AI Data Collection for Legal Teams: One Workflow for Claude, ChatGPT, and Gemini

Enterprise AI is no longer a one-vendor market. TechCrunch's coverage of Menlo Ventures' 2025 mid-year report found that Anthropic held 32% of enterprise large language model usage and OpenAI 25%, with Google gaining ground. For legal teams, that mix means a single custodian may have relevant conversations in Claude, ChatGPT, and Gemini at the same time.

Courts are also treating those conversations as discoverable. In the consolidated copyright litigation against OpenAI, a federal magistrate judge ordered production of 20 million anonymized ChatGPT output logs, and Anderson Kill's analysis of the ruling notes that AI output logs are discoverable when they are relevant to a disputed issue. Collecting from three assistants through three separate processes is slow, inconsistent, and hard to defend.

Three Assistants Create Three Different Collection Problems

Each major AI assistant exposes enterprise data through a different mechanism, with different rules about what is available and for how long:

• Access paths differ. Claude and ChatGPT data is available through each vendor's compliance API, while Gemini data is exported through Google Vault.

• Retention differs. Some content is available only within a vendor's retention window, while other content depends on the organization's own retention policies.

• Content coverage differs. Some sources include uploaded and generated files with each conversation, while others export text only.

• Plan requirements differ. Compliance access is typically limited to enterprise tiers, so personal or lower-tier accounts may fall outside what can be collected.

When each assistant is handled by a separate team or tool, these differences produce inconsistent scope, metadata, and documentation across a single matter. Generative AI eDiscovery becomes much more manageable when one workflow absorbs these differences instead of passing them on to the legal team.

What Legal Teams Can Collect From Each Assistant

Claude Data Comes Through Anthropic's Compliance API

Claude Enterprise data is accessed through Anthropic's Compliance API. Collections can include full conversation threads, prompts and responses, user-uploaded and Claude-generated files, shared artifacts, and metadata such as the model version used for each response. The Onna Claude connector collects this data by custodian and keeps files linked to the conversations they came from.

ChatGPT Data Depends on the OpenAI Compliance API

Teams that need to collect ChatGPT conversations work through the OpenAI Compliance API, which is available for ChatGPT Enterprise and Edu workspaces. Available data includes persistent and temporary chats, prompt edit history, attachments, and model information. Ephemeral and deleted conversations are available only within OpenAI's 30-day retention window, so timing matters. The Onna ChatGPT connector supports collection across multiple Enterprise workspaces.

Gemini Data Is Exported Through Google Vault

Gemini conversations in Google Workspace are exported through Google Vault, which means only content retained under Vault policies or holds is available. Exports cover text-based prompts and responses, along with conversation and message metadata. The Onna Google Gemini connector manages these Vault exports by custodian, and this practical guide to eDiscovery for Google Gemini covers the setup and retention questions in more depth.

One Workflow Covers All Three Assistants

The access paths differ, but the legal workflow does not need to. A consistent approach to AI data collections follows the same five steps for every assistant.

1. Confirm Access Before a Matter Begins

Verify that each assistant is on a plan with compliance access, that the right APIs or Vault settings are enabled, and that an administrator can authorize a connection. Discovering a missing setting after a hold is issued costs time that short retention windows do not allow.

2. Preserve Early

Retention windows for some AI content are short, and users may be able to delete conversations themselves. Starting recurring collection for in-scope custodians as soon as a matter is anticipated captures content before it ages out. This guide on how to preserve AI-generated content in collaboration platforms covers the outputs that land outside the assistant itself.

3. Scope by Custodian, Not by Tool

Define the custodians and date ranges once, then apply that scope to every assistant they use. This keeps a single custodian's AI activity consistent across sources.

4. Normalize Into One Format

Collected conversations should share a common structure: threads as reviewable units, prompts and responses clearly distinguished, and metadata such as timestamps, model version, and custodian identity in consistent fields.

5. Search, Cull, and Export Together

With all three sources in one index, teams can search across assistants at once, remove routine interactions, and export only what is relevant to review.

An AI-Generated Content Collection Platform Should Hide the Differences

The value of an AI-generated content collection platform is that legal teams work with one process while the platform handles each vendor's access method. Useful capabilities include:

• Prebuilt connectors for each assistant, authorized once by an administrator

• Custodian-based collection with one-time, recurring, and archive sync options

• Files and artifacts kept linked to the conversations they belong to

• Audit logs for every collection, across every source

Onna provides this through its Claude, ChatGPT, and Gemini connectors, alongside connectors for the collaboration apps where AI outputs often end up. Teams can request a demo to see all three assistants collected through one workflow.

Consistency Across Assistants Is What Makes AI Collections Defensible

Employees will keep choosing different AI assistants for different tasks. Legal teams do not need to match that variety with a separate process for each tool. When Claude, ChatGPT, and Gemini data is preserved early, scoped by custodian, and normalized into one format, every collection can be explained the same way when opposing counsel or a regulator asks how AI data was handled.

To set up one collection workflow for Claude, ChatGPT, and Gemini before your next matter begins, contact the Onna team.