What goes wrong
Every line in this chapter cost me real time before I wrote it down.
What this does for a shop
The first three chapters get something built. This one is about not getting burned by it. A chat program, a phone agent, a bookkeeper wired into your books, all three will tell you they did something correctly. That report, on its own, is worth nothing.
Here’s what that looked like for me. I had a check that was supposed to run against my own files, and it printed a clean pass every time, looking exactly like success. It turned out to be checking zero files. Nothing was wrong because nothing was actually being tested. The result was the color I wanted to see, and that’s exactly why I didn’t catch it sooner. A result being the color you expected is not evidence that the thing you built produced it.
This applies to everything in the first three chapters, not just to code. The nameplate photo in chapter 1, the booked call in chapter 2, the reconciled bank balance in chapter 3, all of it comes out of something that sounds sure of itself whether it’s right or not. Nothing here tells you to trust it less across the board. It tells you what to check, so you can tell which parts have actually earned your trust and which parts just sound like they have.
What it costs, honestly
Here’s the honest answer to whether any of this saves you time. The writing takes seconds. The checking is the real cost. Every habit below exists to make the checking short instead of long.
What you need before you start
- A place to write down open questions and unresolved claims, a notes app is fine
- The willingness to say “I don’t know yet” out loud, to a customer or to yourself
- An agreement with yourself that no password or bank login ever goes near an AI chat
- One plain question you ask before you act on anything it tells you: would I bet a real job on this
Step by step
- Before you repeat any number, whether the AI said it or you did, label it in your own head: done, building, or hoping. Say which one before you say the number out loud.
- When an AI builds you a rule, a check, or an automation, ask it to also build a test that fails when you delete the rule. If the test can’t fail, it isn’t proving anything.
- Ask every check to print out what it actually checked, not just pass or fail. A clean pass over nothing still looks clean.
- Go look yourself at which file, which step, or which path actually produced the result you’re looking at. Don’t take the report’s word for it.
- When it hands you a number or a claim, ask it for the source, ask what it did not check, and ask it to repeat the number back from the actual document, not from memory.
- Anything about tax, licensing, employment, or money gets a second, independent check before you act on it: a second AI, and then a human professional.
- Never put a password, bank login, or card number into any AI chat. If you paste one by mistake, go change it right away, don’t just delete the message and move on.
- Once one of these habits works for you, write it down somewhere you’ll actually reread it, the same way you’d write down a fix that took you three tries to find the first time.
What goes wrong first
- You accept the first answer because it sounds confident.
- You skip the second check on a tax or legal question because the first answer seemed obvious.
- You paste something sensitive into a chat out of habit, the way you would to a coworker, and don’t think to go change it after.
- You build a check, it passes, and you never ask whether it was checking anything real at all.
- You build the habit for one tool and skip it for the others, because that one’s been reliable so far, and the one you skipped it on is the one that gets you.
- The result looks like the answer you expected, so you stop looking, and that feeling is not evidence of anything.
What I’d do
Write the three words, done, building, hoping, somewhere you’ll actually see them, and say one of them out loud every time you’re about to repeat a number to a customer, a supplier, or yourself. It costs nothing, and it catches most of this before it costs you anything.
Back to the front page · Next: Your CRM, and the shop’s memory