Build an AI Pilot Test Set From Exceptions, Not Just Easy Examples
Before an AI pilot, collect representative exceptions and define acceptable handling. Include missing information and cases that should go to a person.

An AI demonstration looks convincing when every example is complete, familiar and neatly formatted. A small business's working inbox may contain none of those things.
Before proposing a pilot, ask the owner to help define a small, authorized set of examples that includes awkward cases. The purpose is to learn how the workflow behaves, not to manufacture a flattering score.
Choose one decision the workflow must make
For an illustrative inquiry-triage pilot, the decision might be whether a message belongs with sales, support or a person who can clarify it. Do not combine triage, quotation and automatic sending into one vague success criterion.
Use fictional or appropriately sanitized examples unless the business has approved a suitable way to handle real data. Do not copy customer records into a tool merely because they are convenient.
Include the cases that should not produce a confident answer
- A message missing the information needed to route it.
- A request covering two different departments.
- A message outside the business's service scope.
- A repeated message with a correction.
- A request whose intended meaning is unclear.
For each, write the expected handling before running the test. “Ask a person to review” can be a correct outcome. It should not be scored as failure simply because the system did not automate everything.
Record the mistake in operational terms
Instead of “bad output,” write “sent to sales despite an unresolved support issue” or “invented a missing project location.” Then decide whether the problem can be addressed through clearer scope, better inputs, human review or a different approach.
Keep examples used for revisions separate from a later check. Repeatedly improving the same demonstration does not show how the workflow will handle a new case.
Use the results to narrow the pilot
Report which cases worked, which need review and which remain outside scope. Agree on who can pause the pilot and how staff will report a problem.
AI Consultant Launchpad's Starter Kit is a place to begin planning consulting work. A useful pilot recommendation includes the exceptions, not just the appealing demo.
Featured image: AI-generated editorial illustration.
Get the Starter Kit
Get the Starter Kit