Everyone's seen it. A confident answer with a wrong number in the middle. A reference that doesn't exist. A summary that's ninety per cent right, which in business can be worse than wrong, because nobody spots the ten per cent until it costs money.
So the reputation is fair, as far as it goes. But it's a reputation built on one way of using these tools, and it's the wrong way.
Why "nearly right" happens
Picture a brilliant new starter on their first morning. Clever, quick, keen to please. Now ask them to quote a job for one of your customers, from memory, without the price list, without the drawings, without anyone checking their work before it goes out.
They'll give you an answer. It'll sound confident. And some of it will be wrong, not because they're stupid, but because you asked them to guess.
That's how most people use AI. They ask it to guess, with nothing to go on, and nobody checking. Then they blame the tool for guessing.
Ask anything to guess and it will guess. The fault is in the asking.
You already know how to get it right
Think about your best people. They don't guess. They check the drawing. They look up the price. They read the certificate, not the sticky note about the certificate. When something doesn't add up, they say so instead of covering it.
That's exactly what makes AI right too. Not a cleverer model. The same discipline your best people already have.
Where nearly right was never an option
We've spent our careers where a wrong figure costs a fortune: data that had to be right first time, every time, a long way from anyone who could fix it later. When we brought these tools into that world, "nearly right" would have been laughed out of the room. So we learned to use them the way that world demands. Here's how:
- Everything is checked four ways, every time. Every price, date and total is cross-checked as a matter of routine before you ever see it: against your agreed prices, against the arithmetic, against the totals, against the rules. You never see the checking. The checks themselves take a few milliseconds.
- It answers from your facts, not the internet. Your price list, your certificates, your job sheets. It works from what's true in your business, not what's typical somewhere else.
- Every number shows its working. Where did that figure come from? Which sheet, which row, which document? If it can't point to the source, it doesn't get used.
- It says when it doesn't know. A date it can't read, a count that says "approx", a field that's blank: it flags them for a person instead of filling the gap with a guess. (Every example on our Check it out page does this.)
- It's tested on your past before it touches your future. We run it against jobs you've already done, where you know the right answer, and it has to match before it goes live.
- The expert signs it off. Your rules, your judgement, your final say. It works with people, not instead of them.
The one question to ask any AI tool
Whoever's selling it to you, ask: "Where did that number come from?"
If the answer is a shrug, a "the model worked it out", or a long word you don't understand, walk away. If it can show you the document, the row and the rule it used, you've got something you can trust with your business.
Ours always can. That's why we can make the promise we make: if it doesn't do what we agreed, you don't pay. Nobody offers that on "nearly right".
Right is the whole point
The tools are extraordinary. In the right hands, grounded in the right facts, checked the right way, they don't give you nearly right. They give you right, faster than anyone could by hand, every single time, and they tell you the moment something doesn't add up.
Nearly right is what you get when you ask a tool to guess. Right is what you get when you give it the facts, the rules and someone who knows the job. That's a choice. We make it every time.