What Wider Claude Compliance API Coverage Changes for Enterprise Records

On October 8, 2026, Anthropic’s Claude Platform release notes recorded two changes to the Compliance API: the chat endpoints “now also return chats from the unified Claude experience, in beta for Claude Enterprise organizations, with your existing Compliance Access Key,” and the API “can now download Claude Docs documents as Word files, in beta.”

For a compliance, legal, or security team that already treats the Compliance API as a system of record for eDiscovery and data loss prevention, the operative word in both entries is “beta,” and the operative question is whether either change alters a retention or discovery decision you have already made. The short answer from the documentation is that the coverage widens while the retention rules stay put.

Start with the chat change. The documentation states that chats in the unified Claude experience “are returned by these endpoints like any other chat: a chat that continues in a session in the cloud comes back as one chat.” The work performed in the cloud portion appears in each message’s content as tool_use and tool_result blocks. Deletion is documented as cascading: deleting such a chat “also deletes the sessions in the cloud that were started for the chat,” though not sessions those sessions started. If your export pipeline is keyed on chat id and walks the list endpoint, this reads as additional rows rather than a new schema.

That is the reassuring half, and it is also where the caution begins. The documentation never defines what the “unified Claude experience” encompasses. It describes how those chats come back, not which surfaces fall inside the term. It also scopes the content endpoints to Claude Enterprise organizations throughout and says nothing about consumer or personal accounts either way. Neither gap is a conclusion in itself: absence of a statement is not a statement of exclusion, and an article should not manufacture one. Practically, that means the beta coverage should be labelled as beta in any export manifest, because an unlabelled export implies a completeness you cannot yet assert from the documentation alone.

The Claude Docs change is narrower than its headline. The documented path is a ZIP archive “with one Word (.docx) file for each tab that holds text.” Comments, uploaded files, and edit history are excluded. A Word file produced this way is, by the documentation’s own terms, a capture of tab text rather than a replica of the authored document. Teams that have assumed a downloaded Word file is a complete rendition of what an employee wrote should correct that assumption before it reaches a discovery response. The same limitation already applies on the chat side, where documents attached to chats may be stored as the text claude.ai extracted, with the original document unavailable through the API.

On retention, nothing moved. The published figures are unchanged by this release: Activity Feed records are retained six years, chat, file, and project content follows your organization’s claude.ai retention policy unless a user deletes it sooner, and remote session transcripts are retained six years unless a user deletes them. Two independent clocks matter here, because the Activity Feed retains events for six years while it records who did what and when, not prompt or response text. Message bodies require the content endpoints and a Compliance Access Key carrying read:compliance_user_data.

The decision that follows is about what you archive rather than what you can query. Three documented behaviours push toward exporting rather than relying on retrieval. A user deleting a chat in claude.ai removes its content; the chat still lists with a populated deleted_at and an empty name, and its messages return without content. A hard-delete through the API is immediate and permanent, with no recovery window, so a workflow that might delete should retrieve and archive first. And content your retention policy has already removed cannot be returned, so an eDiscovery horizon longer than your retention window requires a parallel archive.

Two smaller planning items are worth folding into the same conversation. The API’s own guidance for keeping an export current is the organization-wide ordered walk, because one paginated loop picks up new chats, chats with new messages, and chats deleted in claude.ai without enumerating users first, with results processed idempotently and keyed by chat id. The docs also note that exports are not verified by a total count: the list endpoints return no total_count and no checksum, and the recommended attestation is to log the starting cursor, the terminal id, the record count, the run timestamp, and the final page’s request-id. If your legal hold or audit horizon exceeds the documented retention periods, that record is what lets you defend the export later.

What this release does not do is settle the boundaries. It does not define the unified experience, does not state whether consumer accounts are covered, offers no output format beyond the Word-per-tab ZIP, and names no general availability date for either beta. Treat the two entries as a coverage expansion worth re-testing against your existing collection scope, not as a retention change or a completeness guarantee.

One constraint the export section does not surface is that the two shapes are not interchangeable. The docs state that user-filtered queries always sort by created_at, that passing order_by=updated_at alongside user_ids returns a 400, and that filtering by project_ids is only available in that user-filtered form – while the same page lists the project_ids filter among the things not available on the organization-wide queries it recommends for keeping an export current. The docs also say combining user_ids with any updated_at.* bound is deprecated and would be rejected with a 400 after 2026-09-22, a date that has now passed, so a team that scoped its pulls to named custodians has to move to the org-wide walk and select those custodians out of its results instead. That is a forfeiture of the project dimension rather than a pure upgrade, and I would want to know how anyone plans to reconstruct project membership from the walk’s output.

The retention discussion treats the archive decision as driven by user deletion and hard-delete, but the local session clock has a third trigger the post does not mention. The docs put local session transcripts at six years by default, at your organization’s custom conversation retention period when a finite one is set, and at 30 days in organizations with HIPAA readiness enabled, and they say the local session endpoints stop returning messages older than the organization’s current period as soon as the setting changes. Lengthening the period afterwards does not restore transcripts that have already expired. So a retention-policy edit is destructive in one direction even though nothing was deleted, which makes a stored copy the hedge against a configuration change as well as against an erasure.

The archive strategy has a throughput number attached to it that I would want in the plan before any backfill starts. All /v1/compliance/* endpoints share one budget of 600 requests per minute per parent organization, and the errors page is explicit that this is a single budget shared across every key under the parent – Compliance Access Keys and the Admin API keys of all linked organizations alike – and across every endpoint, with the remote session endpoints carrying a second budget on top. That means a one-time historical export, the recurring org-wide walk, and any custodian-scoped pull for a hold are drawing on the same 600 rather than three separate pools, and the docs say to contact your Anthropic representative if the integration needs a higher limit. Two details belong in the runbook: a valid key that lacks an endpoint’s scope consumes a quota unit before the 403 comes back, and the 429 rule is to wait the retry-after seconds and not advance the cursor – which is the very receipt you would otherwise log for the attestation the post recommends.

Something the post’s attestation advice leaves out is that the export is itself an audited event. The docs say calls to the Compliance API emit compliance_api_accessed activities, and that you scope them with activity_types=compliance_api_accessed and read actor.api_key_id from the activities whose actor.type is api_actor to attribute a query to a specific Compliance Access Key or Admin API key. That is a useful second record beside the cursor-and-count log, because it shows who ran the pull and not merely that it completed. The same docs say recording begins when the Compliance API is first enabled for the organization and is not backfilled, so a query trail has a start date and cannot be extended backwards – worth knowing if the collection window you are asked about reaches back before enablement.