Am I certain? The question that keeps CFOs in charge of AI

Ever walked into an important meeting with the gut-wrenching feeling that the numbers you’re about to present might not be entirely accurate?

That’s the sort of situation we CFOs are conditioned to have 3am nightmares about.

Some in other professions might be relaxed about presenting their colleagues with ballpark figures in the knowledge that they’re unlikely to be challenged. But we CFOs are used to being personally accountable for every number our departments produce. That’s why many are understandably wary of fully embracing AI.

I suspect most people know two things about AI. The first is that its evangelists talk about it eliminating vast swathes of routine work. The second is that large language models have sometimes had a shaky relationship with accuracy.

Can you achieve those efficiencies without the risk of carrying the can for figures that are just plain wrong? I’m confident that you can – as long as you, the flesh-and-blood CFO, are armed to ask the right questions and remain in charge.

Examining the foundations

In a survey of 250 finance leaders conducted for iplicit in 2025, more than eight out of 10 said they’d partially or fully adopted AI. But a closer look suggested a less confident embrace of the technology.

For one thing, there was a disparity between how CFOs and their financial controllers perceived things. While 51% of CFOs believed they had fully adopted AI, only 19% of FCs said the same. And although 83% of respondents said they had adopted AI fully or partly, only 53% had a formal framework for its safe use. That implies a lot of experimentation is going on without guardrails.

Just as concerning is that, despite this apparently wide adoption of AI, 42% of respondents still cited manual data entry as one of their most time-consuming tasks. It seems AI is being used in departments that haven’t yet automated basic transactional work. Bolting AI onto inefficient, error-prone systems is like putting a sticking plaster on a broken leg.

AI in finance is only as good as the data beneath it. It works best on clean, well-structured data held in a single, unified system, where the numbers are current and trustworthy. Layer it over the limited structure of entry-level software, or over data exported from an ageing on-premises system, and the efficiency gains shrink. Worse, it can automate and amplify the errors and workarounds that are already there.

Probability versus certainty

Sometimes, AI will confidently assert something that’s spectacularly wrong. That might be amusing when it’s telling you that glue makes a good pizza topping, as a Google AI Overview once advised. It’s less funny when it’s arming solicitors with fictitious case law or providing CFOs with reports based on inaccurate assumptions.

These gaffes happen because large language models are probabilistic. They work by predicting, many times per second, the next word (or “token”) in a sequence. That’s very different from the deterministic technology that ledgers depend on, where fixed rules apply and the same inputs always produce the same outputs.

In finance, AI works best on top of a deterministic system of record, not in place of one. The ledger holds the figures you sign your name to and file, calculated exactly, controlled and fully auditable. AI does the work that suits probability and pattern recognition: forecasting and FP&A, matching bank transactions, automating payables, spotting anomalies. It won’t always be right, which is why anything it prepares for the ledger still passes through the system’s rules, controls and approvals before it’s posted. And because AI isn’t calculating the figures you report, there’s always a single source of truth to check against and fall back on.

The other thing to remember is that tokens cost money. You don’t want to be burning them on tasks that could be done by traditional automation, such as accruals, prepayments or depreciation journals. Look for a system that triages your requests and applies the best tool to each – because aside from inaccuracy, the thing that makes us CFOs most anxious is uncontrolled spending.

The key question about AI

For all its time-saving and data-crunching potential, a CFO still has to ask “Am I certain?”

AI can do things instantly that would once have entailed countless hours wrangling data in spreadsheets. But a CFO signs the accounts to testify that they’re correct, not that they were produced quickly – and we put our reputations, and perhaps our careers, on the line every time. I love the benefits AI brings to me, but I choose where it’s deployed and, equally importantly, where it isn’t.

For more resources on using AI with confidence, click here.

Comments are closed.