The question
A records request lands on your desk. It names a solicitation you worked on last spring and asks for all related records, including any conversations with AI tools. You did use one. You drafted the evaluation criteria with it.
So what do you hand over? Maybe the prompt you typed. Maybe the raw output. Maybe the note you wrote about how the draft was made, if you wrote one. Maybe nothing, if the chat never left the tool.
The brand of the tool doesn't decide it. What the chat documents, whether it made it into your files, and which law governs your office do. Federal, state, and local rules get there by different routes, and courts haven't settled the question yet.
This issue plays one AI-assisted task against a records request, gives you a prompt that checks your work note before it goes in the file, and lays out what each level of government is working with.
Field notes: what a records request would find
I used a large language model to draft the technical evaluation criteria for a municipal fleet electrification RFP. It was a complex buy, and evaluation criteria are exactly the kind of drafting where a first pass from AI earns its keep.
I did the data-safety work before anything went in. Department names came out. So did the local garage locations, the internal budget figures, and the names of current vendors. In their place went placeholders: [City_X], [Budget_Y].
Now play a records request against it. Suppose one had come in for that solicitation. What could I hand over?
Three things. First, the prompt, the exact instruction I typed: "Draft a five-point technical evaluation matrix for a public transit EV charging contract..." Second, the raw output, unedited, placeholders and all. Third, the final approved RFP, which shows where a person did the work the machine could not: the local regulations, the legal boilerplate, the real numbers.
That is a decent package. It shows the AI drafted and a human decided. It is not the whole story.
Two things I had to go looking for. The first was the model version. The platform updated overnight, and the update changed how it structured text. "I used AI" does not explain why a draft reads the way it does. Which model, on which date, does.
The second gap was worse. During the session I uploaded background reference files, including the city's green initiative policy. I did not log which version. Later it was hard to prove which version of that policy the AI had been reading. A criteria matrix built on the wrong version of a policy is a problem. A record that cannot tell you which version it was built on is a second problem.
Then the part that went wrong. The AI cited a federal transit guideline that had been replaced two years earlier. It read as authoritative. It was out of date. My process had no step-by-step verification built in, so the error got through drafting. Human peer review caught it. Review did its job. The fix was a time-consuming manual rewrite of the compliance section.
Here is what that file should have carried, and what I would tell anyone to write down. The prompt. The model and its version. The date. Every reference file you uploaded, with its version. And one line saying who checked each regulatory citation against the source, and when.
None of that takes long. All of it is hard to reconstruct after the fact.
Where the chat lived does not, on its own, decide whether it is a record. What it documents does. Mine documented government business. Write yours so you would be comfortable handing it over.
The prompt: write it to be read
Use it on your AI-use note before the note goes in the file.
You are reviewing a short work note that a government procurement
employee wrote about how an AI tool was used on a task. The note
will go into the official file and could be released under a
public records request. Review only the note and the source list
I paste below. Do not use outside knowledge.
Check the note for five things and report each one as OK or FIX:
1. Tool and version: does it name the AI tool, the model or
version, and the date it was used?
2. Inputs: does it list every file or reference uploaded, with
its version or date?
3. Human judgment: does it say what the person changed, kept,
or rejected, and why? Flag a note that records only what
the tool produced.
4. Verification: does it say who checked each regulatory or
legal citation against the source, and when?
5. Tone: flag any sentence that is inaccurate, speculative,
flippant, or that you would not want read aloud by a
reporter or an auditor.
For every FIX, quote the sentence or name the gap, and suggest
a replacement in plain language. Do not invent facts to fill a
gap. Where information is missing, write "MISSING - ask the
author" instead.
WORK NOTE (public or anonymized text only):
[paste here]
SOURCES THE NOTE REFERS TO (titles and dates only):
[list here]The guardrail. Paste only public or anonymized text. If the note names a vendor, a price, an evaluator, or anything from an active source selection, replace it with a placeholder first, the same way the Field Notes did. Titles and dates of your sources go in. The source documents themselves stay out.
The verification step. The tool can tell you a version number is missing. It can't tell you the one you wrote is right. Read the revised note against the actual files yourself, and confirm every date and version before the note goes in the record.
What it returned
On November 3 the prompt was run on Claude with a fictional work note full of deliberate gaps: "used ChatGPT on March 4," an undated policy upload, a guideline that "turned out to be outdated, so we fixed it," and a closing remark that the tool "knows this stuff better than most of our new hires."
It marked all five checks FIX. It wrote "MISSING - ask the author" for the model, the year, and who checked the guideline, instead of inventing them. It noticed that the outdated guideline wasn't among the uploaded files, so the tool had supplied it. It caught an inconsistency I hadn't planted on purpose: the note said "five factors" while describing a sixth that was cut. And it recommended deleting the new-hires sentence outright.
What it missed
It suggested adding a time-saved figure "if the author can support one." Leave that out. Time saved isn't what the file has to show; who decided, and on what, is. It also added an "account type" field the prompt never asked for. That's a sensible question, but check it against your agency's rules rather than taking the tool's word on what belongs in the note.
What could go wrong
Two failures, in opposite directions.
The first is the record you can't produce. The chat that shaped a decision lived only in the tool, the history was cleared or the account closed, and the file holds a final document with no trace of how it got there. Or a note existed and was deleted early because nobody thought of it as a record. Decide where AI-assisted work notes live before anyone asks for them. That's a records management question, and your records officer has the answer.
The second is over-redaction. Faced with an AI chat in a response, it's tempting to black out everything that looks like drafting. Exemptions protect specific interests. They don't cover material because it's awkward or machine-written. At the federal level an agency withholds only when it reasonably foresees harm to an interest an exemption protects, and it has to release the parts that can be separated out.
Close calls go to your records officer or counsel. Not to the tool, and not to instinct.
Your jurisdiction
Federal. FOIA (5 U.S.C. 552) starts from disclosure: since the FOIA Improvement Act of 2016, an agency may withhold information only when it reasonably foresees harm to an interest an exemption protects, or when the law prohibits disclosure. The exemptions you'll meet most in procurement are Exemption 4, for trade secrets and confidential commercial or financial information, and Exemption 5, which covers internal agency material, including the deliberative process privilege. For records retention, NARA's August guidance (see What changed) treats a prompt or output as a federal record when it is captured in an agency system and used for official business.
State. Cite your own act by its current name and section, checked at your legislature's site rather than pulled from memory or a tool. California's Public Records Act now begins at Government Code section 7920.000, after a recodification operative January 1, 2023 replaced the old section 6250 numbering, and Texas's Public Information Act is Government Code chapter 552, which the Texas Attorney General's handbook says covers electronic communications on any device when they relate to official business. Florida's public records law, chapter 119, reaches material "regardless of the physical form" made or received in connection with official business, and it sits apart from the state's open-meetings law at section 286.011.
Local. In most places your state's act applies to you directly, and your clerk or records officer runs the response. Tell them early that AI-assisted drafts exist and where they're kept, so the search covers them. If your office has no format for an AI-use note, the five checks in this issue's prompt are a workable start.
Contractor. At a federal agency, Executive Order 12600 requires the agency to have procedures to notify a submitter of confidential commercial information, and give it time to object, before releasing that information under FOIA. Notice turns largely on what you designated as confidential when you submitted it, so mark it precisely; the order also lists exceptions and creates no enforceable right. State and local rules vary, so check the buying office's public records rules before you submit.
What changed
August 21 — NARA issues guidance on AI materials. The National Archives issued AC 11.2026, Guidance on Applying the Federal Records Act to Artificial Intelligence Materials. It says agency use of AI platforms does not, in and of itself, create federal records. Prompts and outputs become records when they're captured in an agency system and used for official business, and AI-related records still need a NARA-approved schedule before disposal. It covers records management only, not FOIA.
September 14 — No court has ruled yet. Writing in the ABA's Communications Lawyer, media lawyer Steve Zansberg reports that no court has yet published a decision applying FOIA or a state public records act to officials' AI chatbot conversations. Some agencies have released employees' chat logs, and others have refused. Until a court rules, your office's practice is the working rule.
October 19 — FAR overhaul comments close. Comments closed on the four Revolutionary FAR Overhaul proposed rules published September 18 (FAR Cases 2026-003, 2026-006, 2026-010, and 2026-011). As of November 6, acquisition.gov lists no newer proposed rule, and we found no final rule. Eight of the twelve planned proposed rules are out.
From the course
The public records rules in this issue come from Module 2 of Commercial GenAI for Public Sector Procurement, Regulatory Compliance Across Government Levels. Its Video 2.5, Public Records and Transparency, is free to watch before you buy. Module 2 has six videos and runs about 3.5 hours across federal, state, and local requirements, at $395. Eligible for approximately 3.5 CLPs as self-directed study, subject to agency approval — The Wolverine Group is not an accredited CLP/CEU provider. The course is live, with Modules 1 and 2 available now. Module 3 releases Tuesday, November 10.
Your turn
Has anyone ever asked for your AI chat history, in a records request, an audit, or a protest? What did you hand over, and what do you wish you'd written down?
Send a note through the contact page. The answers shape future issues, and anything I write up is anonymized.
Know someone who handles AI-assisted procurement work? Send them this issue. If someone sent it to you, subscribe to The Wolverine Brief for the next one.
About The Wolverine Group
The Wolverine Group is a Woman-Owned Small Business in Washington, D.C., delivering information operations, open-source intelligence, and acquisition solutions to the Department of Defense, the Intelligence Community, and federal civilian agencies. For twenty-five years we have turned complex, multi-source data into decisions leaders can act on. Our acquisition practice supports federal buyers directly and trains the workforce that does the buying — which is where this brief comes from. UEI VK81K1DMC6L5 · CAGE 6BUU5 · CMMC Level 2.