How to analyze a public policy with AI

Translate the policy into an analyzable proposition

Begin with the authoritative policy text and identify the stated problem, objective, affected population, instruments, delivery actors, timetable, and decision authority. Quote or cite the exact provision behind each element. If the policy is only proposed, label it proposed throughout.

For a synthetic municipal program intended to reduce missed health appointments, a control card might read:

| Element | Working description | | --- | --- | | Problem | Missed appointments delay care and waste capacity | | Instrument | Reminder messages plus an easier rescheduling route | | Population | Patients with scheduled outpatient appointments | | Delivery actors | Clinic operations and messaging provider | | Intended outcome | More appointments attended or rescheduled in time | | Decision | Pilot in two clinics or retain current process |

AI can extract candidate elements, but a reviewer checks every one against the policy document. The absence of detail should remain a gap rather than becoming a generated assumption.

Draw the mechanism before judging the result

Write the causal chain as testable links: intervention, immediate response, changed behavior, intermediate outcome, and final outcome. Add necessary conditions and plausible failure modes. In the synthetic program, a reminder must reach the correct person, arrive at a useful time, be understood, offer a workable action, and connect to available scheduling capacity.

HM Treasury’s Magenta Book describes evaluation as a systematic assessment of policy design, implementation, and outcomes and emphasizes asking what worked, for whom, why, and at what cost. Its guidance also distinguishes process, impact, and value-for-money questions. That distinction helps prevent a delivery count from being presented as evidence of impact.

Separate four kinds of statement

Tag each important sentence in the draft:

Consider the sentence “Text reminders will reduce missed appointments.” Without a relevant design and population, that is an unsupported causal claim. A bounded alternative is: “The pilot tests whether reminders and simplified rescheduling change attendance or timely cancellation in these clinics; the analysis will separately measure delivery, response, and appointment outcomes.”

AI can challenge the tagged draft by asking which mechanism link each source supports, whether the population matches, and what alternative explanations remain. It should not assign evidential weight without a documented method.

• evidence: a source directly reports a fact or estimate;
• assumption: the analysis requires a condition that has not been established;
• value judgment: the choice gives weight to an outcome or group;
• forecast: the analysis predicts what may happen under an option.

Compare options on the same decision frame

Define a baseline and at least one feasible alternative. Compare benefits, costs, risks, distributional effects, implementation constraints, and reversibility. The 2026 Green Book describes appraisal as assessing costs, benefits, and risks of options for achieving objectives so decision makers can understand impacts and trade-offs. It also places appraisal within a broader cycle of rationale, objectives, appraisal, monitoring, evaluation, and feedback.

For the synthetic brief, the options could be no change, reminder only, or reminder plus rescheduling. Record who may benefit, who may be excluded, and which operational dependency is most likely to break. Do not turn the matrix into a numerical score unless the weights and evidence justify one.

Release the uncertainty with the recommendation

The brief should contain the policy basis, decision question, source map, mechanism, option comparison, stakeholder effects, uncertainties, and an evaluation outline. The recommendation should state what is proposed now, why it is proportionate, what is not known, and what evidence would trigger expansion, revision, or stopping.

Before release, verify that:

• policy status and quoted provisions match the authoritative text;
• each causal statement has suitable evidence or is labeled as a hypothesis;
• stakeholder effects include groups that may bear costs or face exclusion;
• all options use the same objectives and evidence rules;
• implementation dependencies and alternative explanations are visible;
• the monitoring and evaluation plan can test the proposed mechanism.

Sources and scope

This resource is an educational analysis method. It is not legal advice, an appraisal of a real policy, or evidence that the synthetic intervention would work.

• HM Treasury and Evaluation Task Force, The Magenta Book: used for evaluation scope, policy-life-cycle integration, and the distinction among process, impact, and value-for-money questions.
• HM Treasury, The Green Book 2026: used for option appraisal, objectives, costs, benefits, risks, and trade-offs.

Sources: [1] [2]