AI prototyping is different from traditional prototyping because the technology itself is uncertain. You are not just testing whether users want the thing — you are testing whether the thing is even buildable to the quality bar users require.
The Wizard of Oz Prototype
Put a real user in front of a fake AI interface. A human on the other end (you, an engineer, anyone) provides the responses in real time. The user thinks it is the model.
This tests: does the user understand how to use this? Do they trust the responses? Do they try to use it in ways you did not anticipate?
What it does not test: latency, consistency at scale, or whether the model can actually do this. You will need other methods for those.
The Prompt-First Prototype
Before writing a line of product code, spend a day in the model’s playground. Write prompts. Try every input type you expect in production. Try the adversarial ones. Capture the outputs.
You will learn more about feasibility in one day of serious prompting than in three weeks of engineering exploration. The goal is not a demo — it is an honest picture of the capability ceiling.
The Narrow Slice
Pick the one user scenario where AI would be most valuable and build only that — end to end, in production. No edge case handling, no configuration, no generalisation.
Ship it to a small group. See what happens. The learnings from 50 real users using a narrow feature are worth more than any amount of research or internal testing.
What to Measure in Prototyping
Do not measure satisfaction. Users are polite. Measure behaviour: do they use it again? Do they complete the task they started? Do they share it? Behaviour is honest in ways that survey responses are not.