The AI you can stake a forecast on

The AI you can stake a forecast on

There's a lot of either-or-ism when it comes to the role AI has in our businesses. People seem to think that either you embrace AI and destroy our profession, or you reject it and double down on the traditional ways of doing things. Artisanal accounting anyone?

Something that seems to be getting lost is the various flavours of AI. And because artificial intelligence in the form of free or paid-for GPTs is readily and affordably available to anyone with a web browser, many of us default to thinking that this alone is AI.

And sure, a year ago, I was a great proponent of dipping your toe in the water with whatever AI you preferred. But the time for that is over. We need to be more discerning about which flavour of AI we use, and why.

Recently, I met with an accountant who showed us the AI tool they had built in-house for financial planning and analysis (FP&A) using public AI. It was very cool and very clever. But then they said: “I don't fully trust the numbers, and I'm worried that the AI token costs are going to eat into any benefit I get from this.”

Close, but not quite there yet

What you should be asking

1. Does placing AI at the centre of FP&A exclude the people at the coalface? Doing this ignores vital insight from the people who can see what is happening and predict what is coming. It also erodes any ownership and accountability they have for implementing the resulting budget.

2. How helpful is AI’s reliance on historical trends and patterns? Think about the last 24, 36, or even 48 months. How much of this has been typical? Historical patterns, which is all AI has to go on, are no longer a helpful foundation for your FP&A.

3. Will the tool tell you what you want to hear? I've argued before that the most dangerous AI is the one that agrees with you. This risk is at its highest when you have built the tool yourself. You set the assumptions and frame the questions, and then the model reflects your own thinking with the polish and confidence of an independent answer. A budget that confirms what leadership already expected is not the same as a correct budget, and nothing will tell you the difference.

4. What happens to your data? You wouldn’t typically broadcast your clients’ income statements for all to see. Can you be sure you aren’t doing this by using a public GPT, which uses uploaded data to keep learning and optimising?

5. What will the long-term token cost be? As soon as you move beyond the chatbot interface into more advanced work, you start paying for tokens as you use them. These ramp up very quickly, as the more you use the tool, the more you pay. And although the cost of tokens is, currently, decreasing, this drives more use, and so you spend more. This is not an all-or-nothing choice, though. You wouldn't put your most senior partner on data entry, and the same logic applies here: match the AI to the job, keeping the expensive, high-end models for the work that genuinely needs them and cheaper ones for the routine queries.

6. How do you audit your data in the AI black box? Take the large amounts of data we expect AI to process. How do we track if anything has been dropped, summarised, or misinterpreted?

7. Who is ultimately accountable? If something goes wrong, where does the buck stop? Your client? The junior staff member instructed to use the AI tool? Your leadership team? Your AI vendor? This is a space where regulation is lagging technology, and it needs to be addressed urgently as an industry.

The judgement was never the problem

The accountant I mentioned earlier hadn't done anything wrong. They'd just discovered the gap between AI that's impressive in a demo and AI you can stake a forecast on.

I've written before about guardrails and checkpoints, the boundaries that stop AI overstepping and the moments where a human has to confirm the route is right. The harder question is whether the tool you are using can enforce any of it. A checkpoint that relies on a busy person remembering to look isn't a checkpoint, it's a hope.

What I keep coming back to is that these questions aren't really about technology. They're about judgement and accountability, which have always defined good finance work. AI doesn't remove the need for any of them. If anything, it raises the bar, because now you also have to understand what the tool is doing, where its gaps are, and what it's costing you to run. The inclusion of coalface knowledge, the raised eyebrow when you see tidy historical patterns, the instinct to ask who's responsible when it goes wrong: those aren't obstacles to adoption, they’re essential for success.

So the question isn't whether to use AI for FP&A. It's whether the flavour of AI you are deploying actually fits the job. A free public GPT and a purpose-built, finance-grade tool are not the same thing wearing different badges. The difference is in what happens to your data, whether you can audit the work, what it costs to run, and who carries the can when it goes wrong.

Getting the flavour right

That accountant's instinct was right, by the way. Go ahead and tap the power of AI but make sure you can trust the numbers and justify the cost. Better asked before you commit than after the budget has gone to the board.

What was missing was a tool built with accountancy-grade scrutiny baked in. You can bring every ounce of professional judgement to the party, but if you are working with a system that has no audit trail and no place for a checkpoint to live, the tool will not be able to prove any of it happened.

Get the flavour of AI wrong, and you inherit every one of those problems. Get it right, and AI carries the workload you've always wanted to put down, while you keep your hands on the parts that were never the machine's to take.

As published AccountingWeb - July 2026