Takeaways

Should AI do this, how much do we check, and whose name is on it? Six questions, one decision, before anyone acts.

Pros and cons hand a model half the picture. Tell it what you want and what worries you first.

Name what the label measures and what your decision needs. If they are different properties, the decision is resting on the wrong evidence.

Before the renewal: what we use it for, what could replace it, and whether we have the ground to build on. If not, renegotiate.

Name the decision, name the number that flips it. If both branches read the same, the pilot was never measuring anything.

One spend you approved, one you refused, the same harm on both sides. The algebra brackets the price you say you don't have.

Let the agent propose for at least fourteen days and change nothing. Log the regrets, run the checks, then recommend one notch, or don't.

Six questions about how an AI product changes after you buy it. Anything still unanswered at the deadline, and it isn't ready.

The cheapest test at each end, run twice, before anyone digs up the middle. It cancelled a trench; it will cancel a rebuild.

Compress the output to one plain page and count what is left. The formatting was doing more work than the claims.

Four checks for any AI answer that cites something, and a rule for when it's still only a lead.

Outputs a week times honest minutes to verify one, against the minutes we really spend. One number, the one we were avoiding.

The dashboard says green. Feed the workflow an empty source on purpose and see whether it still does. If so, it never checked.

Every recurring commitment, one row each, with the two columns nobody ever fills: what would end it, and who is checking.

Every episode carries two signals and two subtractions. Fifteen minutes with your last month of work finds yours.

One row per thing that acts without a human deciding first, ending in a named person or a visible blank.

For every place a human is supposed to be able to say no: have they ever, and what happens if they don't press?

Name the mechanism that enforces one limit you rely on, and the verdict writes itself.

List every place an AI speaks to outsiders for you, then write down who answers for it.

One row per AI default in force, ending in the name of whoever set it. Most organisations can't fill the column that matters.

Every AI obligation with a date on it, one row each, ending in a named owner and the qualified person who checked it.

Three surfaces in, one list of where they describe different people out. The model finds the drift and rewrites nothing.

One row per handoff between specialists. Write down what each one protects before anyone collapses the chain.

Ten people whose decision has to change, by name. Ten roles and zero names means you have an org chart, not a plan.

Every AI threat we face, placed three times: as it's felt, as we're responding, and as the evidence says. The gaps are the work.