What an Enterprise AI Assistant Leaves Behind: Prompts, Outputs, and Uploaded Files as Evidence

Courts have started treating AI conversations as ordinary discovery material. In the consolidated copyright litigation against OpenAI, a federal magistrate judge ordered the company to produce 20 million anonymized ChatGPT output logs, rejecting objections based on privacy, burden, and proportionality. As Anderson Kill's analysis of the OpenAI discovery ruling notes, the order signals that AI output logs are discoverable when they are relevant to a disputed issue.

That case involved an AI provider, but the same reasoning applies to the organizations using these tools. McKinsey's State of AI in 2025 survey found that 62% of organizations are at least experimenting with AI agents, on top of broad use of AI assistants across business functions. Every one of those interactions creates records that may later be requested in litigation, an investigation, or a regulatory inquiry.

An AI Assistant Leaves Three Kinds of Records

Each interaction with an enterprise AI assistant produces several distinct records. Each one can carry different evidentiary weight.

Prompts Show Intent and Knowledge

A prompt records what a user asked, when, and in what words. Prompts can show what an employee knew at a given time, what they were trying to accomplish, and what information they shared with the tool. In an employment dispute or internal investigation, the question someone asked can matter as much as the answer they received.

Outputs Show What the AI Produced and What Was Used

AI-generated responses, summaries, and drafts often move into email, documents, and chat messages, sometimes with light edits and sometimes without. Determining what counts as evidence when AI wrote it requires knowing which output a person relied on and how it changed before it was shared. Teams also need methods for identifying AI-generated content in data once it has been copied into other systems.

Uploaded Files Show What Data Left Its Usual Controls

When users attach contracts, spreadsheets, or customer records to a prompt, a copy of that file may be stored with the conversation. Uploaded files can be relevant to confidentiality claims, data protection obligations, and trade secret disputes, because they show what information was shared with the AI tool and when.

AI Records Are Harder to Collect Than Email

Email sits in a known system with long retention and mature collection tools. AI assistant data does not follow those patterns:

• Records are split across locations. The conversation may live in the assistant, the output in a shared document, and the uploaded file in separate storage.

• Retention is short or user-controlled. Chat histories may be deleted automatically after a set period, and users can often delete conversations themselves.

• Context is easy to lose. An output collected without its prompt, or a prompt without the files attached to it, can be misleading or hard to authenticate.

• Ownership is unclear. AI tools are often adopted by individual teams before IT or legal knows they are in use.

These issues reflect broader governance gaps in enterprise AI, where adoption has moved faster than the policies and controls that make AI data collections possible.

Four Steps Make AI Data Collections Defensible

1. Inventory AI Tools and Their Settings

List every sanctioned assistant and AI feature in use, who administers it, how long it keeps prompts and outputs, and whether users can delete history. This record supports defensibility if preservation decisions are questioned later.

2. Include AI Sources in Every Hold

Hold notices and preservation scopes should name AI tools alongside mailboxes and file shares. As this guide to legal holds for AI-generated content explains, notices alone cannot stop automated deletion, so technical preservation needs to apply to AI data as well.

3. Collect Records Together, With Metadata

Prompts, outputs, and uploaded files should be collected as connected records, with timestamps, user identifiers, and conversation IDs intact. Collecting them separately makes it harder to show what was asked, what was returned, and what data was involved.

4. Filter Before Review

AI assistants generate high volumes of routine interactions. Searching and filtering by custodian, date range, and keyword before export keeps review focused on the conversations that matter to the case.

A Collaboration Data Platform Treats AI Like Any Other Source

AI assistant data is easiest to manage when it is handled through the same process as chat, email, and file storage. A collaboration data platform connects directly to these sources, preserves content with its metadata, and keeps a single audit trail across every collection. This avoids one-off exports for each new AI tool and keeps prompts, outputs, and files in their original context.

Onna's prebuilt connectors cover collaboration apps and enterprise AI assistants, so legal teams can search, preserve, and collect AI data alongside the rest of a custodian's records without routing each request through IT. Teams that want to see how this works with their own AI tools can request a demo.

AI Evidence Is Defensible Only When Its Context Survives

A prompt, an output, and an uploaded file each tell part of a story. Collected together, with their metadata, they show what a person asked, what the AI returned, and what data was involved. Collected separately or too late, they can be incomplete or impossible to authenticate. Organizations that bring AI assistants into their standard preservation and collection process are better prepared when enterprise AI evidence is requested.

To build an approach that collects AI prompts, outputs, and uploaded files with their full context intact, contact the Onna team.