Ask an AI to add up a column of figures and it will give you an answer. It will sound sure of itself. Most times it will be right, and some times it won't, and nothing in the answer tells you which time is which.
For a blog post that is a small problem. For a finance team, a builder pricing a job or a business owner signing off on a report, it is the whole problem. So I made one rule, and I build it into every AI system that touches a number.
A confident wrong number is worse than no number
When a person makes a mistake in a spreadsheet, you can usually find it. You follow the formula back and you see where it went wrong. When an AI makes a mistake in a sentence, there is no formula to follow. The figure just sits there, written in clean English, looking exactly like a correct one.
That is why "the AI is usually right" is not good enough for anything a business will act on. Usually right means that some day a wrong total goes into a report, a quote or a tax return, and nobody can say where it came from.
The rule I build in
The engine computes, and the AI explains.
On a platform I built recently, businesses use AI assistants that prepare real work from their own finance and operations documents. A rule based engine reads the client's files and computes every number. The AI never calculates anything. It only explains figures that the engine has already checked, every claim cites the exact quote from the client's own document, and a person approves before anything happens.
So when the assistant says a figure looks wrong, you can see the number, where it came from and the line in your own file that it rests on. You are not trusting the AI. You are reading its explanation of something you can check.
Estimator: a bid you can defend line by line
I used the same rule in Estimator, a construction estimating tool I designed and built that runs inside Claude. The brief came from a pilot for general contractors, pool builders and landscapers, and it set a hard line: the AI never does the arithmetic.
So the AI drafts and explains, and tested code does every calculation. Waste is applied once, packs are rounded up, labour hours, markup and margin are all worked out in exact decimal arithmetic, so a price never drifts by a few cents because a computer rounded it the wrong way. Every number on the estimate shows where it came from, and the pricing engine passes 19 of 19 tests built from the worked examples in the brief.
That job went to someone else in the end. I kept the prototype live in Claude, because the rule was worth proving on a real brief and I'd build it the same way again.
Even a writing tool needs the rule
You would think a tool that writes social posts has no maths in it. My Social Agent proves otherwise.
It reads last week's posts and works out which ones performed best, and that part is arithmetic, so plain code does it. Code also checks the banned words, the required themes and the length for each platform, because those are exact rules and a model could miss one. Claude gets one call to draft next week's posts, because writing needs judgement. Then a separate Claude call scores each draft for voice, and that call never saw how the draft was written, so it can't grade its own work.
A typical run makes at most 6 Claude calls. Nothing is published until a person clicks approve in Slack.
What the rule buys you
But the rule is not only about being correct. It changes what the system costs and how it behaves.
Code gives the same answer every time, for free. Fewer AI calls means your bill is predictable, and a month with more work doesn't surprise you. And when something does go wrong, you can find it, because the numbers live in code you can test and the words live in an AI call you can read.
Before you put AI near your numbers
If you are hiring someone to build an AI assistant for your finance, pricing or operations, ask them one question. Which numbers does the AI calculate itself? If the answer is "it handles everything", walk away.
The right answer is that the AI calculates none of them. It reads, drafts and explains, and tested code does the sums.
Have a process where the numbers have to be right every time? Book a 30 minute call and walk me through it. I'll show you which parts belong to code and which parts the AI can take.