AI and Policy Analyst Memo Craft: One Page That Decides

Policy memos fail at the page-one paragraph; AI structures the lede that earns the next paragraph.

What AI does well here

• Draft a tight executive lede with the recommendation
• Suggest a 3-option comparison framing
• Format risks and uncertainties without burying them

What AI cannot do

• Replace actual policy expertise
• Predict the principal's political constraints

The one-page bar

A policy memo earns its existence in the first sentence. Decision-makers read the BLUF, scan the options, and either commit or kick it back. AI is a strong drafting partner for this format because it forces the analyst to externalize the implicit logic: what is the question, what are the realistic options, what does each cost, and what triggers each one. The risk is that AI produces fluent prose that hides weak reasoning. The analyst's job is to keep the chain of logic visible.

Memo skeleton

1. 1BLUF: one sentence answering the decision being requested
2. 2Three options with cost, risk, and political tradeoffs side by side
3. 3Recommended option with the trigger that would make you switch
4. 4Open questions and the data that would resolve them

How to drive AI without letting it drive you

Start by writing the BLUF yourself. Make AI work backwards from your one-sentence answer to develop the three options that surround it. This forces AI into a supporting role and prevents the failure mode where polished prose disguises a recommendation you have not actually thought through. After AI returns the options, attack each one — what would the strongest opponent of this option say in a hearing? Where does the cost estimate come from? What is the political sponsor who would carry it? AI is excellent at producing the steelman version of an argument when you ask. It is worse at noticing when no one in the actual building wants any of the three options. That is your job.

Citation hygiene checklist

• Pull every cited statute, case, or regulation from the primary source — not from a summary.
• Verify every numeric claim has a published source, dated within the last 24 months when possible.
• Flag estimated figures clearly with the word "estimate" and the methodology in a footnote.
• Never let AI invent the name of a study or report — search to confirm it exists.