What did it produce, and what did it cost, last Tuesday?

In August I switched off every AI worker in our business. Not because one of them had done something wrong. Because the credit card bill and the output had come apart, and I could not put them back together.

That is the failure mode nobody warns you about. It is not the machine doing a bad thing. It is the machine doing a great deal of something, invisibly, on your card.

So here is the question I now ask about any AI tool, ours or anyone else’s, before it gets near a business:

What did it produce, and what did it cost, last Tuesday?

If the answer takes more than a minute, you do not have an AI problem. You have a definition-of-done problem. Those are cheaper to fix.

Why the question works

“Last Tuesday” is specific enough that nobody can answer it with a slide. It forces three things into the open at once:

• A deliverable. Not “it helped with quotes”. Which quotes, in what state, sitting where.
• A cost. In dollars, for that day, not a monthly plan figure divided by thirty.
• A record. Somebody has to have written it down. If nobody did, nobody was watching.

Every tool vendor can tell you what their product can do. Very few can tell you what it did, for one business, on one day, and what it cost. The gap between those two sentences is where the August problem lives.

What was actually wrong in August

Not the models. Three plain operating problems:

1. No definition of done. Workers were given goals, not deliverables. A goal can be worked on forever.
2. No cost line per job. We tracked spend per month, not per task. A month is too late.
3. No “stop and ask” rule. When a worker hit an ambiguous decision, it tried harder instead of asking. Trying harder is expensive.

None of those is a technology problem. All three are the kind of thing a decent operations manager would spot in a new hire inside a week. We just had not thought of the crew as something that needed managing.

The three rules we run under now

Every job ends in a named thing, or it is not a job. A quote. A reply. A report. A page. When the crew is asked to “look into” something, the first thing it has to do is say what it will hand back.

Every job carries a cost. Spend is read per task from the provider dashboards, not from the monthly invoice. The number sits next to the deliverable, so you can see in one glance whether a job was worth doing.

Every decision that belongs to a human stops the work until a human answers. And the answer is recorded in that person’s own words, then applied everywhere. There is no “always allow”. We removed it on purpose, after reading the 2025 incident record: hidden instructions in a web page, a support ticket with commands in it, an approval nobody looked at twice.

Since the reset, the same crew has done a year of our GST preparation in a day, built the research behind our newsletter, and run most of our operations. The difference is not smarter AI. It is that I can now tell you what each job cost and what it produced.

Try it this week

If anyone in your business is using AI tools in anger, ask them the question. Not to catch them out. To find out whether the thing is being managed or just paid for.

Three outcomes are possible:

• They answer in a minute. Good. You have an asset. Make sure it is written down somewhere other than their head.
• They answer in ten minutes, with some digging. You have a useful tool with no record. An afternoon with a spreadsheet fixes it.
• They cannot answer. You have the August problem. It is fixable, and it is mostly a definition-of-done problem.

If you want the same three rules applied to a workflow in your business, that is what the Espresso package is: one outcome, a definition of done in writing, and a cost beside it.