A founder turned his laptop round to show me a forecast his team had built that morning. Twelve months, three scenarios, charts already formatted. It had taken about a minute to produce.
I asked what was driving the revenue line in month seven.
Nobody in the room could answer. Not the founder, not the finance lead, not the analyst who had run it. The number was there, it looked reasonable, and it had arrived without anyone having to believe anything in particular.
That is the whole problem in one meeting. AI-built forecasts are now standard in businesses of every size. Board-credible forecasts are still rare. The distance between the two is where finance credibility quietly goes.
A forecast you cannot defend is an opinion with a spreadsheet attached
Over twenty years I have built forecasts on budgets from £10M to £200M+ and got them to 96%+ accuracy. The maths was never the hard part. The hard part was sitting in front of a board, being asked why the number moved, and having an answer good enough that the next decision could be made in the room.
AI is very good at producing the artefact. It is indifferent to whether anyone can defend it.
And a model gives you no signal when it is wrong. It does not hedge, it does not look uncomfortable, and it will restate a bad assumption in clean formatting with total confidence. A human analyst who is unsure usually shows it. That tell is gone.
The risk isn't that AI gets the forecast wrong. It's that it gets it wrong persuasively, and nobody in the room has a reason to argue.
So the checks below are not anti-AI. They are the same checks I would run on a forecast built by a graduate, a Big Four secondee, or myself at 11pm. AI just makes running them urgent, because the cost of producing a forecast has collapsed and the cost of believing a bad one has not.
The five checks
1. Trace the drivers
Take the top three numbers in the forecast and ask someone to explain, in one sentence each, what moves them. "Revenue grows 8% because these four contracts renew in Q3." If the honest answer is a shrug or a reference to the model, you have an output rather than a forecast. Outputs cannot be challenged, which means they cannot be trusted.
2. Stress the assumptions until something breaks
Ask what happens to cash if your largest customer pays thirty days late. Then ask what happens if two do. A forecast that only works in the good case is a plan for a business you do not run. Break it deliberately in private, before someone breaks it for you in a board meeting.
3. Check the joins
Errors hide where the P&L, balance sheet and cash flow connect. Profit that never becomes cash. A working-capital movement nothing on the balance sheet accounts for. Depreciation quietly double-counted. That junction is the first place I look and the last place a model is designed to police, because each statement can be internally consistent while the three of them disagree.
4. Re-run it
Change one assumption and refresh. If that takes days, you own a relic. A forecast you cannot rebuild in an afternoon is already out of date, and the practical effect is that nobody updates it between board meetings, so decisions get made against a document that stopped being true weeks ago.
5. Sanity-check against the business
When the model and your gut disagree, investigate. Do not overrule the model, and do not defer to it either. One of the two is wrong and finding out which is the most valuable hour in the process. In my experience the gut is wrong slightly more often than the model, and when the gut is right it is usually right about something the model was never given: a customer going quiet, a hire who has not started, a price rise everyone knows is coming.
What AI is genuinely good at here
Speed, volume and first drafts. Ten scenarios instead of three. Pattern-spotting across a data history no analyst would read end to end. Reconciliation grunt work that used to eat a week a month.
Use it for all of that. What it cannot do is carry accountability. When the board asks why the number moved, the model does not answer. You do.
Key takeaway: AI does not make a forecast trustworthy. The person who can defend it does. Trace the drivers, stress the assumptions, check the joins, re-run it, and sanity-check it against the business you actually see.
Where to start this week
Pick the forecast your board saw last. Take the three largest numbers in it and write one sentence under each explaining what drives them.
It costs nothing and it takes twenty minutes. If you can write all three, your forecast is in better shape than most. If you cannot, you have found the exact place the trust is missing - and that is a method problem with a known fix, not a software purchase.
Could your team explain, in one sentence each, what drives the top three numbers in your forecast?