Unit 01.03: Reading a vendor demo critically
A demo is a recording of a good day. Six questions turn it into something you can evaluate.
What the happy path does not show
Failure handling, loop protection, approvals, cost, input selection, and out-of-scope behaviour.
The table below lists all six with the usual honest answer.
ask the demo answer
what happens when the tool call fails? usually: the demo has no failure
what stops it if it loops? usually: nothing shown
who approved the action we just saw? usually: nobody
what did that cost in tokens? usually: not displayed
was that input chosen or typical? usually: chosen
what does it do with a question outside scope? usually: not demonstrated
6 questions. A demo that answers all six is a system;
one that answers none is a recording of a good day.
None of these is a hostile question. They are the properties that decide whether the system survives real inputs, and they are absent from a demo precisely because a demo is optimised to show capability rather than robustness.
"Was that input chosen or typical?" is the one to ask first. Everything else is downstream of it - a demo run on a curated input tells you the system can handle that input, which was never in doubt.
The mistake this prevents
The mistake is evaluating a demo on what it accomplished. Accomplishment is the easy part to arrange. Ask instead what it does when a tool call fails, and watch whether the answer is a designed behaviour or a shrug.
Takeaway
Evaluate demos on failure handling, stopping, approvals, cost and out-of-scope behaviour - none of which a demo shows. Start with whether the input was chosen.
