Before You Trust the AI With the Numbers
Somewhere in your business is a number that takes three seconds to produce and the better part of an hour to trust.
I have done this myself for twenty-odd years. A spreadsheet I built, that has never once been wrong, produces the figure in seconds — and I still walk the calculation by hand. Relying on a thing and trusting it are not the same. The calculation was never the slow part. Being sure of it is.
That gap — between an answer arriving and an answer you will put your name to — is the real reason most owners have not let AI near their routine numbers. Not speed. Trust. And when an owner tells me AI cannot be trusted, they are usually saying two things at once: it makes things up, and it cannot be relied on with the arithmetic. Both are true — when AI is used naïvely, and almost everything that turns a sensible owner off rests on that one word.
It is two objections wearing one coat
"It makes things up and it gets the numbers wrong" sounds like a single complaint. It is two, with different causes and — this is the part the noise skips over — different, well-established fixes.
The first is invention. Ask a general-purpose AI a question about your business and, if it has not been given your facts, it will answer from the average of everything it has ever read. It does not know it is guessing. It produces something fluent, plausible, and disconnected from your actual accounts — which is worse than an obvious error, because it looks right.
The fault is not in the model. It is that it was asked to recall something it was never told. The fix is to stop asking it to recall anything at all. Every figure is pulled from a system of record — the ledger, the bank feed, the payroll — and handed to the model as fact. It is forbidden from being the source. Give it the numbers and it reports them; deny it the numbers and it has to say it does not have them. The invented answer has nowhere left to come from.
The second objection is arithmetic, and it is a different animal. A language model predicts text. Asked to add a column, it produces the most likely-looking total — which is not the same thing as the correct one. Trusting it to do the sums in its head is trusting a very well-read stranger to total an invoice in theirs. Sometimes right. Never to be relied on.
The fix here is not a better model. It is to take the arithmetic away from the model entirely. Its job is to assemble the right inputs and call a calculator — a formula, a query, the spreadsheet, the ledger’s own engine — and the calculator returns the number. The figure is as accurate as the spreadsheet, because it is the spreadsheet. Calculation error stops being a risk you manage and becomes a category you have removed.
That leaves one real risk, and I would rather name it than have you find it. The model can still pick the wrong account, or read an ambiguous instruction the wrong way, and produce a clean, well-formatted, confidently wrong number. That is not invention and it is not bad arithmetic — it is misjudgement. It cannot be designed down to zero. But it is exactly the kind of error a finance function already exists to catch: numbers that don’t reconcile, a figure that has moved further than it should, a second pair of eyes on anything that matters. Visible, contained, checked.
Why the objection survives anyway
If the fixes are this well understood, you would think the objection would have died. It hasn’t, for two reasons, and neither is the reader’s fault.
The people selling AI roll the two failures into one, because "it’s accurate now" is an easier sentence than the true one. And the owner who tried it once — typed a question into a chatbot, got a confident wrong answer, put it in front of someone who mattered — learned the correct lesson from the wrong experiment. They concluded AI cannot be trusted. What they actually proved is that AI used naïvely cannot be trusted. Different finding. Same scar.
I have no interest in arguing anyone out of the scar. It was earned. The point is narrower: the thing that burned them is not the thing on offer.
The question worth asking instead
Once you separate the two, the question changes — and the new one actually has an answer.
"Can AI be trusted?" is unanswerable, because trust is not a property the model has or lacks. It is a property of how the system around it is built. The better question is plainer: "Where does each number come from, and who checks it before it counts?" Ask that, and a vague anxiety turns into a short list you can inspect. This figure came from the ledger. That one was computed by the spreadsheet. This judgement was flagged, and a person signed it off. Nothing was recalled. Nothing was invented. Nothing was added up by a machine doing sums in its head.
That is the foundation. Not because it is clever, but because everything else depends on it — work can only safely leave one person’s desk once you can say, of every number it produces, exactly where it came from and who stands behind it.
What I’m not claiming
Grounding only works if the systems behind it are clean — point a model at a messy ledger and three different definitions of revenue and it will report the mess back to you, accurately, which is its own kind of wrong. Handing the arithmetic to a calculator removes the model’s mistakes, not yours: a wrong formula in the spreadsheet is now wrong every week, faultlessly. And the honest objection — that a good controller with good software already does most of this — is fair. The difference is that the checking becomes explicit and repeatable instead of living in one person’s head. A difference of degree, not kind. I would claim it as exactly that, and no more.
The arithmetic is solved. The facts are grounded. What remains is judgement — which is where you wanted your people spending their attention, not on re-checking a number they should have been able to trust on sight. That is the hour you are buying back. If that is the version of all this you have been waiting for someone to describe honestly, future editions of this newsletter are the rest of the answer. And if you would rather not wait for it — you know where to find me.