Prepare for AI consulting
Acceptance criteria for a small AI project
An acceptance criterion describes an observable outcome for an agreed task. “The AI must work” is insufficient. Specify what it must answer correctly, what it must refrain from doing and what evidence determines acceptance.
Understand the task
One impressive demonstration can hide failures with missing information or unavailable systems. Criteria should cover normal cases, boundaries and failures. Revise requirements when scope changes and distinguish a passed example from a guarantee for every future situation.
Prepare these inputs
- Approved brief and explicitly permitted actions.
- A bounded set of fictional inputs with expected outcomes.
- An approver and a process for deviations.
A practical workflow
- Write each requirement as input, expected answer or action, and evidence.
- Add unknown sources, denied access and changed information. Require visible manual clarification where appropriate.
- Save results and review failures with the approver. Correct the workflow and retest affected cases.
Worked example, not a customer case
An illustrative example
Illustrative criterion: “When a customer asks for an unknown price, the answer must omit an amount, identify missing scope and show the approved contact route.” A screenshot of the actual test can document that case. It does not prove coverage of every possible price enquiry.
Test failures and exceptions
| Test input | Expected outcome |
|---|---|
| A requirement names only a feature, not an outcome. | Add concrete input and evidence. |
| A change fixes one answer but breaks another. | Retest the relevant reference set. |
Evaluate whether it helps
Report passed, failed and untested cases separately. Measure approver corrections too. A percentage without count and scope makes actual capability difficult to assess.
Next step: Write at least one normal case and one stopping case as concrete acceptance criteria before implementation.
Sources and scope
Examples are editorial exercises using fictional information. They do not describe completed customer tests or activated integrations. The authority source below provides AI and data background rather than approval of the exercise.
Want to clarify your own workflow?
Bring one task and a non-sensitive example. Scope, price and timing are agreed individually before work. Reading a guide activates no customer messages or system actions.
Describe your task