The following is a guest post from Mohamed Chaudry, FCCA, a group and operating CFO and board adviser with more than 25 years of experience. Opinions are the author’s own.
A recent survey by planning software firm Board found that nearly half of CFOs would follow an AI recommendation over their own gut feeling. I understand why. The output is quick; it's well written, and most of the time it's right.
A year ago, I'd have been in that half. And I build these systems.
Last year I started building an AI system that takes a founder's numbers and produces a business plan, a pitch deck and an investment committee review. It runs in stages, and each stage passes its data to the next. I was testing it on four or five companies at a time, all in different sectors.
One business plan came back looking excellent. The sections had real depth and the analysis was detailed. Honestly, it was better written than plenty of plans I've reviewed from actual founders. I skimmed it and moved on.
It was only when I sat down and read it properly that I realized it wasn't the company I'd put in. The numbers didn't match either. A mapping error between two stages meant the data never arrived. So the AI did what it was built to do and wrote a business plan. It just invented a company to write it about, despite a hard instruction not to make anything up.
What stays with me is my first reaction. I assumed the mistake was mine. I wondered whether I'd misremembered the data I'd fed in, because the plan was so convincing. After 25 years as a CFO, I doubted my own memory before I doubted the machine.
AI doesn't fail like a spreadsheet
That's the problem every finance leader using AI now has.
When a spreadsheet goes wrong, it usually shows. A total doesn't tie, a formula throws an error, a number is out by a factor of 10. Experienced people spot it because they've seen it a hundred times.
AI goes wrong differently. When it doesn't have what it needs, it fills the gap, and it fills it well. Ask it to expand a 300-word section to 600 words without giving it more to work with, and a lot of those extra 300 words will be made up. They'll read just as well as the rest.
Telling it not to do that doesn't hold. I learned that the hard way. In the end, it took independent verifiers, a corrector and checks against the original data, plus hundreds of test runs before I was comfortable. I still review the output myself before it goes to anyone.
That isn't an accounting problem. But it lands on whoever puts their name to the numbers.
What this means for the CFO's job
For most of my career, a CFO's credibility rested on technical accounting. Can you trust the numbers? Do they reconcile? Can you defend them to a board, an auditor or an investor?
Those questions haven't gone away. What's changed is what produces the numbers. Forecasts, variance commentary and board materials are increasingly drafted by AI, and most CFOs can't see how that output was put together.
I'm not saying CFOs should start writing code or run IT. I mean the CFO now needs to understand AI the way good CFOs have always understood operations: how it works, where it breaks and what it looks like when it has quietly gone wrong. In that sense, the job is moving closer to a CTO's.
The hiring market hasn't caught up. A Datarails analysis of U.S. job postings found only 27% of CFO listings in January asked for AI familiarity, no higher than a year earlier, even as more finance roles below them did. We're asking teams to use AI without asking the people who sign off on their work to understand it.
There's a real risk in that gap. AI makes it easy for someone with limited experience to produce something that looks expert. Anyone can prompt a polished forecast. If the person reviewing it can't tell good analysis from convincing fiction, it will get through.
Questions to start asking now
You don't need to build an AI system to close that gap. You do need to ask better questions, and expect real answers.
Of your team:
- Which parts of this pack did AI draft?
- Where did this number come from, and can you show me the source?
- Who checked it, and was that a different person or process from whatever produced it?
Of your vendors:
- What does your system do when data is missing? Does it stop and flag it, or fill the gap?
- If I run the same inputs twice, do I get the same answer?
Of a CFO candidate:
- What have you actually built or deployed with AI, and what went wrong?
- How would you know if an AI-generated forecast was wrong?
And of yourself, before you sign: if the board asked about any number in here, could I explain where it came from?
There's no clause in a set of accounts that says the AI got it wrong. It's still the CFO's neck on the line.
The question I can't answer yet
This leads me to a question I keep coming back to.
If CFOs now need to understand how AI is built, and AI is doing more of the finance work itself, why couldn't someone who already understands how AI is built learn enough finance to do the job?
I don't think the answer is obvious. Accounting judgment is what told me something was wrong with that business plan. Understanding the system is what told me why.
The next generation of CFOs might be accountants who learned AI. Some might be technologists who learned finance. If you were hiring tomorrow, which would you choose?
