How to develop and assess recommendation formation in AI-enabled work

Form the recommendation after preserving genuine choice

Start with the decision that must be made and the authority of the recipient. Then retain a credible range of options, including continuing current arrangements where relevant. The UK Green Book 2026 says appraisal assesses the costs, benefits, and risks of different options rather than setting objectives or making policy decisions. It also emphasizes genuine choices, explicit deliberation, and the evidence and assumptions behind judgments.

This distinction matters in AI-enabled work. An assistant can make one option sound inevitable by drafting a fluent rationale around it. The worker should instead separate three layers:

Do not start by asking an AI assistant for “the best recommendation.” First state the objective, decision boundary, available options, critical constraints, and evidence standard. Generate a counter-position before converging. If only one option remains, explain whether alternatives failed a genuine constraint or were never developed.

• evidence describes what the available records support;
• judgment interprets trade-offs and uncertainty for this context;
• recommendation proposes an action to an owner who can accept, reject, modify, or defer it.

Build a WARRANT recommendation note

Use WARRANT as a compact decision artifact:

| Field | Question | Evidence to retain | | --- | --- | --- | | What decision | What is the owner being asked to authorize now? | Bounded action, owner, and deadline | | Alternatives | Which credible options were considered, including a baseline? | Option list and exclusion reasons | | Reasons | Which objectives and evidence make the preferred action stronger? | Claim-to-source links and comparison | | Risks | What could make the recommendation harmful, ineffective, or premature? | Risk, uncertainty, and counterargument | | Accountability | Who owns approval, implementation, checks, and specialist input? | Named roles and authority limits | | Next conditions | What must happen first, and what would pause or reverse the action? | Preconditions, stop conditions, and signposts | | Test and revisit | How will the decision be checked and when will it be reopened? | Review date, evidence measure, and fallback |

The note should make proportionality visible. A reversible pilot can tolerate more uncertainty than an expensive, irreversible commitment, but it still needs clear boundaries and a useful test. A recommendation to wait also needs evidence: specify what uncertainty matters, what information is expected, and how the delay affects the objective.

Reasons should correspond to the decision criteria rather than repeat attractive features. Risks should include the strongest plausible case against the recommendation, not a list of generic cautions. If one uncertainty can change the preferred option, show that dependency in the next conditions or revisit field.

Worked artifact: choosing a document-review pilot

Consider a synthetic programme team deciding how to reduce unsupported claims in internal briefings. This example is original and does not describe a Two Prune customer, measured result, or product promise.

The team considers four options:

The available evidence includes a sample of 20 recent briefings, interviews with five reviewers, a time estimate, and current information about the external service. The sample shows recurring missing source notes, but it is too small and selective to establish an organization-wide rate. Reviewer interviews suggest that a full claim ledger would add substantial effort to low-risk updates. The external service has not yet answered a material data-handling question.

The WARRANT note asks the programme lead to authorize a six-week structured-form pilot for one team, not a permanent rollout. Its rationale is specific:

The note acknowledges the strongest counterargument: a form may create administrative compliance without better judgment. The pilot therefore samples the quality of the completed evidence fields, not only whether boxes were filled. It records three preconditions: an approved synthetic test set, reviewer briefing, and a definition of a consequential claim. It sets a stop condition if the form causes restricted information to be copied into the wrong field.

The revisit plan does not promise success. At six weeks, the programme lead reviews completion, sampled claim traceability, reviewer effort, observed workarounds, and any exceptions. The possible dispositions are continue testing, revise the form, return to peer review, or develop a different option. Procurement of the external service remains separate until its evidence gap is resolved.

This is a recommendation because it proposes an owned action with reasons, boundaries, and a decision path. It is not merely “Option 3 scored highest,” and it does not present the synthetic pilot as proven to improve real outcomes.

• continue informal peer review;
• add a mandatory claim ledger to every briefing;
• pilot a structured review form for one team;
• purchase an external review service for all teams.
• the pilot directly tests whether required source and uncertainty fields improve the inspectability of briefings in the sampled workflow;
• it is more targeted than a ledger on every claim and does not depend on unresolved external-service evidence;
• the current process remains the baseline, so the team can compare review completion, returned claims, and reviewer effort during the same period.

Practice by forcing a change in the recommendation

Give a learner a decision brief, mixed evidence, and at least three viable options. Ask for a WARRANT note. Then reveal one new fact chosen to test the reasoning: an implementation dependency is unavailable, a cost estimate doubles, a key source is less applicable, or the decision becomes harder to reverse.

The learner should identify which field changes and whether the recommendation should narrow, switch, pause, or remain. Reward a stable recommendation only when the new evidence does not affect the decisive reasons. Changing an answer is not weakness when the warrant has changed.

AI can help enumerate alternatives, construct a counterargument, check whether reasons trace to evidence, and draft reader-friendly language. The worker must decide what evidence is credible, which trade-offs matter, and what action to propose. Retain the material prompts or outputs that changed the option set or exposed a weakness, not every conversational turn.

Assess the recommendation from observable work

Use the decision process and final artifact:

Retain the option set, evidence map, excluded-option reasons, counterargument, draft revisions, and final WARRANT note. A persuasive executive summary alone is weak evidence because it may hide missing options, unsupported reasons, or an action outside the recipient's authority.

• decision fit: the recommendation asks the correct owner for an action within their authority;
• option integrity: credible alternatives and the baseline survive long enough for comparison;
• evidence traceability: decisive reasons connect to relevant, inspected sources;
• claim discipline: observation, inference, judgment, and recommendation remain distinct;
• counterargument quality: the strongest plausible objection is addressed rather than caricatured;
• proportionality: commitment, reversibility, evidence strength, and consequence are aligned;
• implementation realism: owners, preconditions, dependencies, and stop conditions are concrete;
• adaptation: new evidence changes the note when it changes the warrant.

Verify before sending the recommendation

Run a reverse trace. For every sentence that supports the preferred action, identify the evidence or label it as judgment. Then run a rival test: write the strongest recommendation for the best alternative using the same evidence. Differences should come from declared objectives, criteria, risk tolerance, or uncertainty, not selective sourcing.

Check implementation independently. Confirm that the named owner has authority, prerequisites are achievable, the test can be observed, and the fallback is real. Finally, remove confident adjectives and see whether the logic still holds. If the recommendation depends on words such as “clearly” or “obviously,” the warrant likely needs strengthening.

Recommendation formation supports accountable action under uncertainty. It does not transform limited evidence into certainty or replace the decision owner's judgment. In a Two Prune assessment, the artifact can provide task-bounded evidence about Judgment & Synthesis, Evidence Navigation, and Communication & Delivery. It describes observed performance in the simulation and does not make an employment decision.

Sources

UK Government: The Green Book 2026

• UK Government: Government functional standard GovS 010, Analysis
• U.S. Government Accountability Office: Evidence-Based Policymaking

Sources: [1] [2] [3]