Legal Holds for AI Chats: A Legal Ops Playbook

Litigation holds had to absorb email, then team chat, then text messages. Generative AI is next, and most AI tools keep a history. A federal discovery order in Miami on October 21 and law firm guidance published October 29 point to the same practical conclusion: courts treat AI prompts and outputs as ordinary electronically stored information, so legal operations teams should treat them that way in hold notices, retention settings and policy.

This is operational guidance, not legal advice. The scope of any hold is a call for the lawyers on the matter.

What the Miami Order Decided

In Fontainebleau Florida Hotel, LLC v. Botach, a defamation case in the Southern District of Florida, the hotel sought the defendant’s communications with AI platforms since December 3, 2024, through a document request and through subpoenas to four AI service providers. Magistrate Judge Marty Fulgueira Elfenbein’s second order after a discovery hearing, entered October 21, allowed the requests about the hotel, its counsel and its people. It rejected broader requests about the litigation, court orders and the judges, and it rejected a blanket work product objection while letting the defendant log specific documents on a privilege log.

The court refused to make the defendant consent to the provider subpoenas but allowed narrowed Rule 45 subpoenas to the providers. According to the order, the hotel wanted the providers’ records because of concerns that the defendant may have deleted or failed to preserve responsive AI communications.

“The Court emphasizes that a party’s communications with and use of AI does not automatically establish either discoverability or protection,” the order states.

This is a magistrate judge’s interim discovery ruling in one case, and the order says it does not compel any provider to produce anything. What matters for operations teams is how ordinary the analysis was. The court applied the same relevance, proportionality and work product rules it would apply to email, and it said neither of the two 2026 AI decisions the parties cited, United States v. Heppner and Warner v. Gilbarco, sets a categorical rule.

Why AI Chats Belong in the Hold

Katten’s October 29 advisory makes the preservation case directly. It says AI prompts, outputs and conversation logs are generally treated as ESI under Rule 34, that a failure to preserve them is analyzed under Rule 37(e), and that a user who can log in and export a chat history likely controls those conversations. It points to In re OpenAI, where a magistrate judge in May 2025 ordered OpenAI to preserve output log data that would otherwise be deleted, an order the district judge upheld. It also warns that arguments treating AI interactions as ephemeral are unlikely to succeed, because most platforms save prompt histories.

Privilege is the second reason to get ahead of this. Katten summarizes Heppner, decided February 17, 2026, as holding that a criminal defendant’s prompts to a consumer AI tool were not privileged or work product, because they were not communications with counsel, counsel did not direct them, and the platform disclaimed confidentiality. Prompts written by lawyers have fared better. Katten notes that courts in Tremblay v. OpenAI and Concord Music Group v. Anthropic protected attorney prompts, then found waiver where the parties relied on the outputs in court filings.

The exposure starts inside the legal department. Thomson Reuters’ 2026 Legal Department Operations Report, published October 29, says legal ops functions are viewed as “the first line of defense for proper governance” and names shadow AI, meaning AI use that is not visible to or tracked by the organization, as a growing concern. In its survey data, 16% of corporate legal respondents said they use unauthorized AI tools for work occasionally and 11% frequently.

Building AI Into Holds and Retention

Start with an inventory. List the enterprise AI tools the company licenses, the consumer tools people are known to use and who administers each one. Katten’s companion piece in The National Law Review puts that inventory first among its recommended steps.

Then update the hold notice template. Katten recommends that hold notices expressly cover generative AI tools, both enterprise and consumer, including prompts, outputs and related account settings. Name the actual tools in plain words that custodians recognize, and tell them not to delete chat histories on covered topics.

Next, confirm the technical side with IT. For Microsoft 365 Copilot, Microsoft’s Purview retention documentation says prompts and responses are stored in a hidden folder in the user’s mailbox, can be searched with eDiscovery tools, and are not permanently deleted while a Litigation Hold or eDiscovery hold applies to that mailbox. The same page says new retention policies can also cover ChatGPT Enterprise and, where a collection policy captures content, consumer tools such as ChatGPT and Google Gemini. It adds a warning worth repeating to every custodian: “Messages visible in your AI apps are not an accurate reflection of whether they are retained or permanently deleted for compliance requirements.” For tools outside Microsoft, ask each vendor how long histories persist, whether admins can export them and whether a hold can stop deletion.

Finally, connect the hold to policy. Katten advises routing litigation-related AI use through counsel and approved enterprise tools and keeping confidential or privileged information out of open consumer tools. Those rules belong in your AI use policy, and the hold process is where you learn whether people follow them.

Common Questions

Are AI chats discoverable?

They can be. The Miami order says using AI does not by itself make a communication discoverable or protected. Ordinary relevance, proportionality and privilege rules decide.

When should AI prompts be preserved?

Katten advises preserving relevant prompts, outputs and account settings once a duty to preserve attaches, and naming AI tools in hold notices. Counsel decides the scope for each matter.

Can a deleted Copilot chat still be found?

Often for a time. Microsoft says Copilot messages a user removes stay searchable in a hidden mailbox folder until permanent deletion, and a hold suspends that deletion. Other platforms differ, so ask each vendor.

Are prompts written by lawyers privileged?

Some courts have protected attorney-drafted prompts as work product, according to Katten, but protection was waived where parties relied on the outputs in filings.

What to Watch

Watch whether the AI providers in the Botach case object to the narrowed subpoenas, and whether more courts follow Heppner on consumer tools. Teams that add AI tools to the hold template and test retention settings now will have an answer ready when the first preservation letter names a chatbot.