NIST AI Risk Management Framework
Primary guidance on human oversight, accountability, and managing consequential AI risks.
← Back to Compute College
Compute College
Human approval and reversible actions
Design approval gates and undo paths before an AI system can take consequential actions.
Plain-English definition
Human approval is a deliberate checkpoint before an AI-generated action crosses a consequential boundary. Reversible actions can be undone or safely retried, while irreversible actions require stronger authorization, preview, audit, and escalation controls.
Memory trick: Preview before impact; reverse what you can.
Why it matters
A model can produce a plausible request without understanding authority, side effects, or the user’s intent. Approval and reversibility limit blast radius when model, tool, or context errors occur.
- The design makes a hidden tradeoff measurable before it reaches users.
- A clear boundary reduces retries, unsafe actions, or unnecessary capacity.
- The operating choice should be tested against the workload rather than assumed from a demo.
Simple example
An assistant may draft an email and show recipients, attachments, and text for approval. It may create a reversible calendar hold, but sending a contract or deleting a record requires a separate authenticated confirmation.
- The example names the input, decision, and expected boundary.
- A deterministic or human checkpoint owns the part a model should not guess.
- The result can be measured with quality, latency, safety, and cost metrics.
Example figures are illustrative calculations, not current quoted market prices.
Current example
Primary reference
This source explains the method or risk boundary. It does not replace testing the actual model, data, provider, and workload conditions in production.
Source discipline: examples on this page are illustrative; measure the real workload before making a release or capacity decision.
Common mistake
A vague “run” button gives a human nothing useful to approve. The review must show the exact action, inputs, target, side effects, and identity under which it will execute.
Practical takeaway
What you can do with this
Classify actions by impact and reversibility. Add preview, dry run, approval, idempotency, audit, and rollback behavior before exposing each tool.
- Start with a narrow workflow and explicit failure boundary.
- Instrument the path before optimizing it.
- Review the design whenever model, tool, traffic, or policy changes.
Decision check: can a reviewer understand and undo the exact action before it happens?
Compute College learning path
AI Engineering
Step 34 of 48: Human approval and reversible actions