Somebody asked me recently to list the times AI has let me down at work. I sat with it for a while, because I've been running most of my hard thinking through these tools for a couple of years now and the honest answer is that it hasn't let me down much.

But the times it has, they were all the same failure wearing three different coats.

Not once was it the model being dumb. Every single time, it was me handing it something bad and getting exactly what I handed it. That's a process problem, not a technology problem, and the difference matters because one of those you can fix on a Tuesday afternoon.

Here they are, in the order they'll happen to you.

This is the expensive one and it's the one you won't see coming.

I wrote about it Wednesday: an inventory file said I had 30 of an expensive part when I had none, and the bid built on that came out cheaper, sharper and more competitive than the truth would have allowed. I won the job and ate $2,000 buying the material back at full price.

The tool did nothing wrong. It read my file. My file lied.

What makes stale information the worst of the three is that it doesn't produce an answer that smells off. It produces a confident, well-organized, internally consistent answer with a number in it that happens to have stopped being true. You've got nothing to be suspicious of.

How you catch it: dates on numbers, and a rule about how old is too old. Ask where a figure came from before you act on it.

This one is quieter and it took me longer to understand.

I build meeting packets for big multi-contractor jobs by pulling from everywhere the project lives. Files, chat, email, daily reports. The packet comes out organized and readable and it saves me hours.

Then one time something went wrong on a job, and the packet didn't have it. Not because anybody hid it. It got posted in a channel that wasn't part of what I'd pointed the thing at. The information was sitting in the company, in writing, timestamped, ten feet away in the org chart. Just outside the lane.

And here's the nasty part. A packet missing something doesn't look like a packet missing something. It looks complete. Nothing is flagged, because nothing was seen. You go into a meeting confident and unprepared at the same time, which is a worse place to be than just unprepared.

How you catch it: every so often, ask out loud what it can't see. Not what it found. What's outside the fence. Then go check whether anything important lives out there. It usually does, and it usually got put there by somebody who assumed you were already in that channel.

Nobody warns you about this one, so I'll be the one.

I've built some big reference documents for the projects I run with AI. Long ones. And I got in the habit of grabbing all of them at the start of a session, on the theory that more background can only help.

It doesn't. Past a point the thing spends its energy hauling paperwork around instead of doing the work. It'll go read four long documents, three of which say overlapping things, and by the time it's done there's less room left for the actual job.

I didn't spot that myself. At the end of a session I asked how it went and got told, plainly, that three of the four documents I'd handed over contained the same information twice and two would have done the same job. That was a correction I needed and wouldn't have gone looking for.

The fix isn't a shorter briefing. It's a briefing that points instead of carries — a short front page that says where things are, so the detail gets pulled only when it's needed.

How you catch it: ask at the end of a long session what you gave it that it didn't need. It will tell you. Nobody asks.

Garbage in, garbage out has been around since about 1957 and you've heard it your whole working life. It's true, and it's so worn out that it doesn't change anybody's Monday.

What's useful isn't the slogan. It's knowing the three specific shapes, because each one has a different tell and a different fix:

Stale gives you a confident wrong number. Fixed with dates.

Absent gives you a complete-looking gap. Fixed by asking what's outside the fence.

Excess gives you a slow, wandering session. Fixed by pointing instead of carrying.

Every one of those is a filing problem. A documentation problem. A which-channel-did-you-put-that-in problem. They are all failures of how a company writes things down, and I'd argue every one of them was already costing me money before there was any AI in the building to expose it.

That's the part I'd want you to take away. These tools don't create the mess. They act on it faster, and they say it out loud.

The next time you get an answer you don't like, don't ask what's wrong with the AI. Ask which of the three it was: stale, missing, or too much. It'll be one of them, and you'll know what to go fix.

And all three are headed off by the same boring hour of writing your business down. If you haven't done that hour yet, the Context Starter Pack is free and it's that hour, laid out for you.

Reply and tell me which of the three you'd bet is happening in your business right now. I'll tell you what I'd check first.

Monday: the trick that lets a project survive a long build, and why the handoff note is written for you, not the machine.