Part of: AI Activity Retention
A ChatGPT session that produces a client recommendation, drafts a piece of marketing copy that will go out under the firm’s name, or summarizes regulated communications for a compliance review is producing business records under existing rules. It’s the content that determines the obligation, not the interface.
This page covers what current compliance vendor capture actually preserves of a ChatGPT session, what it misses, and what a firm should expect to produce if an examiner asks.
When ChatGPT activity is likely a record
The existing recordkeeping framework is channel-agnostic. The same logic that applies to email, instant messaging, and mobile messaging plausibly applies to AI tool activity. If the content of the communication or activity relates to:
- Client business, accounts, or recommendations
- Trades, orders, or transactions
- Marketing or solicitation material
- Compliance review of business communications
- Firm operations and personnel decisions in a regulated context
- Any other category of activity covered by the firm’s recordkeeping obligations
The activity may be a record under existing rules. ChatGPT being the interface does not change the underlying obligation. Whether the firm has captured the activity, and in what fidelity, determines whether the firm can produce the record on examiner request.
This is the same logic the SEC and FINRA pursued during the 2021-2024 off-channel sweep. The activity was happening in WhatsApp, Signal, and iMessage. The firms had not captured it. The penalties followed. Whether AI activity follows the same enforcement pattern is not yet established, but the regulatory framework that would support that path is in place.
What ChatGPT actually produces
A modern ChatGPT session is not a simple prompt-response transcript. Depending on configuration, it may include:
- User prompts and responses. The text the employee typed and the text the model produced.
- File uploads. Documents, spreadsheets, and other files attached to the conversation. These often contain client or firm data.
- Knowledge file retrievals. If a custom GPT is configured with knowledge files, the GPT retrieves content from those files during the conversation. The retrievals shape what the model says.
- Tool calls and GPT Actions. A configured GPT can call external HTTP APIs — GPT Actions are integrations that let a custom ChatGPT call external systems on the user’s behalf. These are the AI taking real action against real systems.
- Code interpreter execution. ChatGPT can run Python code in a sandboxed environment. The code that ran and the data it produced are part of the session.
- Web browsing. When enabled, ChatGPT can fetch web pages. The URLs visited and content retrieved are part of the session.
- Memory. Personalized context the model has stored about the user across sessions.
A captured chat transcript covers the first item. Some of the second item (file uploads) is captured by the OpenAI Compliance API. The rest sits in a layer that current vendor capture generally does not address.
What current vendor capture covers
As of public materials reviewed on 2026-05-15, ChatGPT Enterprise vendor capture sits on the OpenAI Compliance Platform. OpenAI lists Global Relay as currently supporting updated conversation logs, with Smarsh and Microsoft Purview listed as in progress for updated conversation logs while supporting other data types. Smarsh publicly documents its ChatGPT Enterprise integration. Global Relay has announced its integration. Microsoft Copilot capture is publicly documented by Smarsh and Theta Lake; we did not find equivalent ChatGPT Enterprise capture documentation from Theta Lake as of this writing.
This is real progress. The chat-transcript layer is increasingly supported.
OpenAI’s own description of the Compliance Platform includes conversations, uploaded files, memories, users, and “workspace GPT configuration and metadata.” That is broader than a simple chat export. The practical question for a regulated firm is what each archive vendor’s integration exposes in examiner-ready form. Coverage of GPT configuration, knowledge file retrievals, GPT Action arguments and results, code interpreter activity, and memory state is not consistently documented across vendor public materials as of 2026-05-15. If a specific capability matters for a firm’s compliance posture, request current documentation and dated screenshots from the vendor.
Two coverage gaps are widely understood today:
- Tool calls and GPT Actions inside the session. A custom GPT can invoke external HTTP APIs (Actions) on the user’s behalf. Whether the resulting arguments and return values surface in an examiner-ready export through the Compliance Platform is not publicly documented by archive vendors in a way that lets a firm rely on it without verification.
- ChatGPT Team and consumer-tier ChatGPT. Compliance Platform integration is an Enterprise-tier feature. Activity on Team and consumer tiers is not captured through the same path. The activity still happens. The recordkeeping consideration still applies.
What regulators have actually said
The SEC’s 2026 examination priorities flag AI governance and AI-assisted activity as a focus area. Examiners are directed to assess whether firms have policies governing employee AI use, whether the policies are enforced, and whether records of AI-assisted activity are being retained.
FINRA’s 2026 Annual Regulatory Oversight Report and Notice 24-09 signal that existing recordkeeping rules apply to AI-generated communications in the same manner as any other business communication.
Regulators have not issued explicit guidance that a ChatGPT prompt is a record, that a GPT Action call is a record, or that a custom GPT’s system prompt must be preserved. The framework that would implicate this activity (Rule 17a-4, Rule 4511, Rule 204-2) already exists. Explicit AI-specific rulemaking does not. A defensible compliance posture, developed in consultation with the firm’s counsel, often preserves activity the existing framework appears to cover rather than waiting for new rulemaking.
What an examiner is likely to ask for
Examination requests typically arrive in plain language and ask for activity, not artifacts.
“Provide all communications and records relating to [matter] between [start date] and [end date], including any AI-assisted output.”
For a ChatGPT session, a defensible answer assembles:
- The chat transcript (captured by current vendor integrations)
- The files attached to the session
- The system prompt of any custom GPT used
- Knowledge file retrievals the GPT made during the conversation
- GPT Action calls with arguments and results
- The model identity and version
- Tool surface (which Actions and capabilities the GPT had available)
- Policy decisions and any lifecycle hooks that fired
If the firm has only the transcript, it can produce a description of what the AI said. If the firm has the full set, it can produce evidence of what the AI did.
Where the gap is widest
Three scenarios are worth flagging because current vendor capture has the least public coverage of them.
Custom GPTs with Actions, used at scale. A firm builds a custom GPT to assist with research, configures it with a detailed system prompt, attaches knowledge files containing internal data, and gives it Actions that call internal systems. An employee uses this custom GPT throughout the workday. The user prompts and the model responses are captured. Whether the GPT configuration, the knowledge retrievals, and the Action arguments and results land in an examiner-ready export depends on the specific vendor integration and is not consistently documented across vendor public materials as of 2026-05-15. A firm relying on these for regulated activity should verify with its vendor before assuming coverage.
ChatGPT Team and consumer-tier ChatGPT. Employees who use ChatGPT outside the Enterprise tier are not covered by the Compliance Platform. This is the same pattern as off-channel messaging when employees use messaging outside of the firm’s approved apps. The activity still happens. The recordkeeping consideration still applies. The vendor capture path does not exist for these tiers. A firm needs to either restrict to Enterprise tier or capture activity at a different layer (network, browser, or internal proxy).
Self-hosted AI. When a firm runs self-hosted AI on Ollama, vLLM, LiteLLM, Open WebUI, or similar infrastructure, there is no vendor compliance API to call. The activity still happens. The recordkeeping obligation still applies.
How Arc closes the gap
Arc is designed to capture both halves of an AI session:
- The chat layer, where current vendor integrations are concentrated. Arc integrates with the OpenAI Compliance API and Microsoft Purview / Graph for Enterprise-tier capture (demo available today).
- The execution layer, where current vendors are not. For ChatGPT-style activity, this means capturing tool calls and Action invocations through an integration point or proxy. For internal AI deployments (Ollama, vLLM, LiteLLM, Open WebUI, internal Copilot-style tools), Arc Bridge handles capture at the infrastructure layer. For MCP-driven agent activity, Arc Relay captures every tool call in production today.
The result is a single archive that holds both the chat transcript and the execution record - the same archive that already holds the firm’s WhatsApp, Signal, iMessage, and 40+ other communication channels. One examiner-ready format. One legal-hold workflow. One supervisor review queue.
Related reading
- AI Activity Retention for Regulated Firms The category-level overview. What AI activity retention means, what the two-layer record format covers, and how Arc spans human channels and AI execution.
- Are LLM tool calls business records? What a tool call is, why the chat transcript misses it, and what an examiner is likely to ask for.
- Arc Relay - Open Source MCP Control Plane The first production piece of Arc. Open source on GitHub. Captures every tool call.
- SEC 17a-4 Messaging Compliance Core recordkeeping rule. Why the channel-agnostic framing applies to AI activity.
- FINRA Rule 4511 FINRA's general books-and-records requirement. The same logic applies to AI-generated activity.