AI-Assisted Analysis Check

Treat what the tool gives you as evidence to weigh, not an answer to accept: what kind of claim each one is, what it rests on, what it assumed about your situation, and whether it moves the work forward. The judgment is the part you cannot hand over.

Practice, not research: CEBMa's rules for evidence-based AI use (2026) are a practitioner framework, and McKinsey's argument that judgment is the new bottleneck (2026) is consulting practice, not checked against the original. Mixed research: a survey of 319 knowledge workers found that higher confidence in generative AI went with less critical thinking (Lee et al., 2025; self-reported, a correlation). The general verification ladder is in 10.9; this page adds the evidence lens.

Use it when:

• An AI tool produced or shaped an analysis that will inform a real decision: a market estimate, an option comparison, a data summary or a "what does the research say" answer.
• The output reads well and you notice you are about to trust it because it reads well.
• A colleague hands you AI-shaped material and you need to know what it rests on before you build on it.

Skip it when:

• The tool only drafted text whose facts you supply yourself: the general 10.9 AI Output Verification Checklist is enough.
• You have not decided whether AI fits the task at all: start with 10.7 AI Task Map.

This check points the core move of evidence-based practice at a new kind of source. You normally trust an analysis partly because you know how it was made; with a generative tool, you did not see it being made. The answer is not suspicion but sorting. Take each claim back to what it is (fact, interpretation or the tool's own reasoning) and what it rests on, following CEBMa's rules, which are practitioner craft (CEBMa, 2026). When analysis becomes cheap to produce, the scarce resource becomes the judgment that reviews it (McKinsey Quarterly, 2026, consulting practice). Your own calibration matters as much as the tool's error rate: in one self-reported survey, more confidence in the tool went with less critical thinking (Lee et al., 2025, mixed evidence). A fluent text hides the errors that slower work would have exposed, so ease of reading is a warning sign.

That sets the order here. Step 1 builds checkability into the request. Steps 2 and 3 sort the claims and expose what the tool assumed. Step 4 keeps the judgment human on purpose.

Name the decision, then build these four demands into the request (see 10.8 Prompt Brief). If the output already exists, tick the ones it meets and ask again for the rest.

Strong answer: "Decision: whether to raise the minimum order for small cafes, owned by the bakery owner, April board. All four demands written into the prompt, including 'say when you cannot find a source'."

The decision this analysis feeds, and who owns it:

• ☐ Separate the claim types: ask the tool to mark facts, interpretation and its own reasoning, and to flag numbers it is unsure of.
• ☐ Ask for evidence levels: for each research claim, what it rests on (review, single study, expert opinion or the tool's reasoning), with effect sizes rather than bare appeals to science.
• ☐ Forbid invented sources: tell it never to make up a citation and to say when it cannot find one.
• ☐ Ask for the case against: what would someone who disagrees say, and where is the field truly divided?

Illustrative example: the business, the roles and the tool's answers are made up.

The sales lead at the Gothenburg bakery from 2.1 asks an AI assistant whether raising the minimum order makes small cafes stop ordering, for the April board report. The answer is fluent and confident.

Step 1: The first prompt asked none of the four demands. The sales lead asks again with all four.

Step 2: The key claim, that higher minimums drive small customers away, is causal but rests on one trade article and the tool's own reasoning. One cited study cannot be found at all.

Step 3: The analysis assumed a large wholesaler with a field sales team. The neutral rerun says the effect "depends on the alternatives customers have", which is far less certain than the first answer.

Step 4: The output moves the work forward only as a source of interview questions. The report says the background was drafted with AI help and that the conclusion rests on the bakery's own interviews, planned with 2.5 Simple Study Design Sheet. Decision: use as a draft only.