The high-stakes AI checklist: six steps before you act on any answer

Treating every AI answer with suspicion is exhausting and unnecessary — most answers are low-stakes, and speed is the point. The discipline that actually protects you is knowing which answers are not low-stakes, and having a fixed routine for those. Here is ours, in six steps. It takes minutes, and it’s built for the moments where being wrong costs real money or real rights.

Step 1: notice that you’re in a high-stakes moment

The trigger list is short: the answer involves money above what you’d shrug off, a signature (lease, severance, settlement, contract), a deadline with legal or financial force, your rights (claims, disputes, immigration, employment), or your health. If any apply, the rest of this checklist is mandatory. The most common failure isn’t skipping the checklist — it’s not noticing the trigger, because the answer arrived in the same casual chat where you plan dinners.

Step 2: split the answer into facts and advice

Every consequential answer mixes factual claims (“this notice type has a 30-day response window”) with judgment (“you should respond with X”). Ask the model to list its factual claims separately. Facts can be verified; advice can only be weighed — and advice built on an unverified fact is worth exactly as much as the fact.

Step 3: make the model state its uncertainty

Ask for its confidence and for what would change its mind. Remember the caveat that applies to every model, everywhere: self-reported confidence is not measured accuracy. You’re not collecting a guarantee — you’re collecting the model’s own map of where its answer is soft, which becomes your verification list.

Step 4: independently verify the load-bearing facts

Take the two or three claims the whole answer stands on and check them at the source: the actual notice, the actual policy document, the agency’s own page, the contract’s actual clause. Not another AI summary of the source — the source. Rules also differ by state on many legal questions, so favor sources specific to where you live, and treat any answer that speaks in confident generalities about “the law” as unverified until checked.

Step 5: get an examination the author didn’t write

For a genuinely consequential answer, self-checking has a ceiling: the model reviewing the answer shares the blind spots of the model that wrote it. Second Opinion exists for this step — paste the question and the answer, and models from four rival labs cross-examine it, returning an action band (act / verify first / don’t act) computed from real cross-lab agreement, plus the specific claims any referee challenged. It’s $9.99 once for 30 days. Use it on the answers that pass Step 1, not on everything.

Step 6: know when this becomes a professional’s job

Some answers, verified or not, precede decisions that need a licensed human — large sums, litigation, diagnoses, anything where the downside is irreversible. The checklist doesn’t replace that moment; it makes it cheaper and sharper, because you arrive with the facts checked and the questions framed.

One place to start: the free tools are built for several of these exact moments — no account needed — and the Calibration Scorecard will show you, in two minutes, why your own confidence needs this checklist too.