The decisions that define, validate, interpret, and act on an outcome measurement should remain manual. Manual does not mean that every calculation must be performed by a person; it means that people retain authority over the judgment that gives a result meaning and consequence. A product team should therefore keep manual ownership of the success criterion, the evidence judgment, the acceptance decision, and any consequential follow-up.
Two relevant statements in the NIST AI RMF support this boundary: processes for human oversight should be defined, assessed, and documented, and system performance or assurance criteria should be measured qualitatively or quantitatively and demonstrated under conditions similar to the deployment setting.
Which decisions should remain manual?
- The definition of success. The team should decide what outcome is being measured, in what context, and what would count as an acceptable result. A convenient proxy should not silently replace the intended outcome.
- The measurement plan. The team should decide which evidence is relevant and whether the test conditions are sufficiently similar to intended deployment.
- The judgment about evidence. Missing, conflicting, or ambiguous results still require an interpretation. A calculated metric can show what was measured without resolving what the result means.
- The acceptance decision. The team should decide whether the result meets the approved criterion and document the rationale, limitations, and uncertainty.
- Consequential follow-up. Decisions to revise the use case, gather more evidence, pause, escalate, or proceed should not be triggered solely by an unexplained score.
How to check that the boundary is real
- Separate calculation from judgment. A tool may help calculate or summarize a measurement, but the acceptance decision and follow-up action remain assigned to people.
- Make the decision record reviewable. The record should show the evidence, the measurement conditions, the decision owner, the rationale, and any known limitations.
- Test deployment fit. The team should compare the measurement setting with the intended deployment setting and document meaningful differences.
- Provide an override path. An authorized person should be able to reject a result, request more evidence, or escalate the case.
- Review the oversight process. The process for human oversight should be defined, assessed, and documented rather than left to an informal expectation.
What the team must still confirm
These NIST statements do not, by themselves, establish a universal list of manual decisions, fixed approval thresholds, required roles, or a requirement that deployment conditions be identical. The team must confirm that the metric represents the intended outcome, that the evidence is sufficient, and that the documented authority matches the team’s policies and applicable constraints.
Any legal, regulatory, contractual, or domain-specific obligation must also be checked against the relevant official source. The cited statements do not verify those obligations.
The practical boundary is therefore straightforward: automation can help produce a measurement, while people retain the decisions that determine what the measurement means and what happens next.