Quick answerOnce litigation is reasonably anticipated, whether or not a lawsuit has actually been filed, a company generally must suspend routine deletion of AI agent conversation logs that could be relevant, and preserve them in a form that is retrievable and unaltered. This duty exists independent of any specific subpoena or law enforcement request, and it can apply broadly across all customers, not just the ones directly involved in the anticipated dispute, if the logs are relevant to the underlying issue. Treat the AI agent's conversation store the same way you would treat email or any other business record once a legal hold is issued.
This obligation exists before anyone asks for the logs
A subpoena or a law enforcement request, the scenario covered in Handling a Legal Subpoena for AI Agent Conversation Logs, is a specific external party asking you to produce specific records. A litigation hold obligation is different and comes earlier: once your company reasonably anticipates litigation, whether that is a demand letter, a serious internal incident, or credible knowledge that a dispute is likely, you generally have a duty to preserve relevant records even before anyone has formally requested them, and often before a lawsuit has even been filed.
For AI agent conversation logs specifically, this duty is easy to miss because conversation data often lives in a system that was built for operational purposes, like debugging, quality review, or model improvement, not for legal preservation. If your retention policy quietly rolls off conversation data after some number of days for storage cost reasons, that policy needs an override mechanism the moment a hold is triggered, or you risk destroying records a court would have expected you to keep.
What "reasonably anticipated" means in practice
You do not need an actual complaint filed against you for the duty to attach. Courts generally look at whether a reasonable person in your position would have concluded that litigation was likely, which can be triggered by things like a serious customer complaint referencing legal action, an internal incident significant enough that legal counsel gets involved, or a regulator inquiry that could turn into enforcement. This is a lower bar than many teams assume, and it means the trigger for a hold is often a business event, not a legal filing, so legal and the team operating the AI agent both need a clear, fast channel to flag it.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
What preservation actually requires for a conversation log system
Suspend routine deletion for the relevant scope, not necessarily your entire conversation history. Work with legal counsel to define the scope precisely, since preserving everything indefinitely is expensive and can itself become a liability, while preserving too narrowly risks missing something later found to be relevant.
Preserve the log in a form that shows what the customer and the agent actually exchanged, including any tool calls or retrieved context that shaped the answer, not just the final displayed message. If your agent's behavior depends on retrieval or tool outputs, and those are not part of what gets preserved, you may end up unable to reconstruct why the agent said what it said, which is precisely what a dispute might turn on.
Preserve metadata alongside the content: timestamps, the model or prompt version in effect at the time, and any relevant configuration state. This connects directly to the discipline described in Versioning and Rolling Back AI Agent Prompts: if you cannot tell which version of the agent produced a given conversation, you cannot reliably explain its behavior later, in litigation or otherwise.
Document that the hold was issued and followed, including who was notified and when. A hold that exists only as an informal request in a chat message is much harder to demonstrate compliance with than one that is logged and acknowledged.
Where this differs from your routine data retention policy
A standing data retention policy, the kind that governs how long you keep conversation logs by default, is designed around operational and privacy tradeoffs: keep enough to support the business, delete enough to limit exposure and cost. A litigation hold does not replace that policy, it overrides it temporarily and narrowly for the specific scope in question. The two systems need to interact cleanly: your deletion jobs, whether automated or manual, need a way to check for an active hold before they run, and that check needs to happen at the record level for the relevant customers or time window, not just as a company-wide switch someone has to remember to flip.
This also intersects with any commitments you have made to customers about deleting their data, such as the process described in What Happens to an AI Agent's Conversation History When a Customer Asks to Be Forgotten. A litigation hold can legitimately override a pending deletion request for the specific records in scope, but that exception needs to be handled carefully and disclosed appropriately, since silently ignoring a deletion request without explanation creates its own compliance problem.
FAQ
Does a litigation hold apply only to the specific customer involved in the dispute? Not necessarily. If the underlying issue is systemic, for example a claim that the agent behaves incorrectly for a class of customers, the hold may reasonably extend to all conversations that could be relevant to demonstrating the pattern, not just the individual claimant's conversations. Scope should be set with legal counsel based on the nature of the anticipated dispute.
Who is responsible for making sure the hold actually stops automated deletion? Whoever owns the data retention and deletion pipeline for the conversation log system needs to be a named party in the hold process, since a hold issued only to legal or leadership without reaching engineering will not actually stop an automated job from running on schedule.
What happens if relevant logs were already deleted before the hold was issued? This is generally treated differently from a failure to preserve after the duty attached, but it still needs to be documented honestly and disclosed if it becomes relevant, rather than discovered later. Work with counsel immediately if this happens rather than treating it as a routine gap to quietly patch.