Key takeaways
- The base score is likelihood multiplied by consequence; detectability, scope, reversibility and sensitive data add visible modifiers.
- Safety, rights or a critical irreversible effect creates a hard stop regardless of the numerical score.
- The result is an educational triage prompt, not permission, legal advice, a security review or a compliance classification.
- Inputs remain in your browser and the report can be copied without an account.
Triage one proposed action
Risk triage · 4 base + 3 modifiers
Moderate · score 7
The base is likelihood × consequence. Detectability, affected scope, reversibility and sensitive data each add visible modifiers. A rights/safety flag, or a critical irreversible effect, overrides the total.
Controls to investigate
- Define an observable success check and alert before widening scope.
- Cap batch size, permissions and affected audience for the pilot.
- Keep a before-state and test rollback; use blocking approval where rollback is incomplete.
The visible calculation
Likelihood and consequence each use four levels, producing a base from 1 to 16. Detectability, scope and reversibility each add zero, one or two. Sensitive data adds one. Scores below 6 are low, 6–10 moderate, 11–15 high and 16 or more critical. A hard stop makes the result critical regardless of the total.
NIST's AI Risk Management Framework treats risk management as contextual and continuous across govern, map, measure and manage activities. This small matrix borrows that inspectable risk logic but does not implement the framework. It cannot know affected people, legal obligations, system dependencies or actual failure rates; those require evidence and qualified owners.
| Input | Question | Evidence |
|---|---|---|
| Likelihood | How often would this failure occur under the proposed controls? | Representative tests and comparable production events. |
| Consequence | What is the worst credible effect before recovery? | Affected records, people, money, access and obligations. |
| Detectability | Would the team know before or only after harm? | Validation, postconditions, alerts and sampled misses. |
| Scope | How far can one wrong run propagate? | Permissions, batch caps, audience and downstream consumers. |
| Reversibility | Can the exact prior state and downstream effect be restored? | Tested rollback and evidence of residual exposure. |
Why some cases override arithmetic
A low estimated likelihood does not make a safety or rights decision suitable for automatic execution. Nor does a small batch make an irreversible critical effect ordinary. The hard-stop flags prevent a reassuring average from hiding one dimension that requires policy, legal, safety, security or subject-matter authority.
Hard stop does not always mean the wider process must remain manual. It may mean prohibiting that action, narrowing permissions, replacing judgment with deterministic validation or placing a qualified human decision immediately before the effect. The Approval Gate Designer turns a chosen review point into a runtime contract.
Use the report in a risk review
- 01Name one effectUse an action verb and target: send a message, change a status, grant access, publish a result or issue a payment.
- 02Challenge every estimateAsk the process owner, control owner and affected operator for evidence, especially about detection and rollback.
- 03Apply controls and rescoreReduce permissions, batch size and ambiguity; add preconditions, postconditions, monitoring and safe failure states.
- 04Record the decisionKeep assumptions, evidence, owner, permitted boundary and review trigger with the workflow version.
Risk assessment versus readiness and approval design
The readiness assessment asks whether the process is stable and specified enough to pilot. This tool asks what happens if one proposed automated effect is wrong. The gate designer then asks how a selected human review point should pause, decide, expire and resume. Use them in that order; no single score answers all three questions.
Limitations and when not to use this
- This educational matrix is not legal, safety, security, privacy, compliance, employment or financial advice and is not a substitute for a formal risk assessment.
- The thresholds are intentionally simple and are not calibrated to an industry loss scale. Governing policy can require stricter treatment at any score.
- Inputs are not submitted to all-agents, but planning descriptions should still exclude real personal data, credentials and confidential records.
Sources
- AI Risk Management Framework Core — NIST Accessed 1 August 2026
Design the control
Use consequence, reversibility and uncertainty to draft an approval-gate contract.
Design the controlUli Prantz
Builds and operates all-agents
Uli Prantz builds all-agents, the process-automation platform this site documents. He writes about the operational side of automating recurring business work: where deterministic code beats model judgment, where it does not, and where a human still has to approve.