Where CFOs Draw the Line on AI Authority

AI can help the team close the books, but it won’t be the one explaining a bad number to the board. 

Here, the distinction matters as finance teams use AI across processes that eventually feed the financial statements. At Garner Health, CFO Jake Shuster is building toward public-company rigor, even though he stressed that a stock market debut isn’t imminent.

“We are not IPOing today or anytime soon,” Shuster told CFO Upside. The healthcare company, which helps businesses address healthcare costs, recently moved from QuickBooks to NetSuite and switched auditors to PwC as it builds more sophisticated finance operations.

Shuster shared how there’s plenty of AI in that plan. He said Garner, which is approaching 750 employees, has just four people in accounting and three in FP&A. “Everything from bank reconciliations to monthly close processes to building the reforecast to coming up with the right dashboards to review performance is all really being generated through Claude,” he said.

That’s a lot of ground for a small team to cover, and Shuster sees accounting expertise as essential to making the arrangement work. “They still have to have the expertise to understand what good looks like,” he said. “But then they need to take their expertise of what good looks like and translate it into really automated processes.”

Pretty Please, CFO?

Kenneth Paqvalén, CFO of Aptitude Software, which provides finance data and accounting technology, said finance chiefs need to spell out AI’s role before putting it to work.

“There is a difference between asking it to analyze data, make a recommendation, prepare a journal for a person to approve and act autonomously,” Paqvalén told CFO Upside. “Those uses carry different levels of risk and need different controls.”

He wants finance teams to be able to trace an output back to the underlying transaction and the accounting rule behind it. A process that can produce an entry, he said, should also have a way to handle an exception or reverse a mistake. “I’d want a human approval step for AI-prepared journals before considering more autonomous action,” he said.

Though Paqvalén’s company sells technology to finance departments, he said its own team still has work to do. Aptitude has begun evaluating its technology internally, and he would be reluctant to deploy AI where its scope was unclear or its output couldn’t be traced.

“The CFO remains accountable,” he said. “If AI produces a number and it’s wrong, it’s still my name on the financial statements, not the AI vendor’s.”

For Dean Quiambao, partner and Northern California market leader at Armanino, the questions extend to whoever sets up the agent in the first place. He says the promise of a faster close still comes with familiar responsibilities around who can change the system and who checks its work.

“Companies need to establish clear controls governing what AI agents can do, who can modify their behavior and how their actions are reviewed and approved,” Quiambao told CFO Upside.

An agent’s journal entry needs enough supporting information for an accountant to evaluate its accuracy. Someone also needs to keep track of changes to the technology producing it. “There should also be appropriate separation between employees who configure the AI tools and those responsible for approving the accounting outcomes,” he said.

Putting a person in the process creates its own assignment, too. That person needs to understand the calculation and have access to the information behind it. The business then needs documentation showing the review happened. “Ultimately, auditors need to be able to understand how financial decisions were made, whether the appropriate approvals occurred and whether the controls operated effectively,” Quiambao said.

Access Denied

The stakes become especially easy to understand when the agent gets near the bank account. Irina Berkon, CFO of Metallicus, which develops blockchain infrastructure for financial institutions, said existing ideas about payment authorization need to carry into AI-driven workflows.

“An AI system might be permitted to prepare or initiate a payment request, but that shouldn’t automatically give it the authority to approve and execute the transaction,” Berkon told CFO Upside.

She described an agent identifying an invoice that meets payment criteria. The transaction would still have to satisfy spending limits and approved recipient requirements, alongside the institution’s other controls. “For higher-value or unusual transactions, human approval or multiple independent authorizations should remain mandatory,” she said.

Adding another agent to check the first one’s work doesn’t automatically settle the oversight question, either. “We should also recognize that multiple AI agents do not necessarily provide independent oversight if they share the same underlying permissions or vulnerabilities,” Berkon said.

That warning complicates the promise of AI freeing up the finance team’s time, since adding another agent to review the work could still leave the same mistakes slipping through. 

Accountants Still Count: Shuster’s approach at Garner points to the people needed alongside that technology, whose accounting judgment becomes part of making the automation work. “We need accountants who are thoughtful about what would be an optimal process versus just doing the process,” he said. 

Designing that process includes deciding which actions require approval and making sure someone can explain the result afterward. Those responsibilities travel with the automation, however quickly the software finishes. Paqvalén put it plainly: “The controls need to be built into the accounting process, not left to a prompt.”