Chapter 8 · Infrastructure, Incentives, and Real Limits

Extended content

Liquid part: this text keeps growing after the printed edition.

Three services, one sheet

Write down the three AI services your work depends on most. Under each, four lines: what the provider can change without asking your permission, with how much notice, where the data you hand it lives, and what you would do tomorrow if it stopped working for a week.

Nearly always there is one line you can't fill in. That is your task for the next conversation with whoever signs the contract, and the first step toward having a second model tested on your critical tasks.

Add a fifth line, which isn't about the provider but about you: of what you've automated with that service, what do you do because it adds something and what do you do because it costs almost nothing today? Cross out the second kind in pencil. The day the price goes up or the usage limits shrink, you will already know what disappears without anyone missing it.

A case: Carmen and the five o'clock proposal

Carmen runs operations for a small chain of veterinary clinics. One Friday at five she receives two things. A proposal from a vendor promising to handle "most" appointments and inquiries with no human involvement. And a message from a colleague with a similar headline and a "we need to move now."

Carmen takes out the filter. The proposal shows a demo with one clinic and clean data; her chain runs three different scheduling systems. The error rate is nowhere to be found. In the contract, responsibility stays with the client. The vendor charges by volume, so it pays the vendor for her to believe in volume. Last question: if the claim were half as spectacular, would she decide the same? No. She wouldn't sign the full rollout; she would ask for a pilot at a single clinic and first measure how things work today.

Carmen hasn't rejected the tool. She has changed the order: baseline first, then the pilot, then the decision. The vendor, if its product is good, can wait three months. If it can't wait, that is information too.

A week of inbox

For five days, every piece about AI that reaches you (an email, a message, a headline, a video) gets logged with two facts: who produced it and what they live on. On Friday, count how many come from someone selling you something, even if only attention.

Add a third column with what each piece wanted you to do: buy, fear, rush, or understand. Few will land in the last one. The result is a census of beneficiaries built from your own sources, and it is worth more than any generic list because you made it from the noise you actually receive.

Questions to work with the filter

• Who benefits from my believing this, and how do I know?
• Which part of my use of AI would still make sense if the price multiplied?
• Do I have a second model tested on my critical tasks, or only the intention of having one?
• Which function of my work would I declare today, in writing, as "this does not get automated"?
• What decision did I make this quarter because "everyone else is already doing it"?

A three-minute habit

When a claim about AI changes your mood, stop before you forward it. Run it through the filter's five questions with a timer in your hand: conditions, error rate, accountability, beneficiary, decision. If after three minutes your decision hasn't changed, you have saved yourself an afternoon of unease. If it has, you've found something to verify before acting.

Over time the habit leaves you a quiet advantage: you stop being the echo of the latest news. Your team notices before you do, because in meetings your questions begin to arrive ahead of your opinions, and questions make everyone else look at the data.

Where to next

With the board in view, what remains is bringing it down to the ground you earn your living on: what is really happening to jobs, teams and careers while all this moves. That is the matter of Chapter 9.