Owner decision
Which decisions can use AI output, and which still require accountable human judgment?
AI can support research, options, drafting, comparison, and bounded execution when inputs and checks are controlled. A named human must still own decisions involving people, money, customer contact, legal or safety exposure, irreversible action, and any consequence the tool cannot carry.
Use this now: Create an AI decision boundary with allowed support, required evidence, human owner, approval point, stop rule, and incident path.
Owner worksheet
AI-to-human decision routing matrix
| Check | Write down |
|---|---|
| Decision and consequence | Name the business decision, affected people or customers, financial or legal consequence, reversibility, and urgency. |
| Permitted AI role | Specify whether AI may retrieve, summarize, classify, draft, calculate, recommend, or act, plus the data and system boundary for that role. |
| Verification requirement | List sources, calculations, assumptions, tests, and human expertise required before the output can influence or execute the decision. |
| Human decision owner | Name the accountable role that can approve, reject, stop, and repair the action and the authority that role actually holds. |
| Approval and stop gate | Define the risk or uncertainty that requires higher review, the prohibited action, emergency stop, rollback, and incident path. |
| Decision record | Log model or tool, version, inputs, output, verification, human judgment, final action, date, and observed result. |
Close the decision: Use AI only inside its written role; an accountable human must own consequential approval, stop, repair, and the decision record.
Worked example
Illustrative AI use, assist, human-decide, and prohibit cases
Illustrative only: No AI output is declared safe by category alone; actual consequence, privacy, security, employment, legal, financial, and human-authorization rules control.
| Step | Illustrative input | Replace with your evidence | Completed test or status |
|---|---|---|---|
| Low consequence: USE | AI drafts an internal meeting summary; human owner checks decisions, names, dates, and source transcript before filing. | Transcript-to-summary comparison and filer identity. | AI may draft; verified record remains human-owned. |
| Medium consequence: ASSIST | AI compares three vendor responses; procurement verifies price, term, security answer, and omissions against original proposals. | Proposal citations and comparison receipt. | AI structures evidence but cannot select the vendor. |
| High consequence: HUMAN DECIDE | AI outlines a financing scenario; finance owner recomputes numbers and qualified advisers review legal/tax/accounting implications before any decision. | Actual model, assumptions, and specialist review. | No AI recommendation releases financing action. |
| State-changing: PROHIBIT AUTONOMOUS ACTION | Contract termination, payment release, employee action, or destructive system change requires authenticated human authorization and rollback/notification plan. | Policy, authorization event, and system guardrail. | AI cannot self-authorize the action. |
| Close rule | Route by consequence, reversibility, affected party, evidence quality, authority, and recoverability; retain the decision receipt. | Risk classification and owner readback. | Decision: use the least authority consistent with the verified case. |
Decision produced: Permit drafting and evidence organization under review, retain accountable human judgment for material decisions, and prohibit autonomous high-consequence state changes.
