Pressure-test one decision before you own it.
State the call, separate what AI can support from what a human must judge, and identify what must be true before you decide. Everything runs in your browser.
It produced the recommendation — which means you met its view before fully forming your own. It can refine that recommendation endlessly. It cannot decide whether the recommendation deserves your name.
It analyzed the evidence it was given. It cannot know what it was not given — the missing input is yours to name, and it is usually the one that matters.
It drafted the work, so the work looks finished. Looking finished and being decided are different states — the polish is exactly what discourages one more question.
It challenged your thinking — the strongest seat to keep it in. The moment it drafts the answer instead, the anchor flips and your independent view gets harder to recover.
Nothing yet — which makes this the cheapest moment you will ever have to form your own view. Write it down before AI enters. Afterward, it is gone.
Whether the upside of this person justifies the uncertainty you are accepting.
Which customer behavior you are willing to risk in exchange for margin.
How much uncertainty justifies committing capital that is hard to get back.
When the evidence is sufficient to act, rather than keep learning.
Which trade-off between capability, cost and human consequence you are prepared to own.
How much dependence on someone else’s performance you will accept in exchange for the gain.
How much uncertainty you are willing to own in exchange for moving now.
Not yet named — and that may be the finding. Until this sentence exists, every further round of analysis is polishing a call whose foundation nobody has stated.
Not yet named. A decision with no reversal condition is not a decision — it is a commitment to agree with whatever arrives next.
You. Then finish this sentence before the recommendation moves anywhere: “I am making this call because ___, despite uncertainty about ___, and I would reconsider if ___.” That sentence is the decision record almost nobody writes.
Someone else — so your job is not to pre-decide it. It is to make the reasoning visible enough that they can genuinely decide rather than merely approve.
A group. A call a group owns together is often a call nobody owns — ask which single name attaches to it when the consequences arrive, and get that said out loud.
Unclear — and that is the finding. Before this decision gets one more round of analysis, it needs an owner. Everything else on this card is secondary to that.
Before you own an AI-assisted decision
Are you ready to own it?
Then say whose call it is — out loud, to a colleague, before the recommendation moves. Not as a formality: because in six months, “who decided this and why” will be the only question that matters, and today is the cheapest day to answer it.
Good answer — it is the honest one more often than yes. Do one thing next, not five: take the unchecked line nearest the top of the list above and spend your next ten minutes there. If it is the assumption — ask your AI to make the strongest case that its own central assumption is false, then verify the most material claim yourself, independently.
Run your next consequential decision through the same six questions before AI touches it. The card does not change — your speed does. That is the routine; this page was just the first repetition.
No decision handy? Run the 90-second sample instead — you make a call, the information changes under you, and you see what moved your judgment.
AI can make the call. It can't take the fall.
Forty people. The same discipline. Their own decisions.
This took you two minutes, alone. The session runs it live with your team — people decide under pressure, watch what moves their judgment, and agree what AI may do, who owns the call, what reasoning stays visible, and when to escalate.