Score each question 0 (no), 1 (partly) or 2 (yes). Total the column.
| # | Question | Score |
|---|---|---|
| 1 | Is the decision written down, with its inputs and outputs? | |
| 2 | Do we know what a wrong answer costs, in money or trust? | |
| 3 | Can the action be undone within a window? | |
| 4 | Do we have 50+ real past cases with a known good answer? | |
| 5 | Is there a named human who owns this decision today? | |
| 6 | Can we enforce hard bounds (caps, floors, step limits) in code? | |
| 7 | Can the agent be given only the tools it needs, nothing more? | |
| 8 | Can we log inputs, proposal, guardrail results and action for every decision? | |
| 9 | Is there a clear rule for when to escalate to a human? | |
| 10 | Can we measure accuracy and cost per decision weekly? | |
| 11 | Would compliance, finance and customers accept the audit record? | |
| 12 | Is there a kill switch a non-engineer can pull? |
Reading the score
- 0 to 11: suggest-only. Let the agent draft; a human acts.
- 12 to 19: act within bounds. Tight step limits, daily review.
- 20 to 24: act and report. Sampling audits, weekly review.