Marisol Tran is a constituent services specialist in a U.S. congressional district office. On a normal Tuesday she opens 140 emails, 30 voicemails transcribed to text, and a stack of casework files about delayed passports, denied veterans' benefits, and a flooded basement nobody at the county will return calls about. Her job is to read all of it, summarize the urgent items for the chief of staff by 10 a.m., and draft replies that sound like a human who cares wrote them. For years that meant arriving at 6:30. Last month her office approved a general-purpose AI writing assistant. Marisol cut her morning triage from three hours to forty minutes. She also nearly sent a constituent a letter citing a benefits deadline the AI invented. This lesson is about getting the forty minutes without the invented deadline.

The Volume Problem Every Agency Has

Marisol's inbox is a small version of a problem the whole sector has. Government runs on text: citizen complaints, legislative testimony, agency reports, grant applications, casework files. Budget analysts need to review hundreds of financial reports. Caseworkers need to process thousands of complaints. Policy analysts need to synthesize research across hundreds of sources. Reading all of it thoroughly is impossible, and missing something important is risky. That gap between what arrives and what a person can absorb is exactly where an AI writing assistant is genuinely useful, and it is also exactly where it is most tempting to misuse.

Be honest about the shape of the help. AI cannot eliminate the work. What it can do is augment human capability: handle the initial pass through a document, extract key points, and flag what needs a person's attention. The payoff shows up in four places. The work moves faster. Important items are less likely to be missed. Human hours shift toward judgment and decision-making instead of skimming. And the same staff can process more volume with the same resources. None of that requires trusting the tool. All of it requires supervising it.

What These Tools Are Actually Good At

A modern AI writing assistant is a large language model, software trained on enormous amounts of text to predict what words should come next. It is not a database and it does not "look things up" unless you give it the source material. Knowing that one fact tells you exactly which government tasks it helps with and which it endangers.

It is strong at transforming text you give it: shortening, reorganizing, changing tone, pulling out action items. It is weak at producing facts from its own memory: dates, dollar figures, statute numbers, names. The rule Marisol now lives by is short enough to keep in your head. The AI may rearrange your words, but it may not supply your facts.

Three task families dominate frontline government work, and the AI helps with all three when you feed it the source.

• Summarizing long documents, briefing papers, and constituent letters into something a busy person can act on.
• Drafting correspondence, talking points, and routine reports from a few bullet points.
• Reformatting messy input, such as a rambling voicemail transcript, into a clean intake note.

When Summarizing Earns Its Keep, and When It Does Not

Summarization is not free. Prompting, reading the result and checking it against the original all cost time, so spend that time where it pays. The four situations below on the left are where the trade works in your favor. The four on the right are where it does not, and each of them has burned somebody: a paraphrase changes wording by definition, which is precisely the problem when the wording carries legal weight.

The Summarization Workflow That Holds Up

Marisol's worst near-miss came from a lazy prompt: she pasted a 22-page Department of Veterans Affairs policy memo and typed "summarize this." The AI returned a confident summary that compressed two different appeals deadlines into one wrong number. The text was in the memo. The AI just smoothed it over. The fix is to make the AI point at the source instead of paraphrasing freely, and a government-grade summarization prompt has four parts.

1. Role and audience: "You are summarizing for a chief of staff who has 90 seconds."
2. The source, pasted in full: the actual document, not a description of it.
3. The shape you want: "Five bullets, each under 20 words, plus a one-line 'so what.'"
4. A grounding instruction: "Only use information in the text above. If a deadline or dollar figure appears, quote it exactly. If something is unclear, say so rather than guessing."

That last line is the difference between a tool and a liability. With it, the same VA memo produced a summary that flagged two deadlines and quoted both verbatim, because Marisol forced the model to anchor to the page instead of its imagination.

Underneath those four parts is a four-step loop worth naming, because every skipped step is somebody's incident report. First, give clear instructions: "Summarize this policy document in 3-4 bullet points. Focus on eligibility requirements and restrictions. Use plain language" beats "Summarize this" every time. Second, provide the document, pasted in or uploaded if your tool supports uploads. Third, review the summary: read it, check it for accuracy, ask what is missing and what is wrong. Fourth, use it appropriately. A summary is a starting point, not a replacement for the original. If you are making a decision based on the document, read the relevant sections yourself.

Synthesizing Across Several Documents

The multi-document case deserves its own mention, because that is where the hours actually come back. Say you have 5 policy documents and you want to understand how they relate. The prompt is: "I'm attaching 5 policy documents. Summarize how they relate to benefits eligibility. What are the key differences? Are there any conflicts or contradictions?" The AI can synthesize across documents in ways that would be time-consuming for a human, and it is reading your documents rather than its memory, which is the safe configuration.

Asking explicitly for conflicts and contradictions is the part people leave out, and it is the part that earns its keep, because a conflict between two authorities is the single most useful thing a policy reader can be handed early. Treat what comes back as a set of leads. Each claimed conflict points you at two specific passages, and you go open both. What the tool has saved you is the search, not the reading that follows it.

The Drafting Workflow: From Bullets to a Letter

Drafting runs the opposite direction. Instead of compressing a long source, you expand a few facts into a polished document. Here the danger is not invented facts, since you are supplying those, but tone, the accuracy of your own bullets, and the model quietly adding claims you never made.

Marisol's drafting recipe: give the AI the verified facts, the tone, and the constraints, then read every word before it leaves the building. A real example, lightly disguised. A constituent named Mr. Okafor was denied a small-business disaster loan and wants help. Marisol's input to the AI:

• Facts (verified by me): SBA loan #, denied for incomplete income documentation, 60-day reconsideration window, our office can submit a congressional inquiry.
• Tone: warm, plain, no jargon, eighth-grade reading level.
• Do NOT: promise an outcome, invent any deadline beyond the 60 days I gave you, or cite any statute.
• Length: under 200 words.

The draft came back usable in one pass. Marisol changed two sentences and sent it. The forty-minute morning is built on dozens of small wins like this, not on letting the machine run free.

Her recipe generalizes into five steps you can run on anything. Provide a model or template: "I'm attaching a previous constituent response letter. Please draft a similar response to this new inquiry about [topic]." An example teaches tone and structure faster than any adjective. Be specific about content: "This letter should: thank the constituent, explain why we can't approve the request, suggest alternatives, provide contact information for appeal." Set the tone and audience: "Write this for a non-technical audience. Use plain language. Be respectful but clear." Review and edit, fixing errors and customizing to the specific situation, because your name goes on it. Fact-check anything in the draft that asserts a fact.

Where Drafting Belongs and Where It Does Not

Drafting has a comfortable zone and a forbidden one. It works well on constituent response letters, where the AI drafts and you customize; on internal memos, where the AI drafts and you edit; on policy briefing papers, where the AI drafts and you fact-check and refine; and on meeting agendas, where the AI drafts and you reorganize. Notice the shape of that list. In every case a human takes the last, accountable pass, and that pass has a specific verb attached to it rather than a vague intention to "look it over."

It does not belong on legal documents that carry specific statutory language, where the exact words are the whole point. It does not belong on high-stakes policy decisions; the AI can draft supporting material, never the decision itself. And it does not belong on communications that need absolute accuracy, because fluent text and correct text are different things and only one of them is visible on the page. If a document falls into one of those three categories, the tool can help you think. It should not hold the pen.

The Rules That Keep You Out of Trouble

Frontline government work carries obligations a private-sector marketer never thinks about. Three matter most for everyday AI use.

Never paste sensitive data into an unapproved tool. Constituent Social Security numbers, medical details, immigration status, and case identifiers are protected. If your agency has approved a specific enterprise tool, use only that one, and follow its data-handling rules. A free public chatbot may store and reuse what you type. When in doubt, redact before you paste: replace "SSN 123-45-6789" with "[SSN]" and the AI summarizes just as well.

You are accountable for the output, not the AI. A letter on agency letterhead is an official act. "The AI wrote it" is not a defense to a constituent harmed by a wrong deadline. Every AI-assisted document gets human review before release.

Records still apply. Official correspondence is a federal or state record regardless of how it was drafted. Save the final version where your agency's records system expects it. The AI conversation itself may also be a record, so check your office policy.

A Usable Artifact: The Two-Minute Pre-Send Checklist

Marisol printed this and taped it to her monitor. Run it on every AI-assisted summary or draft before it leaves your hands.

Worked Example: Triaging a Month of Complaints

Here is the whole pattern on one real task. Your agency receives 200 citizen complaints per month. They range from one paragraph to five pages. You need to categorize them, meaning work out what each complaint is actually about, and escalate the most serious ones. Right now you read every one manually and it takes hours. The AI-assisted version starts, counter-intuitively, with more careful human reading rather than less.

1. Pick 20 recent complaints and read them carefully yourself. Build the categories from what you actually find: service delays, staff behavior, billing issues, access issues, other.
2. Use AI to summarize each of the 200 complaints.
3. For each summary, ask the AI: "Which category does this complaint fit? Is it high-priority?"
4. Review the AI categorization. For complaints the AI marked high-priority, read the original. For the ordinary categorizations, spot-check them.
5. Based on the accuracy you observe, refine your instructions and try again.

Once you are confident in the AI categorization the steady state is much faster: read originals only for high-priority complaints, and rely on the AI for routine sorting. Two details separate this from recklessness. The complaints you read first are what make the categories real and give you any basis at all for judging the tool's accuracy. And "confident" describes a sample you checked, not a certificate the tool has earned. Keep spot-checking after the process settles, because the mix of incoming complaints changes and the error rate moves with it.

Building the Habit

The professionals who get the most from these tools are not the ones who trust them most. They are the ones who have internalized where the tool is brilliant and where it lies. Marisol now treats the AI like a fast, eager intern: wonderful at first drafts and tidying, never to be trusted with a fact she has not checked, never given the keys to the official record. That mental model is the whole skill, and it is portable to every tool your agency approves next year.

Anti-Patterns to Avoid

• Using summaries as replacements for reading. You summarize a document with AI and never open the original. Later, a detail matters that was not in the summary. The risk is that you miss important information and never learn that you missed it.
• Sending AI drafts directly without review. The AI produces a draft and you send it to a citizen without reading it first. The draft contains an error or wrong information. The risk is damage to public trust and incorrect information given to the public, in writing, from the government.
• Using AI to summarize sensitive information. You summarize confidential personnel files or medical information with AI, even an AI approved for general use. The risk is sensitive data exposed to a system that is not approved for it. General approval is not approval for every category of data.
• Summarizing without providing context. "Summarize this," with no statement of which aspects matter most. The risk is a summary that leaves out the details that mattered for your purpose while looking perfectly complete.
• Treating a clean spot-check as a clearance. A batch of correct categorizations tells you the tool did well on those items. That is evidence, not a guarantee, and it certifies nothing about the next batch. Keep sampling.

Practice Prompts

• Draft a prompt asking an AI tool to summarize a document you actually have on your desk today. Apply the CRAFT framework from Prompt Engineering Basics, then run it and read the summary against the document line by line.
• Take a letter you sent last month. Give the AI only the verified facts, the tone, and an explicit list of things it must not do, then compare its draft with what you actually sent.
• Identify one high-volume text-based task in your work. Write the prompt, then sketch the workflow around it: who reads the original, who spot-checks, and where the final version gets filed.
• Run the pre-send checklist on your next AI-assisted document and note which row caught something. That row is where your tool is weakest.

Reflection

• Is there a type of high-volume text work in your role where summarization or drafting could help? How would you approach it, and what would the workflow look like end to end?
• What types of documents in your work should never be summarized by AI, and why?
• If a constituent were harmed by a wrong figure in a letter you sent, what would your honest account of your own review look like?

Glossary

• Summarization: Condensing a long document into key points and main ideas.
• Draft: An initial version of a document, meant to be reviewed and edited before finalization.
• Synthesis: Combining information from multiple sources to create a unified understanding.
• Fact-Check: Verifying that claims in a document are accurate.
• Large language model: Software trained on large amounts of text to predict what words should come next. It generates language; it does not look facts up unless you supply them.
• Grounding instruction: A line in your prompt telling the model to use only the supplied text, quote figures exactly, and say when something is unclear rather than guessing.

Related Lessons

• Your Agency's Approved AI Tools tells you which tool you are allowed to paste into before any of this workflow applies.
• Prompt Engineering Basics covers the CRAFT framework the summarization and drafting prompts here are built on.
• Evaluating AI Outputs is the systematic pass to run on a draft before the pre-send checklist.
• AI for Research and Data Organization extends the same discipline from documents to datasets.

Closing

Summarization and drafting are practical applications where AI can genuinely help government work, so use these tools. But remember what they are. They augment human work; they do not replace it. You are still responsible for accuracy and quality, and that responsibility does not thin out because a machine typed the first version. Marisol's forty-minute morning is not a story about a clever machine. It is a story about a specialist who learned exactly where to point one, and exactly where to keep her own hands on the page.

Key Takeaways

• Transform, don't originate. The AI excels at reshaping text you provide; it invents facts when asked to supply dates, dollars, or statutes from memory.
• Ground every summary in the source. Paste the full document and instruct the model to quote figures exactly and flag uncertainty instead of guessing. Vague requests produce vague summaries.
• Summaries are starting points, not replacements for reading. If a decision rests on the document, read the relevant sections yourself.
• Not everything should be summarized. Short documents, legal analysis, text where every word carries weight, and anything needing absolute certainty stay with a human.
• Feed drafts verified facts plus explicit "do not" rules. Supply an example for tone, state what the document must contain, then edit before sending.
• Protect constituent data. Use only your agency's approved tool, and redact personal or protected information before pasting anything.
• You own the output. An AI-assisted letter on letterhead is an official act; human review before release is non-negotiable, and records rules apply to the final document and sometimes to the AI conversation itself.
• Run the pre-send checklist every time. Facts, invented claims, sensitive data, tone, promises, record: two minutes that prevent the wrong-deadline letter.

Frequently Asked Questions

How long should a document be before AI summarization is worth it? Roughly ten pages and up, or any time you have a high volume of similar documents or several related ones to synthesize. A one or two page document is faster to read than to prompt, check and correct.

Can I ask the AI to include the statute or deadline it is summarizing? Yes, if the statute or deadline is in the text you pasted, and you instruct it to quote the figure exactly. Never ask it to supply a citation from memory. That is the failure mode that nearly sent Marisol's invented deadline to a constituent.

What do I do when the AI's multi-document summary reports a contradiction between two policies? Treat it as a lead, not a finding. Open both passages it points at and read them. The tool saved you the search, not the reading.

My agency approved a tool for general use. Can I put personnel or medical files into it? Not on the strength of general approval. Approval for general use is not approval for every class of data. Check what the tool is cleared to handle, and redact before you paste when you are unsure.

Does the AI conversation itself have to be retained? It may be a record. The final correspondence certainly is, regardless of how it was drafted, and it must be saved where your records system expects it. Check your office policy for the conversation.

How do I stop a draft from promising something we cannot deliver? Put the prohibition in the prompt ("do not promise an outcome") and check for it again on the way out. If a sentence commits the agency, soften it to what you can actually do: "we will request," "we will inquire."