Before this document, decision, or deliverable moves forward, answer these four questions in writing — plainly, no hedging: 1. Who is the named person taking ownership of this — by name, not "the team," not "AI-assisted"? 2. What did that person actually check themselves, and how? Name the specific thing they looked at. 3. What is this person accountable for if it turns out wrong — and do they know that going in? 4. If someone asks "how do you know this is right?" six months from now, what is the one-sentence answer — and does it name a person? If you can't answer all four without hand-waving, this isn't ready to ship. "The AI said so" is not a sign-off. A name is.

Run this on any AI-produced claim, citation, or assessment before anyone relies on it — three questions, in order, out loud if you're in a room: 1. SOURCE — What is this actually reading? Point at the specific document, record, or data source — not "the general body of knowledge." 2. AUTHORITY — Is that source the one that actually governs this decision? A source can be real and still be the wrong one. 3. OWN-EYES — Has a human who knows this area actually read it — not skimmed the summary, read the thing itself? If any answer is vague, stop. A control that can't answer these three isn't a control — it's an opinion with a timestamp. Budget three minutes per document; it's cheaper than the alternative.

Before starting this piece of work with AI assistance, write down the boundary — don't discover it mid-task: 1. What is the AI allowed to decide on its own, with no review? List it specifically. 2. What must a human touch before it moves forward — and who is that human, by name or role? 3. Where does the handoff actually happen — what artifact gets passed, and in what state? 4. What would tell you the boundary was drawn in the wrong place — too much human friction, or too little check? Write the boundary before the work starts, not after something goes wrong. The seam is the skill — not either side alone.

Capture this decision in three lines, written down somewhere a colleague — or a future AI session — could find it: CHOSEN: What did we actually decide? One sentence, specific enough to act on. REJECTED: What was the real alternative we didn't pick, and why did it lose? WHY: What has to stay true for this decision to still be right? If that changes, this decision is due for review. Undocumented intent dies with the conversation it was made in. Three lines now is cheaper than reconstructing the reasoning under pressure later.

Before rolling this out past a pilot, answer these — the pace should be set by trust, not by what the tool is capable of this month: 1. TIMEBOXED — What is the actual end date of this pilot, and what happens on that date whether or not it "feels ready"? 2. REVIEWABLE — Who looks at the results, and what specifically are they checking for — not "how did it go," but what measurable thing? 3. REVERSIBLE — If this pilot fails, what does rolling it back actually involve — and have we confirmed that's true before starting? 4. The tool sets the speed it CAN go. What sets the speed it SHOULD go here — whose trust are we actually pacing to? If you can't answer "reversible" with confidence, you're not running a pilot — you're already committed.