Anything you write down at work is potentially a federal record. Anything that is a federal record is potentially FOIA-able. AI did not invent this rule — it just made the speed at which you create records terrifyingly fast.
Build a personal summarization workflow you can use to extract themes and findings from large document collections without (a) creating new FOIA exposure, (b) paste-ing protected information into the wrong tool, or (c) violating your agency's records schedule.
A federal employee can save themselves dozens of hours by summarizing inboxes, comment dockets, contract files, and meeting transcripts with AI. The same employee can also create a six-figure FOIA headache if they paste the wrong thing into the wrong tool. The goal of this lesson is to give you a workflow that captures the upside without creating the downside.
The Freedom of Information Act (5 USC §552) requires federal agencies to disclose records on request unless one of nine statutory exemptions applies. The most relevant exemptions for AI work are:
"Agency record" is broad: emails, draft documents, calendars, instant messages, and yes, the prompt and the response inside an AI tool that you used for official work. The Federal Records Act (44 USC Chapter 31) and 36 CFR Part 1236 reinforce that records do not lose their character because they are stored in a chat interface.
Bo, here is the rule that catches most people: an AI prompt you typed at work and the response Claude returned are agency records if they were created or used in the conduct of agency business. They can be requested under FOIA. They are subject to records retention. They can be subpoenaed. They can be reviewed by your IG.
Practical implications:
Before you do anything else, label every document in your batch with: classification level (Public / FOUO / CUI / classified), exemption flags (any Exemption 4 commercial info? Exemption 6 PII?), and retention status. If anything is classified or CUI, you stop and route it to a tool that is approved for that level. Do not proceed with mixed-classification batches in a tool you are not certain handles all of them.
For documents going into a tool approved for sensitive but unclassified work, redact PII (names, SSNs, dates of birth, contact info), confidential commercial info, and any draft pre-decisional language that should remain inside the agency. Even when the tool is approved, minimum-necessary handling reduces FOIA exposure.
I am summarizing [N] documents for [purpose]. The documents are attached/pasted below. Produce: 1. A 200-word executive summary in plain language 2. A bullet list of every distinct concrete fact, with the source document filename and page or paragraph reference next to each 3. A bullet list of every recommendation, decision, or commitment found, with source 4. A flagged list of: (a) any apparent PII you noticed, (b) any apparent commercial confidential markings, (c) any classification banners 5. Open questions the documents do not answer RULES: - Do not introduce facts that are not in the source documents - If a claim is ambiguous, mark it [UNCERTAIN] and quote the original sentence - Do not paraphrase classification banners or CUI markings — reproduce them verbatim - If the documents appear to contain anything I should not have pasted here, say so.
Save the prompt, the input filenames (not the input text if it was sensitive), and the output to your records-managed shared drive under the appropriate retention category. This is your audit trail. If a FOIA request later asks about this analysis, you have a clean record of what was done, when, with what tool.
This is non-negotiable. Pasting classified information into a system not approved for that level is a security incident under Executive Order 13526 and component-specific spillage policies. Pasting CUI into a system not approved for CUI is a violation of 32 CFR Part 2002 and your agency's CUI program. Both can end careers.
If you think you may have pasted classified or CUI into the wrong tool: stop, do not delete anything (deletion can compound the problem), and report immediately to your security officer. The reporting path matters more than your embarrassment.
Pitfall 1 — Treating chat as ephemeral. AI chat sessions are records under the Federal Records Act (44 USC Chapter 31). Your agency's records schedule applies whether or not the tool surfaces the history visibly to you.
Pitfall 2 — Assuming Exemption 5 covers everything pre-decisional. Exemption 5 (deliberative process) protects pre-decisional and deliberative records, but factual material is generally not protected, and once a decision is made the deliberative protection narrows. Consult your FOIA officer before relying on Exemption 5 to shield AI-assisted work.
Pitfall 3 — Cross-contaminating retention buckets. If the source documents have a 7-year retention but you summarize them into a chat that gets purged in 30 days, you have created a records-management gap. Save the output to the same retention bucket as the inputs.
Always check with your records officer, FOIA officer, OGC, or IG before relying on this lesson for handling sensitive collections. This is general guidance.
Before moving on, make sure you can answer: