Make a reviewed decision the unit of AI value

Make the reviewed management decision your unit of AI value, so that evidence, alternatives and accountability determine usefulness before prompt counts enter the conversation.

A busy conversation can produce useful exploration. It can also produce six polished explanations without resolving what the team should do. The relevant question for a leader is whether the work supports a commitment that someone can examine, approve and revisit. Counting prompts cannot answer that question.

Begin with the commitment on the table

State the decision as a choice with a boundary. “Understand our sales performance” is an inquiry. “Choose which sales constraint to investigate before the next operating review” identifies a management decision. The second formulation gives the work a stopping point and limits how much evidence is needed now.

Specify who can make the commitment and when. A functional lead may authorize an internal analysis but need another person’s approval for a customer promise or a change in spending. AI assistance does not alter that authority. Put the boundary in the brief before asking for recommendations.

Define an acceptable deliverable in ordinary language: a short explanation of the signal, the strongest competing explanations, comparable options, a recommended next step and the evidence that could change it. Use that brief to make the evidence, alternatives and responsibility for the decision clear.

Treat reliability as specific to the task

Evidence of strong performance on one activity does not establish reliability on another. The HBS research team’s account of the jagged technological frontier, published September 21, 2023, describes uneven AI capability across knowledge tasks. That finding supports evaluating the actual work you intend to delegate.

Break a management assignment into activities with different review needs. Organizing supplied notes, calculating a rate, judging a causal explanation and authorizing action are distinct jobs. A clean summary might justify a light factual check. A proposed cause of a revenue decline needs scrutiny of definitions, timing and alternative explanations.

Anthropic’s guidance on building effective agents, published December 19, 2024, recommends starting simply and adding complexity when performance warrants it. Applied here, that is a reason to establish a reviewable workflow before asking for greater autonomy. It is not evidence that a manager can outsource judgment.

Keep the evidence visible enough to challenge

A decision brief should let a reviewer distinguish an observation from an interpretation. Give each consequential observation a source, period and limit. “Conversion fell” needs a denominator and comparable cohorts. “Customers want a new offer” needs context about who was asked and whether they could authorize a purchase.

Record missing evidence beside the claim it limits. A source list at the end does little if the recommendation silently assumes that all sources describe the same population. When two inputs conflict, keep the conflict visible and explain whether resolving it would change today’s choice.

Then compare alternatives using the same questions. What would each option require? What could it teach the team? What is the cost of being wrong? Include a bounded investigation or a temporary hold when those are credible choices. Otherwise the brief may present a favored action with two decorative alternatives.

Make the human review consequential

A reviewer should be able to reject the reasoning, reduce the commitment or request a specific check. “Looks good” is too vague when the important uncertainty is whether the team is measuring the right cohort. Ask the reviewer to confirm the evidence boundary, the tradeoff being accepted and the next action’s authority.

Keep approval tied to the reviewed version. If a new input changes the recommendation, return that change to the decision owner. Editing the summary after agreement should not quietly expand what was authorized. This can be managed with a dated document and a short change note; the practice does not require an elaborate system.

Name the person responsible for the next step separately from the person approving it. A decision becomes actionable when the owner knows the scope, the checkpoint and the conditions for stopping or escalating. Accountability cannot be inferred from a participant list.

Measure what happened after review

Track whether the deliverable was accepted, how much preparation and review it required, and what was changed before use. Later, record whether the approved action happened and what evidence it produced. Acceptance, execution and business impact are separate observations.

Use the five-step decision method to move from signal to structure, options, a bounded pilot and measurement. For a concrete starting point, inspect the operating review example and identify the decision, source limits, owner and review date. Those are useful places to begin an AI evaluation because each can be checked by a person accountable for the work.