A couple of years back a client cancelled a large project and got stuck with a pile of expensive material they no longer needed. I bought a chunk of it off them. Permanent reference cells, which is a part my work uses a lot of, and I got them at a price I'll probably never see again.

For a while that was a quiet advantage. We normally buy those around $750 each and sell them at $800 to $850 on a job. That's a $50 to $100 margin on a $750 part, 6 to 13 percent on something that's essentially passing through the business. Nobody gets rich on reference cells.

With the surplus on the shelf I could put them on a proposal at $350 and price the work around them. So I did. Sharper number, better odds, still made money.

Then I won a job on that pricing and went to pull the material, and the shelf was empty.

They'd been used. Another project had drawn them down, and nobody updated the inventory spreadsheet. The sheet said thirty. The building said none.

So I bought five at $750 and sold them at $350, because that's the number I'd already put in front of a customer and won with. $400 underwater on each one. $2,000, gone, on a job I was pleased to get.

Run that against the normal margin and it stops being a rounding error. At $50 to $100 a unit, each one of those replacements wiped out four to eight correctly-priced ones. Five of them erased what twenty to forty good sales would have earned. One row in a spreadsheet did that.

And the discounted labor I'd priced around that stock never shows up as a line item anywhere. There's no place on the P&L that says sold cheap because you thought you had parts.

I use AI on bid work now, and it does a lot of the heavy lifting. It reads the plans, counts what's on them, pulls what we charged the last five times something similar came through, checks what's on the shelf, tells me what has a lead time. The half day with highlighters and a ruler is mostly gone.

On this bid, it did every bit of that correctly.

It read the inventory file and reported thirty in stock, because the file said thirty in stock. It wasn't confused and it wasn't making things up. It was reading a document my own company maintains, and my own company had let that document go stale.

Now here's the part I want you to sit with.

Bad data did not produce an answer that looked wrong. It produced a cheaper answer. A tighter, more competitive, more attractive bid, delivered with confidence, that I was happy about. There was no moment where I squinted at the screen and thought that doesn't seem right, because everything about it seemed right. I won.

Failures that make you look stupid get caught immediately. Failures that make you win don't get caught at all, until the material doesn't show up.

That's the whole reason I'm writing this one down instead of the ten-minute-bid story everybody wants to hear.

For a while my answer to this was our AI policy, which says you own the output. That's true and I stand behind it, but a posture is not a control. Telling yourself to be careful is what you do instead of fixing something.

So here are the three rules that came out of it.

Every number in a shared file gets a date next to it. Not the file's date. The number's date. When was this row last true.

Anything older than thirty days gets flagged before it prices anything. Not blocked, flagged. It's allowed to say this count is from June, verify before you bid. $2,000 buys a lot of reminders.

Surplus never gets bid without eyes on the shelf. Regular stock gets reordered and self-corrects. Surplus is finite by definition, and the moment it runs out is precisely the moment the paperwork is most likely to be wrong, because it was never a normal part of the ordering rhythm in the first place.

None of that is technology. It's three sentences of process, and any one of them would have caught this.

The instinct after something like this is to trust the tool less. That's the wrong lesson and it'll cost you the upside without buying you any safety.

The tool read the file correctly. Trusting it less wouldn't have helped, because it wasn't wrong. What needed to change was the file, and the habit that let the file rot.

The uncomfortable version: AI didn't create this problem. That spreadsheet was already stale before I ever plugged anything into it. Somebody would have quoted off it eventually. All the AI did was act on my bad information faster and more confidently than a human would have, and hand me a number I liked.

Before it prices anything, ask where the number came from and when it was last true. If the answer is a file, go look at the file.

And tell me what yours is. Every business has one file everybody trusts and nobody updates. Mine was inventory. I'd bet you already know which one is yours.

Friday: the three ways this actually breaks, all of them mine, none of them the machine's fault.