AI for policy analysis

A policy brief is not an essay — it is a decision instrument. Someone with a budget and a signature reads two pages and chooses. AI can compress days of drafting and analysis into hours: it reads the documents, structures the argument, generates options, and stress-tests your logic. What it cannot do is own the decision. In policy work the division of labour is strict — AI drafts, you own the judgement, and you certify every figure. A brief on a county budget, a national tax measure, or a fuel subsidy that carries one hallucinated statistic can misdirect real money and real lives. That is the standard this module holds you to.

Turning data and evidence into insight

The work runs in stages: gather the evidence, structure the argument, draft, appraise the options, stress-test, verify, and package for the decision-maker. AI touches every stage; you gate the last two. The single most important habit is grounding. Never ask a model for 'Kenya's 2024 fiscal deficit' or 'the poverty rate in Turkana' from its memory — it will answer plausibly and sometimes wrongly. Instead, feed it the source: the KNBS Economic Survey table, the CBK data release, the KRA revenue statement, the National Treasury Budget Policy Statement. The model becomes a fast reader and summariser over a source you control, not an oracle you trust.

The policy brief: problem → evidence → options → recommendation

Every good brief follows the same spine, with a one-paragraph executive summary at the top that states the recommendation first — BLUF, bottom line up front. A busy principal secretary reads the first paragraph and the options table; everything else is there to be checked, not read. AI is excellent at producing this shape quickly and at keeping each section disciplined. But the problem definition, the weighting of trade-offs, and the recommendation itself are judgement calls the analyst owns. Ask the model to draft; never ask it to decide.

• Problem: one paragraph — what is the decision, why now, and what is at stake if nothing changes.
• Evidence: the data and analysis, every figure cited to a named source; this is where grounding matters most.
• Options: three to five realistic choices, each with costs, benefits, and who is affected — usually a matrix scored against explicit criteria.
• Recommendation: the option you back, why, what it costs, the main risks, and the first implementation steps.

You are a policy analyst supporting Kenya's National Treasury. Draft the EVIDENCE section of a two-page policy brief on the fuel subsidy.

Use ONLY the sources pasted below. For every figure or claim, add a bracketed citation to the source and its page or table, e.g. [EPRA pump-price release, Table 2]. If a claim is not supported by the pasted sources, do NOT make it — instead write [NOT IN SOURCES — verify]. Do not invent, estimate, or round figures. British spelling.

Structure: four to six short paragraphs — (a) what the subsidy costs the exchequer, (b) who benefits by income group, (c) the effect on pump prices and inflation, (d) the fiscal and IMF-programme context. End with a bulleted list of the three most important numbers a decision-maker needs, each with its citation.

SOURCES:
[paste the EPRA release, the National Treasury BPS extract, the KNBS CPI table, the KRA petroleum-tax note]

Evaluating options and scenario analysis

Good options analysis is structured, not narrative. Define the criteria first — annual fiscal cost, distributional impact, effect on inflation, administrative feasibility, political risk — then score each option against every criterion in a matrix. AI is strong at building this matrix, surfacing options you had not considered, and articulating the trade-offs. Scenario analysis goes further: hold an option fixed and vary the world around it — oil price up 25%, the shilling down 10%, uptake half of forecast — and trace what breaks. The value is not the model's numbers, which you must supply and check; it is the model's speed at working through the logic and flagging which assumption each conclusion hangs on.

You are helping appraise policy options on the fuel subsidy. The four options are: (1) keep the subsidy as is, (2) cap it at a fixed monthly budget, (3) taper it over 12 months, (4) remove it and redirect the saving to targeted cash transfers.

Build an options-versus-criteria matrix. Criteria: annual fiscal cost, distributional impact (who gains most), effect on headline inflation, administrative feasibility, and political risk. For each cell give a short qualitative rating and state the ASSUMPTION behind it. Do not put a specific figure in any cell unless I have supplied it in the sources — where a figure is needed but not supplied, write [NEEDS SOURCED FIGURE].

Then run three scenarios on Option 3 (taper): oil price falls 15%, oil price rises 25%, the shilling depreciates 10%. For each, describe the direction of the effect on the fiscal cost and on pump prices, and flag which of my assumptions the scenario is most sensitive to. British spelling.

Stakeholder analysis: who wins, who loses, who blocks

Every policy creates winners and losers, and a brief that ignores them is naive. Stakeholder analysis names three things: who gains, who loses, and who can block. On a fuel subsidy the list runs long — matatu and boda operators, long-distance hauliers, farmers running diesel pumps, urban commuters, manufacturers, the National Treasury, EPRA, the county assemblies, and the IMF programme team. AI will enumerate stakeholders, infer their interests, and war-game the objections you will face in the committee room. What it cannot supply is the local political read — which MP chairs which committee, which county is an election battleground, which lobby quietly funds which campaign. The model maps the logic; you supply the politics.

• Interest: what does this actor stand to gain or lose in concrete terms?
• Power: can they accelerate, delay, or block the decision — and through what channel?
• Position: supportive, opposed, or undecided, and how strongly?
• Mover: what single change would shift them — a carve-out, a transition period, a targeted transfer?

Long government documents, and the policymaker deck

The highest-leverage use of AI in policy is reading what you do not have time to read: a 300-page Budget Policy Statement, a Finance Bill, an Appropriation Bill, a county fiscal strategy paper. Load the actual PDF into a grounded tool — NotebookLM, or a long-context model such as Gemini or Claude — and ask it to summarise, extract the money clauses, and answer specific questions. Grounded tools cite the page, and that citation is the whole point: trust a figure only when the model can point to the page and you open that page and see it. A summary you cannot trace back to the document is a rumour, however fluent. Verify against the source, always.

The last step is packaging. Policymakers decide from decks as often as from briefs, and the same discipline applies: bottom line up front, one message per slide, the options and the ask visible without scrolling. Ask AI to turn your verified brief into a slide outline — a title that states the recommendation, a problem slide, an options matrix, a cost-and-risk slide, and a single clear ask — then build it in Gemini in Slides, Copilot in PowerPoint, or by hand. Never let the deck introduce a figure the brief has not already verified; the deck inherits the brief's evidence, it does not create new claims.