Practical guide
How to create a policy brief with AI
An AI-assisted policy brief should help a named decision-maker understand a bounded question, compare feasible options, and see where the evidence is strong, weak, or missing. AI can help organise source material, test structure, and challenge unclear prose, but the author remains responsible for source selection, analysis, recommendations, and release. Use BRIEF: bound the decision, retrieve.
An AI-assisted policy brief should help a named decision-maker understand a bounded question, compare feasible options, and see where the evidence is strong, weak, or missing. AI can help organise source material, test structure, and challenge unclear prose, but the author remains responsible for source selection, analysis, recommendations, and release. Use BRIEF: bound the decision, retrieve and register evidence, inspect options, explain implications, and finish with provenance and review. The output is a decision-ready brief plus an evidence register, not a persuasive summary detached from its sources.
B: Bound the decision and audience
Write a one-sentence decision statement before drafting. Name the decision-maker, the decision they can actually take, the population or service affected, the geography, the time horizon, and constraints. A brief about whether a city should change summer cooling-centre hours is different from a general explainer about heat risk.
Record the brief's stance. Some briefs neutrally describe evidence and options; others make a recommendation under an explicit mandate. Do not let AI infer the mandate from source language. The World Health Organization's policy-brief template asks authors to state the policy question, provide key messages, describe methods, present evidence, and address implementation and monitoring. Its structure is useful beyond health, but each organisation should adapt it to its own decision process.
Use a commissioning card:
If no real decision is available, use another format. A background note, technical report, consultation summary, or implementation plan may fit better than forcing every analysis into a policy brief.
- decision and authorised decision-maker;
- intended readers and what they already know;
- affected groups and relevant context;
- options that are genuinely open;
- evidence cutoff date and source rules;
- length, accessibility, review, and release requirements;
- questions the brief will not answer.
R: Retrieve evidence and build the source register
Collect evidence before asking AI to synthesize it. For each source, record title, publisher, date, version, jurisdiction or setting, method, population, main finding, uncertainty, and relevance to the decision. Separate law or official policy, research evidence, local operational data, stakeholder accounts, cost information, and expert judgment. These source types answer different questions.
The WHO checklist for routine evidence use emphasises organisational processes that make evidence accessible, appraised, and usable in policy-making. A brief should therefore reveal how sources were identified and assessed rather than presenting citations as decoration. The WHO template also calls for a concise methodology describing evidence searches, appraisals, consultation, local data, and other inputs.
AI can extract candidate passages or propose groupings, but verify every quotation, statistic, date, and source claim against the original. Preserve excluded material and the reason for exclusion. A source register should make visible when the evidence base is narrow, old, indirect, contested, or unavailable.
I: Inspect the problem and compare options
Structure the analysis around the decision, not around the order in which sources were found. Define the current situation, who experiences the problem, what happens without a change, and which mechanisms might connect each option to an outcome. Then compare a manageable set of feasible options, including the status quo when appropriate.
For every option, test:
The UK government communications toolkit describes an effective brief as audience-specific, transparent about evidence, concise, and linked to feasible action. Treat that as communication guidance, not proof that a well-written brief causes policy uptake.
- What would change in practice?
- Which evidence supports the proposed mechanism?
- Who benefits, bears cost, or faces risk?
- What capacity, authority, funding, data, and time are required?
- What could fail during implementation?
- Which uncertainties could change the decision?
- What monitoring would reveal an early problem?
E: Explain implications without hiding uncertainty
Lead with the decision-relevant conclusion, then show the reasoning. Use plain language and define technical terms. Distinguish measured facts, estimates, assumptions, stakeholder views, and author judgment. When evidence conflicts, explain why: populations, methods, dates, settings, or definitions may differ.
Do not collapse uncertainty into a confidence adjective with no basis. State what is uncertain and how it matters. “Local evening attendance data cover only two sites” is more useful than “data confidence is medium.” If a result is an association rather than a causal estimate, say so. If a cost is incomplete, identify excluded categories.
Recommendations should be proportionate to the evidence. Options may include proceed, pilot, sequence, revise, collect evidence, or hold. Add conditions and review triggers. A recommendation for a bounded pilot should not be rewritten as support for full implementation.
F: Finish the artifact, review it, and preserve provenance
Use this compact brief template:
| Section | Required content | |---|---| | Decision | The exact decision, owner, date, scope, and constraints | | Key message | The conclusion in two or three bounded sentences | | Problem | Scale, affected groups, current response, and baseline | | Options | Status quo plus feasible alternatives | | Evidence | Findings, source fit, uncertainty, and gaps | | Implications | Benefits, harms, cost, feasibility, distribution, and dependencies | | Recommendation | Action, conditions, owner, timing, and review trigger | | Monitoring | Indicators, data owner, cadence, and stop or change conditions | | Methods | Search, selection, appraisal, consultation, and cutoff date | | Sources | Traceable citations and version information |
Run three reviews. A subject reviewer checks factual and methodological fidelity. An operational reviewer checks feasibility, ownership, and implementation dependencies. A reader unfamiliar with the work checks whether the decision and reasoning are understandable. Resolve comments in the evidence register, not only in the prose.
Worked example: cooling-centre hours
A synthetic municipal team must decide whether to extend cooling-centre hours during forecast heat alerts. The initial AI draft recommends opening every site until midnight, citing increased heat risk and a neighbouring city's programme. BRIEF exposes three defects: the cited study covers daytime attendance, the local budget excludes security and transport, and the neighbouring programme serves a different population.
The team bounds the decision to a six-week pilot at two sites. Its register includes local hourly attendance, public-health alert criteria, transport schedules, staffing constraints, resident consultation, and evidence about heat exposure. Options are unchanged hours, later hours at every site, or targeted extensions at two accessible sites. The brief explains that local late-evening demand is uncertain and that outreach may affect attendance.
The recommendation becomes a conditional pilot with named owners, accessible transport information, an evening safety review, hourly attendance reporting, and a review after the first two alert periods. A stop condition applies if safe staffing is unavailable. The artifact supports a local decision under stated assumptions; it does not establish a universal policy effect.
Verification and Two Prune boundary
Before release, select five decision-relevant statements and trace each to the source register. Recalculate one numeric claim independently. Confirm that every recommendation has an owner, condition, and evidence link. Check that AI did not invent a source, merge unlike findings, or turn an assumption into a fact. Preserve the draft, source version, review comments, and final disposition.
Two Prune can place a participant in a synthetic policy task and observe framing, evidence navigation, AI orchestration, verification, judgment, and communication. It can report task-bounded assessment evidence with confidence notes. It does not certify policy expertise, determine whether a public policy should be adopted, prove organisational impact, predict future workplace performance, or make employment decisions.
Primary sources
World Health Organization, Policy brief template: how to write an effective policy brief
- World Health Organization, Supporting the routine use of evidence during the policy-making process
- UK Government, Communications toolkit: writing a policy brief
Sources: [1] [2] [3]