Every AI product demo is a performance, rehearsed on data chosen to behave. That is not an accusation, it is stagecraft, and the buyer’s job is to get backstage. This is the field guide: the two requests that end most theatre, then the questions and the sounds of good and evasive answers.

The two requests

Run it on our data, now. Bring a realistic, non-sensitive sample and ask to see it processed live. Watch for the hedge: setup required, needs configuration, our engineers would tune that. Some setup is legitimate; total inability to touch unfamiliar data means you are watching a video with a salary.

Show us the log. What did the system do for a real customer last week: the actions, the errors, the human interventions. Real products have logs and proud vendors open them, suitably anonymised. No logs is a finding that outranks everything else on the screen, and for agentic products it is disqualifying, per the governance guide.

The questions, and the sounds of answers

  • What happens when it is wrong? Good: a specific error path, review step, and a story. Evasive: accuracy percentages without denominators, or “it learns”.
  • What exactly is the AI here? Plenty of solid products are ordinary software with a model at one step, which is fine; the AI-washing guide covers the ones where the label is the product.
  • Which model underneath, and what happens when it changes? You are listening for a plan, not a brand.
  • What does the bill look like at our volume? Have your numbers ready and make them model it live; metered products fail by succeeding.
  • Where does our data go? The contract-reading guide holds the seven clauses this answer must survive.
  • What do we take with us if we leave? Data, configurations, logs; watch the pause before the answer.

After the demo

  1. Write down the promises while they are fresh, because the contract will not contain them.
  2. Structure any trial with numbers and an end date, borrowing the scorecard from the agent job description template.
  3. Score the meeting itself with the due-diligence sheet: vague answers are findings, not gaps.

Where I fit in

I sit on the buyer’s side of these meetings for clients: asking the questions above without needing the seller’s goodwill, translating the answers, and structuring trials that produce evidence. It is a small, cheap engagement that routinely pays for itself in one avoided subscription; the Automation Audit includes it wherever tool decisions are live.