How to develop and assess decision framing in AI-enabled work

Frame the choice before asking AI for a recommendation

HM Treasury’s 2026 Green Book describes appraisal as assessing the costs, benefits, and risks of different options for achieving stated objectives. It distinguishes objective setting from appraisal and says advice should give decision-makers genuine choices supported by evidence. Its supplementary multi-criteria decision-analysis guidance also emphasizes a transparent record of the evidence and assumptions used to consider choices against policy objectives.

These are useful principles beyond public investment. A knowledge worker should be able to state what decision is open, who owns it, what is fixed, and what evidence could change the choice. “Research the market” or “write a strategy” names activity, not a decision.

Build a CHOICE decision frame

Use one CHOICE record for the decision the work product must support:

| Field | Framing question | Evidence to retain | | --- | --- | --- | | Choice | What specific commitment, approval, prioritization, or next step is open? | Decision statement and excluded decisions | | Holder | Who has authority, who advises, and who is affected? | Named roles and decision rights | | Objective | Which outcome matters, and how will options be compared? | Objective, critical success factors, and trade-offs | | Information | What evidence is required, available, missing, or uncertain? | Source list, assumptions, and evidence gaps | | Courses | Which viable options, including the current state or minimum action, deserve comparison? | Option set and elimination rationale | | Event | By when is the choice needed, what triggers it, and how reversible is it? | Deadline, review point, dependencies, and exit route |

Write each field as a checkable statement. Mark uncertain information explicitly. If two stakeholders appear to ask different questions, do not merge them into a vague objective. Identify whether there are two decisions, a decision plus an implementation question, or a disagreement that the decision owner must resolve.

AI can help challenge a frame, generate missing questions, or test whether an option set is artificially narrow. The worker remains responsible for authority, source interpretation, and the final framing judgment.

Worked example: a synthetic supplier renewal decision

Imagine a synthetic operations team preparing a recommendation about a document-management supplier. The contract expires in four months. The brief says “review performance and propose next steps,” while one email asks for a replacement shortlist. This example is invented and does not describe a Two Prune customer.

The first AI response produces a broad market comparison. The participant stops and completes a CHOICE record:

The frame reveals that a product feature comparison is not yet the main need. The immediate artifact should compare the three courses against the agreed objectives, show missing security and exit-cost evidence, and identify the latest date for a reversible decision. The email requesting a shortlist becomes one stakeholder input rather than the decision itself.

The revised recommendation can still use AI for structuring evidence or challenging assumptions. Every option and claim remains tied to the supplied documents. The result gives the director a defined choice rather than a polished market essay.

• Choice: renew for one year, renegotiate with conditions, or begin a replacement procurement;
• Holder: the operations director approves the path after security, finance, and user-support advice;
• Objective: maintain reliable document access while meeting security requirements and controlling transition cost;
• Information: current service incidents, active-user patterns, renewal terms, security review, exit obligations, and migration capacity;
• Courses: current arrangement as the baseline, conditional renewal, and replacement procurement, with immediate termination excluded because the evidence pack shows no supported trigger;
• Event: a direction is required before the notice deadline, and replacement is harder to reverse once migration begins.

Practice with a frame-shift drill

Give a learner an evidence pack and an ambiguous request such as “assess our AI rollout.” Ask for a CHOICE record before any recommendation. Then introduce a change that alters the decision, such as a new deadline, a constrained budget, a changed owner, or evidence that one option is unavailable.

Require the learner to show which parts of the frame, research plan, and output changed. The useful behavior is not merely editing the conclusion. It is recognizing that new evidence can change the option set, decision timing, evidence threshold, or authority route.

A focused practice cycle includes:

• rewriting an activity request as a decision statement;
• separating the accountable owner from advisers and affected groups;
• defining objectives before scoring or comparing options;
• keeping a baseline or minimum-action option visible;
• recording uncertainty and evidence that could change the choice;
• identifying deadlines, triggers, reversibility, and follow-on decisions;
• testing whether the planned artifact answers the framed choice.

Assess observable framing behavior

In an assessment, provide a brief with incomplete authority, several plausible objectives, and evidence that supports different courses. Ask for a decision artifact plus the CHOICE record or equivalent planning evidence.

Review observable evidence such as:

Do not award credit just because the document contains a recommendation. A confident conclusion can follow a weak or invented frame. Likewise, a participant can frame the right decision while reaching a debatable conclusion. These are related but separate judgments.

• the participant states one decision precisely and names what remains outside scope;
• the accountable owner and advisory roles are identified from the evidence;
• objectives and critical conditions appear before option comparison;
• the option set includes a credible baseline rather than only attractive new actions;
• missing evidence and uncertainty influence the recommendation or next step;
• timing and reversibility change the analysis appropriately;
• AI-generated options or criteria are checked against the source environment;
• the final deliverable answers the decision that was framed.

Verify with a decision-readiness check

Give the artifact and CHOICE record to a reviewer who did not perform the analysis. Ask the reviewer to state the open decision, owner, options, objectives, evidence gaps, and decision date. If the reviewer cannot recover those elements, the frame is not carrying through to the deliverable.

Then trace each material recommendation back to an objective and evidence source. Test whether a new fact would change the choice or only its implementation. This exposes hidden assumptions and prevents the frame from becoming decorative metadata.

Decision framing is task-dependent. It does not by itself establish that a recommendation is correct or that someone will perform well in another setting. In a Two Prune assessment, the frame, source review, AI interactions, revisions, and final artifact can contribute bounded evidence about Context Framing, Evidence Navigation, Judgment & Synthesis, and Communication & Delivery. They do not determine suitability for employment.

Sources

HM Treasury Green Book 2026

• HM Treasury supplementary guidance on multi-criteria decision analysis

Sources: [1] [2]