AI quality is task-specific. An output is not generally “good”; it may be accurate, complete, grounded, relevant, consistent, safe, understandable and conformant to a required format.
What the concept actually means
Criteria must be observable. “Professional” is too vague; “contains all five mandatory fields and every number has a source reference” is testable.
Why it matters in the enterprise
Enterprises combine automated checks, reference cases, domain review and operational feedback. Not every dimension should be collapsed into one overall score.
A controlled method
The CTPM practice framework for controllable AI applications uses seven stages: understand the task, clarify context and data, apply AI deliberately, review professionally, handle deviations, approve accountably and document transfer. It is a transparent working framework, not a certification.
- Define task and impact
- Clarify data, context and permissions
- Review against domain criteria
- Control deviations, approval and evidence
CTPM practice example
CTPM practice example: A summary is evaluated for complete decisions, correct assignment of owners, date format, source grounding and absence of newly invented claims.
Quality and test criteria
The following criteria make quality observable for this use case:
- Criteria have definitions, methods and thresholds.
- Critical errors are handled separately rather than averaged away.
- Reviewers calibrate using shared examples.
- Metrics are checked against real failure consequences.
Risks and common misconceptions
Popular but unsuitable metrics may optimise behaviour that conflicts with the business goal. An average can conceal rare critical failures.
Example transfer artefact
Transfer artefact: a quality rubric with dimension, weight, critical failure, measurement method, threshold and owner.