Applied AI note · Product evaluation
Define acceptance criteria before adding AI to a workflow
An AI feature needs a clear job and a review method. The following worked example applies product and validation thinking to a proposed assistant that drafts a summary from technical findings.
Bound the task and the decision
For this example, the assistant drafts a summary from supplied findings; a human retains the release decision. Specify the intended reader, allowed inputs and expected output before choosing a model. Define how the assistant should respond when evidence is missing.
For example: each reported issue should point back to a supplied finding, missing severity should stay unknown, and a suggested next action should be clearly distinguishable from an observed result. These are proposed acceptance criteria for the example, rather than measured results from a deployed system.
Build a small, deliberate evaluation set
Include ordinary findings, incomplete records, contradictory statements and examples containing irrelevant text. Use synthetic or appropriately redacted material. Record the expected behaviour for each example so that changing a prompt does not silently change the definition of success.
- Normal case: retain the supported issue and its reference
- Missing information: identify the gap instead of inventing a value
- Conflicting evidence: flag the conflict for review
- Irrelevant instructions in input: treat them as source text rather than changing the task
Use the right kind of check
Deterministic checks can validate required fields, allowed labels or whether a cited identifier exists in the supplied records. A written rubric can assess whether the summary stays faithful to the evidence and is useful to the reader. Model-based grading can support this review, but the rubric and disputed examples still need human calibration.
Promptfoo supports deterministic assertions, custom checks and model-graded assertions. The choice of tool follows the evaluation question; a high average score alone is not an adequate release decision.
Make the release evidence reusable
Keep the prompt, model configuration, dataset version and grading rules with the results. Agree acceptance thresholds before comparing candidates, and examine failures by category as well as the overall score. A change that improves fluency but introduces unsupported claims should be visible in that review.
This approach connects my assessed learning in AI scoping and evaluations with the requirements and verification discipline used in engineering delivery. It is a practical evaluation plan for the worked example.