Consequence
Who can be helped or harmed if the result is wrong?
Trust is an operating choice
Design the accountability
Good intentions are not a control. Put ownership, evidence, review, override, monitoring, and recourse inside the workflow.
Use the review map
Six tests before a task changes lanes.
Higher consequence, lower reversibility, weaker evidence, and less recourse call for more active human control.
Reversibility
Can the decision or action be safely undone?
Evidence
Can a reviewer trace the claims to reliable sources?
Uncertainty
Does the workflow recognize novelty and missing context?
Contestability
Can an affected person question or correct the result?
Accountability
Is one qualified human authorized to accept the outcome?
A continuous loop
Govern. Map. Measure. Manage.
NIST's AI Risk Management Framework organizes responsible AI work around these four connected functions. The work continues after launch.
Open the NIST AI RMF ↗- 01
Govern
Name purpose, policy, owners, prohibited uses, and recourse.
- 02
Map
Understand people, data, context, dependencies, and likely effects.
- 03
Measure
Test quality, reliability, fairness, security, and human impact.
- 04
Manage
Control risk, respond to incidents, and revisit the design.
The minimum viable accountability record.
If the team cannot answer these questions in plain language, the workflow is not ready to scale.
- Purpose
- What outcome is this workflow allowed to pursue?
- Owner
- Who can accept, reject, pause, or change the result?
- Evidence
- What information may be used, and how is it verified?
- Review trigger
- What risk, novelty, or uncertainty requires human attention?
- Recourse
- How can a worker or customer question and correct the result?
- Monitoring
- Which quality, risk, value, and human outcomes are reviewed?
Scope
This resource is educational guidance, not legal, employment, cybersecurity, privacy, safety, or regulatory advice. Requirements vary by use, industry, location, and affected population. Involve qualified reviewers for consequential decisions.