthrough of their month-end close checklist, because no one could find the company’s.
• The third was a panel with stakeholders who each described a different chart of accounts.
• The final round tested executive presence by asking the candidate to explain why an automated journal entry still needs an owner.

The close is ultimately completed by an ancient technology

On the fifth business day, after the AI had issued 411 requests for evidence and a color-coded risk memo about unapproved manual entries, the accounting manager closed the books. She did so using a spreadsheet, a folder of PDFs, two calls to operations, and the accumulated institutional memory of having seen the same problem last April.

Vellum celebrated the result as a successful human-machine collaboration. The AI had not posted the final entries, but it had generated a detailed summary of the items preventing it from posting the final entries, which finance described as “directionally useful.” A steering committee is now evaluating whether the system can take over the quarterly forecast, pending receipt of the latest headcount plan, sales pipeline definitions, pricing assumptions, approved hiring requisitions, and an explanation of why every department’s forecast begins with the phrase “we should probably be conservative.”

The joke is not that software asks for support. Any tool that handles financial data responsibly eventually runs into the same fact: faster analysis does not create approvals, documentation, ownership, or judgment. It can make a strong accounting and finance team more effective. It cannot close a process that the business has declined to finish.