Build a defensible evidence check [VERIFY: not defined in source] by selecting and justifying verification steps for a consequential workplace scenario in SurveyJS activity

Interactive scenario

Build a Defensible Evidence Check

What you are doing

Practise choosing verification steps you could explain to a colleague or manager. Build a habit you can apply to any AI-generated claim that matters at work.

Situation

You manage operations at a mid-sized logistics firm. Your team uses a generative AI tool to draft quarterly risk assessments. This quarter, the AI output states:

"Seventy per cent of transport firms in our region experienced supply-chain disruption due to extreme weather in the past two years. We should renegotiate our insurance coverage immediately."

The claim sounds specific. It could justify a costly decision. Before you pass it upward, you need to check whether the evidence is sound enough to act on. You have fifteen minutes.

Your decision

Which verification steps do you take first? Select the approach you would actually use.

• A. Search the AI output for the source of the "seventy per cent" statistic, then check whether the cited study exists, when it was published, and whether "our region" matches the study's geographic scope.
• B. Flag the claim for a junior analyst to verify in the next monthly cycle, since the statistic seems plausible and aligns with recent news reports you recall.
• C. Forward the AI output to your insurance broker as context for renegotiation, noting that the statistic requires formal verification but treating its plausibility as sufficient interim justification.

What happens

If you chose A

You separate the claim from its supposed evidence. You locate that the AI has cited no specific study, author, or dataset. The "seventy per cent" appears to be a synthesis or hallucination. You find one relevant industry survey from 2021 covering a different state, which reported disruption for forty-five per cent of respondents. Your region is unnamed in that survey. You document this gap precisely: claim unsupported, partial match found, geographic scope mismatch. Your manager can now see what you checked, what you found, and why you are not yet recommending action. This is a defensible evidence check.

If you chose B

You delay verification and rely on memory of news reports as informal corroboration. The claim gains circulation through your organisation without anchor to source material. Two weeks later, the statistic appears in a board briefing, attributed to "operations analysis." When challenged, you cannot reproduce the supporting evidence. Deferred verification becomes undetected misinformation. The monthly cycle does not protect you; your passivity enabled the gap.

If you chose C

You act on plausibility while nominally flagging a future check. The broker begins renegotiation using your forwarded output as client-provided context. The "seventy per cent" enters commercial discussion as if substantiated. Your caveat is forgotten in email chains. You have converted an unchecked claim into organisational action, and your hedging language does not shield you from accountability when the figure is later disputed.

Takeaway

A defensible evidence check is inspectable. Someone else can trace what you looked for, what you found, and where the gap remains. Build this habit for every consequential AI claim.