Fionnuala Donnelly is a senior legislative analyst at a state health department: a team of nine, a $1.1 billion Medicaid budget to track, and a legislative session that produces roughly 400 bills per year that her team must assess for health policy implications. During the last session, a bill with a 72-hour turnaround requirement arrived on a Thursday afternoon. The bill was 140 pages. Inside that window her team had to produce a section-by-section analysis, flag any provisions that conflicted with existing state Medicaid waiver terms, identify fiscal impacts, and draft a two-page summary for the Secretary. They used an AI document analysis tool for the first time that weekend. It took 40 minutes to generate a structured draft analysis. It took another four hours for her team to verify every claim, correct three errors where the model had misread a provision's effective date, and add the fiscal impact analysis the tool could not produce. The final work product was better than what the team typically produced under that kind of deadline, and it was done before midnight on Friday rather than at dawn on Sunday. She is now rebuilding her team's workflows around AI-assisted drafting and analysis. But she is clear about what the tool is and what it is not.

What AI Can Do for Government Documents

Government work runs on documents. Policy briefs, program descriptions, legislative testimony, regulatory guidance, legal opinions, technical documentation, responses to congressional inquiries. Producing them is time-consuming and cognitively demanding, and it happens under deadlines that do not move. The work falls into two broad categories: drafting, meaning creating a new document from guidelines, source material, or oral direction, and analysis, meaning extracting, comparing, or assessing information from existing documents. AI tools are useful for both, but in different ways and with different risk profiles.

On the drafting side, six capabilities recur across government writing tasks. Outline generation proposes a structure for a document given its topic and purpose. Draft generation produces an initial full draft from guidelines or source material. Section writing handles a specific component, the background, the analysis, the recommendations, rather than the whole. Editing suggests improvements to clarity, grammar and tone in text you already have. Summarization condenses long documents into executive summaries. And citation and reference surfaces potentially relevant regulations, precedents or supporting material, which is the capability that most needs checking rather than trusting.

A model given a clear brief, for example "draft the background section for a legislative testimony on expanding telehealth coverage for rural Medicaid beneficiaries, referencing the three program summaries provided", can produce a usable first draft in minutes. That draft will typically need significant editing for tone, accuracy, and the specific requirements of the document type. But it provides a starting point that is faster than a blank page, especially under deadline pressure.

For analysis, the most useful applications are summarization of lengthy source documents, extraction of specific information across multiple documents, and cross-document comparison, including comparing versions of the same document to see what changed. Fionnuala's 140-page bill analysis is the canonical example: the model can read and structure a document at a speed no analyst can match. The analyst's role shifts from reading to verification, which is faster and, arguably, produces higher quality, because human reviewers are better at catching errors in a structured output than at catching gaps in their own original reading.

The Analysis Tasks in Detail

Analysis is where the productivity case is strongest and the failure modes are least obvious, so it is worth separating the tasks rather than treating "read this document" as one job. Four distinct tasks recur in legislative, regulatory and program work, and each puts a different demand on the human afterward.

Classification sorts incoming documents into categories: which of these public comments raise a legal objection, which of these submissions belong to the waiver program, which of these constituent letters concern eligibility. The model is grouping rather than judging, and the verification question is whether the categories it used match the ones your process actually needs. A model that invents a plausible taxonomy of its own creates rework, not savings.

Summarization condenses a long document to its substance. The verification question here is not "is this accurate" but "is this complete for my purpose", because a summary is defined by what it leaves out. For the Secretary's two-page brief, the risk was never that the summary would assert something false; it was that a provision affecting waiver terms would simply not appear.

Extraction pulls specific fields out of many documents at once: effective dates, dollar amounts, named programs, cross-references to existing statute. This is the task where the effective-date errors in Fionnuala's analysis occurred. Extraction failures are attachment failures, where a real value gets bound to the wrong provision, and they survive a read-through precisely because every individual element is genuine.

Comparison sets documents side by side, either different documents on the same subject or successive versions of one. Version comparison is the highest-value analysis task in legislative work, because a bill that has moved through committee differs from its predecessor in ways that matter enormously and are tedious to find by hand. The check that matters is whether the model reported the changes it found or the changes it expected, which is why a comparison output should be spot-checked against the actual redline rather than accepted as one.

Across all four, the pattern holds: the model produces structure and the analyst supplies judgment about whether the structure is the right one. What the analyst never gets to skip is the question of what is not in the output, because none of these tasks can report an absence the model did not notice.

The Government Document Standard

Government documents operate under standards that do not apply to commercial writing. They must be accurate, well-reasoned and legally sound, and they represent the agency's official position rather than an individual's view. An error is not just a mistake; it is potentially costly and damaging to institutional credibility. Understanding these standards is essential before deploying AI assistance, because every one of them shapes what the tool may be trusted to do.

Legal accuracy is non-negotiable. A policy memo that misquotes a statute, a testimony that misstates a regulatory deadline, or a procurement document that omits a required certification can have legal consequences for the agency and for the people who signed it. AI models hallucinate, generating text that sounds authoritative but is factually wrong. In government document work, the hallucination risk is primarily about citations and specific factual claims: statute numbers, effective dates, dollar amounts, regulatory thresholds. Every such element in an AI-generated draft requires explicit verification against the primary source.

Official voice matters. Government documents have specific tone requirements that vary by document type. A congressional testimony is formal and measured. A grant application is structured and evidence-forward. A plain-language public notice has different requirements than an internal rulemaking memo. AI models trained on general text tend to produce competent but generic writing. The analyst must actively shape the output to match the required register, and must know what that register is.

Classification and handling restrictions apply. Before any agency document, draft or final, is processed through an AI system, the analyst must verify the document's classification level and whether the AI tool has been authorized for documents at that classification. Processing a Controlled Unclassified Information (CUI) document through an unauthorized commercial AI tool is a security incident, not a productivity strategy. Most commercial AI tools are not authorized for CUI without specific agency authorization. The general rule that follows is to understand how a tool handles the data you give it and to use government-controlled systems for sensitive content.

Authorship accountability is assigned to the human who signs. The official author of a government document is the person or office that approves and issues it. AI assistance does not transfer or diffuse that accountability. If an AI-drafted memo contains an error that leads to a policy problem, the accountability rests with the official who approved the memo, not with the AI tool or the analyst who used it. This is not a theoretical concern; it is the framework under which congressional testimony, Inspector General cooperation, and administrative law proceedings operate.

How AI Drafts Go Wrong

The failure modes are specific and worth naming individually, because each one is caught by a different check. Naming them is also what stops a team from treating "review the draft" as a single undifferentiated instruction that everyone performs differently.

• Hallucination. The system generates plausible-sounding but false citations or facts. In government writing this concentrates in exactly the elements that carry the most weight: statutory references, dates, thresholds and figures.
• Bias. The system's language reflects biases present in its training data, which can show up as gendered or demographic assumptions in how people and populations are described.
• Inaccuracy. The system oversimplifies a complex issue or gets details wrong, producing text that is not fabricated but is still not right.
• Tone misalignment. The document does not sound appropriately governmental or formal for its type and audience, which is usually obvious to any experienced reader and damages credibility on sight.
• Confidentiality exposure. The document, or portions of it, passes through an external service that the agency has not authorized for that material.

Fionnuala's weekend produced one instance of the third failure and one near-miss on the second. The three effective-date errors were misreadings rather than inventions: the provisions existed, the dates existed, and the model attached them to each other incorrectly. That is precisely the kind of error a fluent draft hides best, because nothing about the sentence looks wrong.

Bias deserves separate attention in government writing because it is the failure least likely to be caught by a factual check. A draft that describes a beneficiary population in terms that carry an unstated assumption, or that consistently reaches for one framing of a group and not another, will pass every citation verification cleanly. The text is not false. It is loaded. In documents that describe the people a program serves, or that will be read by them, this is a substantive defect rather than a stylistic one, and it needs a reader who is looking for it specifically. The general risk of AI-generated content is that it is uniformly fluent, and fluency is exactly what makes a reader stop reading critically.

The Verification Protocol

Every AI-assisted government document needs a structured verification pass before it can be approved. What that pass covers depends on the document type and stakes, but the minimum standard for any consequential document includes the checks below. The governing rule sits above all of them: never publish AI-generated content without human review and approval.

Citation check. Every statutory reference, regulatory citation, case reference, and data point in the AI draft should be verified against the primary source. This means pulling up the actual statute, regulation, or dataset, not a secondary summary, and confirming that the citation is accurate and that the AI's characterization of what it says is correct. This is the check that catches the errors most likely to be consequential, and it catches only those. It does not touch the logic and completeness failures that the remaining checks exist for, so a clean citation pass is not a clean document.

Logic check. Does the argument hold? Does the analysis follow from the evidence? AI models can produce text that sounds logically coherent but contains non sequiturs or unsupported leaps. A human reviewer who knows the subject matter will catch these; a reviewer who is not paying careful attention will not. The verification pass should explicitly include a step where the reviewer asks: "Does this conclusion follow from these premises?"

Completeness check. Does the document address everything it needs to address? AI models excel at writing about what is in the source material and are poor at noticing what is missing. For Fionnuala's bill analysis, the completeness check meant asking whether the team had addressed every section of the bill, every applicable waiver term, and every fiscal element the Secretary would ask about. The AI tool had missed one provision entirely because it appeared in an amendment rather than in the main text.

Tone and voice check. Does the document sound like it was written by this agency for this audience? This check is especially important for public-facing documents and for documents that will be attributed to a specific official. Tone misalignment is usually obvious to any experienced reader and damages the agency's credibility.

Accountability check. Is it clear who is approving this, and does that person know what they are signing? The last line of the protocol is not about the text at all. It is the confirmation that a named human has read the document, understands the claims in it, and is prepared to defend them, because that is the person who will be answering for them.

AI assistance in government document work is not about writing faster. It is about giving analysts more time to think carefully, by handling the mechanical first-draft work and freeing human attention for verification, judgment, and the parts of the document that genuinely require expertise.

Building the Human-AI Workflow

The most effective implementation pattern is not "AI writes, human approves." It is "AI drafts, human verifies and edits, human owns." This means the analyst is engaged throughout the process, not just at the end.

A practical workflow for analysis tasks: the analyst provides the AI tool with the source documents and a structured prompt specifying what needs to be extracted or assessed. The model produces a draft analysis. The analyst works through the draft section by section, using it as a scaffold rather than a finished product, verifying each claim, filling gaps, and rewriting where the model's output is imprecise or off-register. The analyst saves the verified version as the working document. The draft analysis, with its errors, is not the product; it is the tool.

For drafting tasks, the workflow is similar but the analyst's role shifts: the analyst provides a detailed brief with all the relevant facts and requirements, reviews the model's draft for structure and completeness before editing for accuracy and tone, and inserts the specific evidence and citations that the model cannot reliably generate on its own. The analyst's time savings come from not having to produce the document structure from scratch under deadline, not from skipping careful review of the content.

Deadline pressure is the condition under which all of this is actually tested. Turning around a briefing by tomorrow morning, revising legislation on committee feedback, drafting a response to a congressional inquiry: these are the moments when AI assistance helps most and when the discipline erodes fastest. Used appropriately, the tools help with speed without sacrificing quality, and the qualifier is doing real work in that sentence. The key is treating AI as an assistant that accelerates human work rather than a replacement for human judgment and accountability. A workflow that only holds when there is time to spare is not a workflow; it is an aspiration that will be abandoned on the Thursday afternoon it was designed for.

Practices that make the workflow durable

Seven habits separate a team that uses AI well from one that merely uses it. Use AI for initial drafts rather than final versions. Have domain experts review and revise those drafts, not just any available reader. Verify all facts and citations. Try the tool on low-stakes documents before you rely on it for important ones, treating that as a way to learn the tool's failure modes rather than as clearance for higher-stakes work. Understand the tool's limitations explicitly, so you know where its output is weakest. Maintain version control that tracks which passages were AI-drafted and which were human-written. And document that AI was used in drafting, both for transparency and for quality assurance, so a later reviewer knows which parts of the record to scrutinize hardest.

A Worked Example: Five Regulations, Five Briefings

A policy analyst needs to draft briefings on five new regulations. Working with an AI tool, the sequence looks like this. The system generates an outline for each briefing, and the analyst reviews and revises it. The system drafts the background sections, and the analyst edits for accuracy and agency voice. The system summarizes each regulation, and the analyst verifies the summary against the regulation itself. The system suggests implications, and the analyst evaluates them and enriches them with organizational knowledge the model does not have. The analyst then produces the final document.

The reported result is that the briefings were completed in 60 percent of the time manual drafting would have taken, with quality that was good and perhaps better than usual because the review was systematic. The analyst retained full accountability and control throughout. Read the ordering carefully, because it is the whole lesson: at every step the model produces and the analyst disposes, and the analyst's contribution is heaviest on exactly the elements the model handles worst, namely accuracy against the primary source, agency voice, and the organizational context that appears in no document at all.

Anti-Patterns to Avoid

• Publishing without review. Treating the AI draft as a finished product and pushing it out with a glance. This is the failure the entire protocol exists to prevent, and it happens under deadline rather than out of ignorance. Require documented review before any document is published.
• Confidentiality breaches. Pasting sensitive or controlled material into a cloud service the agency has not authorized for it. Understand each tool's data handling before you use it, and route sensitive content through government-controlled systems. A tool being useful is not the same as a tool being authorized.
• Over-reliance. The analyst becomes dependent on the AI and stops exercising independent judgment, accepting drafts that read well rather than drafts that are right. Frame the tool as an assistant and maintain a critical stance toward its output as a standing habit, not an occasional audit.
• Inaccuracy not caught. The model produces plausible-sounding but false information and a busy analyst does not catch it. Systematic fact-checking, citation verification against primary sources, and peer review are the defences, and they only work when they are scheduled rather than intended.
• Treating a low-stakes trial as clearance for high-stakes work. The tool performed well on routine internal memos, so it is adopted for testimony and rulemaking. Performance on low-consequence documents tells you the tool works on low-consequence documents. It is evidence about the tool's behaviour, not authorization for a different risk tier, and the classification question has to be asked again for every tier.
• Treating the checklist as the guarantee. The verification protocol is a set of prompts for human attention, not a certification. Every box can be ticked by someone who skimmed, and the ticks will look identical to the ones made by someone who pulled up the statute. What makes a document defensible is the reviewer who can say where each claim came from, not the completed form that says a review occurred.
• Letting the citation check stand in for the completeness check. Citations are the most satisfying thing to verify because verification either succeeds or fails cleanly. Completeness has no such signal, which is why it is the check most often skipped and the one that caught the missing amendment provision in Fionnuala's analysis.

Practice Prompts

• Audit your own document work. Which of the documents you create regularly could benefit from AI assistance, and which are too sensitive or too complex for it? Write down what puts each one in its category.
• Design a quality assurance process for AI-assisted documents in your office. What would trigger rejection of a draft? What must be verified before approval, and by whom?
• For your most sensitive document type, assess the risks of using AI to draft it. Is it appropriate at all? If so, under what safeguards, and who decides?
• Evaluate the AI writing tools available to you. Which would be appropriate for government use in your context, and what are their limitations for the documents you actually produce?
• Design an end-to-end process for using AI to draft policy documents in your organization. How would review work, who holds final approval, and how would the record show which passages were AI-drafted?
• Take one AI-generated paragraph containing a statutory citation and verify it against the primary source. Time how long it takes, and note whether the model's characterization of the provision was accurate as well as the citation itself.

Reflection

• In your last three documents, which specific claims would have caused real harm if they had been wrong, and who would have absorbed that harm?
• When deadline pressure is highest, which verification step in your workflow is the first to be dropped? What would have to change for it not to be?
• If an Inspector General asked you today which passages of a recent memo were AI-drafted, could you answer from your records?
• Your agency's approval signature sits on documents you drafted. Does the person who signs understand which parts of the draft were generated and what was verified before it reached them?

Glossary

• Hallucination. When an AI system generates plausible-sounding but false content, most often citations, dates, figures or factual claims.
• Controlled Unclassified Information (CUI). Information that requires safeguarding or dissemination controls, and that may not be processed through an AI tool the agency has not authorized for it.
• Primary source. The actual statute, regulation, case or dataset itself, as opposed to a summary or characterization of it. Citation verification means checking against the primary source.
• Official voice. The tone and register required by a specific government document type and audience, which differs between testimony, guidance, grant applications and public notices.
• Completeness check. The verification step that asks whether the document addresses everything it needs to address, as distinct from whether what it says is accurate.
• Authorship accountability. The principle that the official who approves and issues a document is responsible for its content regardless of what tools were used to produce it.
• Version control for AI drafting. Tracking which portions of a document were AI-generated and which were human-written, so that later reviewers know where to concentrate scrutiny.

Related Lessons

• AI for Government Tasks: Summarization and Drafting covers the same tool set at task level, with more detail on how to brief the model.
• AI Confidence and Hallucination explains why a fabricated citation arrives sounding exactly as confident as a real one.
• Data Sensitivity and Classification is the prerequisite for the classification question this lesson says to ask before every document.
• Data Leakage: When Sensitive Info Enters AI takes up what actually happens when controlled material reaches an unauthorized service.
• Quality Assurance for AI Work Products generalises the verification protocol beyond documents.
• Prompt Engineering Mastery: Structured Prompts is how you write the detailed brief that determines whether the first draft is usable.
• Your Agency's Approved AI Tools answers the authorization half of the question this lesson raises.

Closing

The point of AI-assisted document work in government is not speed for its own sake. Fionnuala's team did not get their weekend back because the tool wrote the analysis; they got it back because the tool absorbed the mechanical first pass and left the four hours of verification, judgment and fiscal analysis to people who could actually do it. That is the trade worth making, and it only holds if the verification actually happens. The tool changes where the analyst's attention goes. It does not change who signs, who is accountable, or what happens when a statute number in a document bearing the agency's name turns out to be wrong.

Key Takeaways

• AI drafting saves the most time on first-draft production under deadline. Generating a document structure and first draft in minutes rather than hours is where AI assistance delivers the clearest time savings in government document work.
• Citation verification is mandatory for every AI-assisted document. AI models hallucinate statute numbers, regulatory citations, and data points. Every specific factual claim in an AI-generated draft must be verified against the primary source before the document is finalized.
• Name the failure modes separately, because each has its own check. Hallucination, bias, inaccuracy, tone misalignment and confidentiality exposure fail differently and are caught differently. "Review the draft" is not an instruction anyone can follow consistently.
• Classification restrictions apply before processing. Documents containing CUI or other restricted information may not be processed through unauthorized commercial AI tools. Verify the document's classification and the tool's authorization before using AI assistance.
• The human who approves the document owns the accountability. AI assistance does not distribute or reduce authorship accountability in government. The official who approves a document is responsible for its accuracy, regardless of what tools were used to draft it.
• Completeness is the verification gap AI drafts most often leave. AI models write about what is in the source material and miss what is absent. The completeness check is the step human reviewers most often skip and most often should not.
• Official voice requires active shaping by the analyst. AI-generated text defaults to a competent but generic register. Government documents require the analyst to actively adjust tone for the specific document type, audience, and issuing official.
• The workflow is "AI drafts, human verifies and owns." The most effective model keeps the analyst engaged throughout, using the AI output as a scaffold rather than a finished product, rather than treating AI assistance as a process that reduces human involvement in content.
• Record that AI was used, and where. Version control that distinguishes AI-drafted from human-written passages serves transparency and gives later reviewers a map of where to concentrate scrutiny.

Frequently Asked Questions

Can I paste an internal draft into a commercial AI tool if it is not classified? Not on the basis of "not classified" alone. The question is whether the material is controlled in any way, including CUI, and whether the specific tool has been authorized by your agency for material at that level. Most commercial tools are not authorized for CUI absent specific agency authorization. Check the document's handling requirements and your agency's approved tool list before the paste, not after.

The model gave me a citation. Do I still need to look it up? Yes, every time. Citation and reference generation is the capability that most needs checking rather than trusting, because a fabricated citation arrives formatted exactly like a real one. Pull the primary source, confirm the reference is real, and confirm that the model's characterization of what it says is accurate. Those are two separate failures and both occur.

If AI drafted the memo and I approved it, who is accountable for an error? You are. The official author of a government document is the person or office that approves and issues it, and AI assistance neither transfers nor dilutes that. This is the framework under which congressional testimony, Inspector General cooperation and administrative law proceedings already operate, and using a new drafting tool does not change it.

Should we tell people that AI was used? Document it in the record as a matter of course, for transparency and for quality assurance. Tracking which passages were AI-drafted lets a later reviewer, an auditor, or your own successor know where verification mattered most. Whether and how that fact appears in a published document is a policy question for your agency rather than a decision for an individual analyst.

We tested the tool on routine memos and it did well. Are we ready to use it for testimony? That test tells you how the tool behaves on routine memos, which is genuinely useful for learning its failure modes. It is not clearance for a higher-consequence document type. Testimony carries different accuracy stakes, a different voice requirement, and potentially different handling restrictions, so the appropriateness question and the authorization question both have to be asked again.

Our verification checklist is complete on every document. Is that enough? A completed checklist records that someone said the checks were done. It cannot record whether they were done well, and a skimmed citation check produces the same tick as a verified one. Treat the checklist as a set of prompts for attention, and treat the reviewer who can say where each claim came from as the actual control.