Model delegation: decide what AI may produce, check, and never own

Start with the consequence, not the prompt

Describe the decision, output, audience, source material, deadline, and harm if the output is wrong. Then separate the work into operations such as extraction, classification, calculation assistance, drafting, critique, recommendation, communication, and action.

Do not delegate a task merely because the model can generate an answer. Ask whether the output is reversible, whether the evidence is available to the reviewer, whether an error will be noticed, and whether the model would cross a privacy, professional, financial, safety, or authority boundary.

The NIST AI Risk Management Framework Core calls for organizations to define human and AI roles, document system limits and oversight, specify intended scope, and evaluate performance under conditions similar to deployment. The framework is voluntary and is being revised. Use it as a source of risk-management questions, not as certification of an individual workflow.

Sources: [1]

Build a delegation ladder

Classify each operation on one of five levels:

For each operation, record permitted inputs, expected output, principal failure, verification method, reviewer, and stop condition. As the ladder rises, require stronger evidence, more independent checking, and clearer action separation. Many workflows should stop at Transform or Analyze. A model recommendation need not receive permission to act.

• Suggest: generate questions, outlines, or alternatives with no claim of correctness.
• Transform: summarize, format, translate, or extract from named sources while preserving traceability.
• Analyze: compare evidence, calculate, or identify patterns under explicit assumptions and checks.
• Recommend: propose an option with visible criteria, uncertainty, and human review.
• Act: send, publish, approve, purchase, update a system, or otherwise change external state.

Use the delegation envelope

A delegation envelope is the compact artifact that travels with the task:

| Envelope field | Question | | --- | --- | | Purpose | What decision will this output inform? | | Permitted operation | Suggest, transform, analyze, recommend, or act? | | Inputs | Which sources and data classes may be used? | | Exclusions | Which content, claims, and actions are prohibited? | | Acceptance test | What observable conditions must the output satisfy? | | Verification | Which source, calculation, comparison, or person checks it? | | Authority | Who can approve and take the next action? | | Stop condition | What uncertainty or failure ends the delegation? |

The envelope prevents a narrow request from expanding silently. If an assistant asked to summarize a contract begins proposing obligations, the operation has moved from Transform toward Recommend and requires a new boundary.

Worked example: a synthetic executive brief

A strategy analyst receives a board pack containing an approved forecast, draft meeting notes, a market report, and a disputed customer figure. The analyst asks an AI assistant to prepare an executive brief.

The first result treats the draft notes as approved, uses the disputed figure, and recommends entering a new market. The analyst rewrites the delegation envelope:

| Field | Recorded instruction | | --- | --- | | Purpose | Prepare an internal evidence summary for executive review | | Permitted operation | Transform approved facts and analyze stated scenarios | | Inputs | Approved forecast and named market-report sections | | Exclusions | Draft notes, disputed customer figure, final market recommendation | | Acceptance test | Every number traces to a source; assumptions appear beside scenarios | | Verification | Recalculate forecast totals and open each cited section | | Authority | Analyst reviews; executive team decides | | Stop condition | Conflicting figures, missing source, or unsupported causal claim |

The revised output presents two scenarios and an unresolved evidence gap. It does not recommend market entry. The example is synthetic and does not represent a customer outcome or professional advice.

Test the delegation boundary

Use safe historical, synthetic, or approved material to test boundary pressure. Include an ambiguous instruction, unavailable source, conflicting figure, sensitive field, request to send externally, and tempting but unsupported conclusion. Observe whether the user notices when the task changes level and whether the workflow stops when verification becomes impossible.

NIST's AI RMF appendix on human-AI interaction explains that human roles and responsibilities should be clearly differentiated and that human-AI outcomes can vary with context. It also notes the importance of examining how people challenge or overrule output. This supports testing the configuration and behavior, not assuming that nominal human review is effective.

Sources: [2]

Verify the operation that was delegated

Different levels require different checks. For Suggest, test relevance and diversity. For Transform, compare against the source and check omissions. For Analyze, reproduce calculations and test assumptions. For Recommend, examine criteria, alternatives, uncertainty, and counterevidence. For Act, require explicit authorization and controls appropriate to the external change.

Record corrections, abstentions, escalations, boundary crossings, and cases in which the reviewer could not reproduce the result. A person clicking approve is not evidence of meaningful oversight. The reviewer must have time, competence, and access to the underlying evidence.

Assess delegation skill without overstating it

A controlled assessment can provide mixed-quality evidence, an AI assistant, an external-action request, and a deadline. Observable evidence may include whether the participant decomposes the task, chooses an appropriate ladder level, writes a usable envelope, checks the output, catches boundary drift, and preserves human ownership.

Report only the behavior and evidence observed in the scenario. One exercise does not establish universal safety, professional competence, future job performance, or authorization for higher-consequence work. Two Prune provides assessment evidence and does not make employment decisions.