Litigation Holds and eDiscovery: When Your AI Chat Logs Are Subpoenaed

What a Hold Notice Does, and Who Has to Receive One

A litigation hold, also called a legal hold, is a written instruction issued under counsel's direction telling specified people inside the organization that they must preserve information relating to an identified matter, and that routine deletion of that information must stop. It is not a filing, it goes to no court, and in most matters the other side never sees it. Its function is practical and evidentiary at once: it actually stops the loss of information, and it demonstrates afterward that the organization took reasonable steps, which is precisely the question Rule 37(e) asks.

A notice that works tends to share a short list of features. It identifies the matter in plain language, states the subject matter and the relevant time period concretely enough that a program coordinator can tell whether a given document is in scope, and lists the categories and systems covered rather than gesturing at relevant documents. It instructs recipients to suspend any personal deletion habits and not to edit, reorganize, or clean up anything in scope, including seemingly harmless tidying. It names a single person to ask questions of, which matters enormously in a small organization where people will otherwise guess. It requires written acknowledgment. And it stays in force until counsel releases it in writing, because an unreleased hold that everyone has quietly stopped observing is nearly as bad as no hold at all.

Who receives it is where organizations most often get the scope wrong, and AI tools make the error easier. The notice goes to the custodians who hold or generate potentially relevant information, which is the obvious cast of characters in the dispute, plus the people who administer the systems where that information lives. In a nonprofit, the second group is routinely overlooked because it is often one part-time IT contractor, a database administrator who is really the operations manager, and whoever happens to be the billing owner on the AI subscription. That last person matters more than their job title suggests: in many small organizations, the only human with the administrative access needed to suspend a retention setting or export a workspace is the person whose credit card set up the account.

Acknowledgment tracking is the unglamorous mechanical half. Counsel needs a record of who received the notice, when, who confirmed, and who was chased. Departing employees need a step in the offboarding process that preserves their accounts and devices rather than wiping and reissuing them, which is the default in most organizations and the source of a great deal of avoidable loss. New staff who inherit a role in scope need the notice too. Periodic reminders keep a hold alive through the many months a matter can take, and in a sector with high turnover and heavy use of part-time and contract staff, reissuing the notice at intervals is not bureaucratic excess. It is how you keep the hold true as the roster changes.

Auto-Deletion Is a Setting Until the Day It Becomes Spoliation

Most AI tools delete things by default, and for ordinary privacy reasons that is often a sensible configuration. Short retention windows limit exposure in a breach, keep sensitive material from accumulating, and reduce the volume of stale data nobody is governing. A nonprofit that set its notetaker to purge transcripts after thirty days, or turned on a ninety-day conversation expiry in its assistant, probably made a defensible privacy decision. The problem is that the same setting, left running after a preservation duty has attached, becomes a mechanism that destroys evidence automatically, at scale, while everyone believes the organization is behaving responsibly.

The federal framework that governs what happens next is Federal Rule of Civil Procedure 37(e), amended in 2015 to replace a patchwork of inconsistent circuit standards. Its opening condition is a three-part test. The rule applies where electronically stored information that should have been preserved in the anticipation or conduct of litigation is lost because a party failed to take reasonable steps to preserve it, and it cannot be restored or replaced through additional discovery. All three elements matter. If the information was never within the preservation duty, the rule is not engaged. If the organization took reasonable steps and lost it anyway, the rule is not engaged. And if a copy exists somewhere else, in a colleague's export, in an attachment, in the output pasted into a document, the loss may be curable without any finding at all.

Where the test is satisfied, the rule splits into two very different tiers. Under subsection (e)(1), on finding prejudice to another party from the loss, a court may order measures no greater than necessary to cure the prejudice. That is a deliberately modest remedy: additional discovery at your expense, permission for the other side to present evidence about the loss, cost shifting, an adjusted schedule. Under subsection (e)(2), only on finding that the party acted with the intent to deprive another party of the information's use in the litigation, the court may presume the lost information was unfavorable, instruct the jury that it may or must presume as much, or dismiss the action or enter default judgment. The severe sanctions are gated on intent, not on carelessness, which is the single most important structural feature of the rule and the reason the Federal Judicial Center's summary of the 2015 amendment is worth ten minutes of a board member's time.

Read that structure carefully, because it tells you what to do and what never to do. An organization that discovers after a trigger that an auto-deletion setting ran for three weeks and purged some conversations is in the territory of reasonable steps, prejudice, and curative measures, which is survivable and frequently survived. An organization that turns a deletion setting on after a trigger, shortens a retention window, bulk-deletes a conversation history, or quietly reconfigures an export has moved toward the intent analysis, and that is a categorically worse place to be. There is also a criminal dimension that nonprofits specifically should know about. Two provisions of the Sarbanes-Oxley Act apply to nonprofit organizations, and one of them concerns the destruction of records in contemplation of a federal investigation or proceeding, which is why the National Council of Nonprofits treats document destruction policy and whistleblower protection as paired governance obligations rather than separate ones.

The operational lesson is simple to state and requires advance work to deliver. The moment counsel says a duty has attached, someone must be able to walk the list of AI systems and suspend deletion in each. That means knowing the list, knowing who has the administrative rights, and knowing whether suspension is even possible in each tool, which for some consumer tiers it is not. The time to discover that a product has no hold capability is during procurement, not during a dispute.

Are Prompts and AI Drafts Business Records?

Staff often assume that a chatbot conversation is more like thinking out loud than like writing a memo, and that assumption shapes behavior in ways that matter later. People type things into a prompt box they would never put in an email: candid assessments of a colleague, speculation about why a grant was pulled, blunt descriptions of a participant, half-formed theories about an incident. The mental model is a scratchpad. The legal model is closer to a document stored on a company system, which is what it is.

Discoverability does not turn on whether your retention schedule calls something a record. It turns on whether the material is nonprivileged, relevant to a claim or defense, within your possession, custody, or control, and proportional to the needs of the case. A prompt in which a manager describes an employee's performance in terms that contradict the written evaluation is relevant in an employment dispute whether or not anyone ever classified chatbot conversations as records. An output the organization relied on in making a decision is relevant to how that decision was made. A thread showing that a draft donor communication was generated and then altered may be relevant to a misrepresentation claim. Courts have shown no particular difficulty treating this category as electronically stored information, and the 2026 litigation over model providers' own conversation logs has made the point in public.

There are harder questions underneath the easy one, and they are genuinely unsettled rather than merely unfamiliar. Whether a retrieval index that encodes your documents in numerical form is itself discoverable, or whether producing the underlying documents suffices, is the kind of issue that will be worked out in ESI protocols and motions over the next several years. Whether a model's intermediate reasoning, where a tool exposes it, is a record of anything the organization did is similarly open. Whether an output nobody ever read or relied on is relevant in the same way as one that drove a decision is a question of fact in each case. These are not questions a nonprofit answers unilaterally, and the appropriate posture is to preserve broadly and let counsel argue scope later, rather than to resolve the ambiguity in your own favor by deleting.

The practical consequence is a training point rather than a technical one, and it should be in every organization's AI onboarding. Anything typed into an organizational AI tool should be written as though it may be read later by someone adverse to you, because it may be. That is not a reason to use AI less. It is the same discipline every competent professional already applies to email, extended to a surface where the informality of the interface has misled people into dropping it. Our article on AI notetakers in nonprofit meetings makes a parallel argument about transcripts, where the gap between what people think is being captured and what is actually being captured verbatim is even wider.

The Vendor Problem: Can They Actually Hold and Export?

Here is the question that decides whether everything above is theory or practice. When your counsel tells you to suspend deletion in your AI tools and produce the relevant conversations, can your vendors actually do it, and does your contract say they must? For a meaningful share of the tools nonprofits use, the honest answer is no, and the answer is only discovered under time pressure, which is the worst possible moment.

The failure modes are specific and recognizable. A consumer tier with no administrative console, where each user controls their own history and the organization controls nothing. A product whose retention is a global account setting rather than a per-matter hold, so the only way to preserve one conversation is to preserve everything, and the only way to stop deletion is a change that affects the whole workspace. An export that produces a format nobody can review usefully, a wall of unsegmented text with no reliable per-message timestamps, no user attribution, and no way to filter by date. A vendor whose support pathway for a legal request is a general contact form answered in a week. A subprocessor relationship where your vendor genuinely does not control the retention of the underlying model logs. And a contract that says nothing at all about preservation, legal holds, export, or cooperation with discovery, which describes most nonprofit AI agreements signed to date.

All of this is cheap to address at contracting time and expensive to address later, which puts it squarely in the procurement conversation rather than the legal one. The questions are not exotic, and a vendor who cannot answer them clearly has told you something useful. Can the organization place a legal hold that suspends deletion for specified users, date ranges, or the whole workspace, and how quickly does it take effect. Who inside the vendor can act on a request, and what is the committed response time. What does an export contain, in what format, with what metadata, and can we see a sample before signing. How is a subpoena served on the vendor for our data handled, and will we be notified before anything is produced. Where do the records sit, including any subprocessors, and whose retention policy governs them. What happens on termination, and is there a period during which we can still retrieve data. Our nonprofit AI vendor evaluation checklist covers the broader diligence frame, and our piece on AI vendor contract management goes into the terms themselves.

One related point is easy to miss. A subpoena may go to your vendor rather than to you, in a matter you are not even a party to, and whether you hear about it before your data is produced depends on what your contract says about notice. That provision costs nothing to negotiate and is almost never in a standard nonprofit software agreement. Ask for it. It is the difference between participating in a decision about your own records and learning about it afterward.

The Work to Do Before You Need It

Everything that makes a hold go badly is a thing that could have been prepared in advance, cheaply, in a quiet month. The preparation is not a project with a steering committee. For a small organization it is a few documents, kept current, that convert a frightening legal event into an administrative sequence.

Start with a data map, which is a plain inventory rather than anything grander. One row per system. What it is, what it holds, who administers it, which account tier, where the data sits, what the retention setting is today, whether deletion can be suspended and by whom, whether it can export and in what format, and the vendor contact for a legal request. Include the AI tools people actually use rather than the ones you approved, which means asking honestly and not punishing the answers. Include free tiers, trials, and tools contractors bring. A two-page version of this document, genuinely current, is worth more than a sophisticated one that was accurate eighteen months ago. Our guide to a nonprofit AI footprint audit walks through how to find what is really in use.

Next, a custodian list keyed to roles rather than names, so it survives turnover: who holds program records, who holds personnel records, who holds financial records, who administers each system, who is the AI subscription owner. Then a draft hold notice template, prepared with counsel while nothing is pending, with blanks for the matter, the subject, the date range, and the systems. Drafting it in advance is worth the small legal fee because it is the thing you will otherwise be writing at speed on a bad afternoon. Alongside it, a simple acknowledgment tracker, which can be a spreadsheet, and a written procedure for the first seventy-two hours: who calls counsel, who suspends deletion in each system, who pauses offboarding and device reissue, who issues and tracks the notice.

Finally, scoping and search preparation. When a matter arrives, counsel will need to search a defined population, and that search is usually built from a keyword list, a date range, a custodian set, and increasingly a set of concepts rather than exact terms, because AI conversations use natural language and the words that matter are often not the words a lawyer would guess. Knowing in advance which systems can be searched, by whom, with what filters, and which require a full export followed by review outside the tool, determines whether scoping takes two days or six weeks. The organizational knowledge that makes this fast, where things live and what they are called internally, is the same asset described in our work on retention schedules for AI records, which is why the two documents should be maintained together and reviewed on the same cycle.

The Lines, Stated Plainly

Three statements belong in your written procedure in roughly these words, because they are the ones people get wrong under pressure.

Once a hold is live, counsel directs the process. Not the executive director, not the IT contractor, not the board chair who has done this before at a previous organization. What the duty covers, how broadly to preserve, what gets searched, what gets produced, what is withheld, and when the hold is lifted are legal determinations. The organization's job is to execute promptly, document what it did, and tell counsel immediately when something cannot be done or has already gone wrong. Internal improvisation in this area, however well intentioned, creates the problems that turn a manageable matter into a sanctions dispute.

Deleting anything after a trigger is the worst available outcome. Worse than producing something embarrassing. Worse than a large and burdensome production. Worse than a privilege fight. Rule 37(e) reserves its severe measures, the adverse inference instruction, dismissal, default judgment, for a finding of intent to deprive, and post-trigger deletion is precisely the conduct that invites that inquiry. This applies to tidying, to reorganizing, to clearing a chat history because it felt cluttered, and to letting a known auto-deletion setting continue running once you have been told to stop it. If something has already been lost, that goes to counsel the same day rather than becoming a quiet hope.

An AI relevance call never replaces attorney review for privilege. Assisted review is a triage instrument that makes human attention go further. It does not make privilege determinations, and treating a model's output as a decision rather than a signal is how organizations waive protections they were entitled to. This is also the clearest illustration of the general principle running through this site's legal and compliance coverage, and set out at more length in our article on AI and litigation risk for nonprofits: AI is excellent at reducing a large pile to a reviewable one, and has no business making the judgment at the end.

A fourth point sits slightly outside those three but matters just as much for small organizations. None of this preparation requires a legal department. A data map, a custodian list, a template notice, an escalation rule, and a seventy-two hour procedure are within reach of a three-person administrative team working with an hour or two of counsel's time. The organizations that come through a hold well are not the ones with the most sophisticated systems. They are the ones where someone, at some point, wrote down where everything lives.