Loading
In this lessonBy the end, you’ll be able to
  • Explain what the Compliance API records, what it doesn’t, and how it differs from audit logs and full data exports
  • Decide whether to route OpenTelemetry export to your observability stack as an operational complement to the Compliance API
  • Choose a retention window matched to your existing data policy and propose inference hooks for content that shouldn’t reach Claude at all

This is the fifth of the five decisions, Visibility. You’ve set what your rollout costs. This lesson covers visibility: the three pieces (the Compliance API, OpenTelemetry, and retention) and the one decision among them that’s hard to walk back.

What visibility is for

The first four decisions you made configure what Claude does. Visibility is how you determine whether they’re working. Visibility also lets you answer specific after-the-fact questions with evidence: Which members touched a particular account? What did Claude produce in a sensitive conversation? How long is that conversation kept? These answers can only come from records you were already keeping.

That’s why you set visibility up before members get access, not after: the stretch between first access and the day you turn recording on leaves gaps in the record that can’t be filled in later. Three pieces give you the coverage: the Compliance API (conversation content, plus audit log events), OpenTelemetry (how the surfaces are running), and retention (how long content is kept). Some of that record is captured only while the Compliance API is on: activity feed events, and transcripts of sessions on members’ machines, such as Cowork and Claude Code.

The Compliance API: your record of conversation content

The Compliance API is the content-carrying record. It covers Claude chat and extends to Cowork and Claude Code sessions on your organization’s existing Compliance Access Key and settings, so there is nothing extra to integrate. Check which plans it’s available on in the help center’s Compliance API article, and which apps’ sessions it records in Retrieve session transcripts on Claude Platform Docs. It carries activity feed events, chat data, and file content, so it can answer “what was said” and “what was shared,” not just “who logged in when.” Whether it meets a specific legal or regulatory requirement is your legal and compliance owners’ call. Integrate it with the CASB, DLP, eDiscovery, and SIEM tools you already run — many of those vendors have built integrations on the Compliance API, listed in Get started with Claude Compliance API integrations.

The Compliance API is easy to confuse with two other records:

  • Audit logs are metadata-only. They record events (who did what when: admin and member actions) with no prompts and no completions. The Compliance API includes audit log events too, so one record carries both, but the distinction still matters: when someone asks for content, metadata alone is the wrong answer.
  • Full data exports are a separate Primary Owner capability, for pulling your organization’s data wholesale rather than reviewing it case by case.

The Primary Owner enables the Compliance API. Access keys are scoped when they’re created: a Primary Owner key can cover every organization under the parent, while an Owner key covers that Owner’s organization only. A Primary Owner can also create a key scoped to a single organization (a child of the parent), not only a parent-wide one — scope is a per-key choice at creation, not fixed by role. The decision that matters after that is who reads it: a record no one reviews isn’t actually a control. Many companies already have a place where security logs go (a SIEM or your security tooling’s central collection point) with an owner who reviews them. Route the Compliance API there, and Claude becomes one more source in a review process that already exists.

OpenTelemetry: how the surfaces are running

OpenTelemetry export sends operational telemetry to your observability stack: usage patterns and performance, and on some surfaces conversation content. Claude Code has a documented opt-in for logging prompts if your engineering team turns it on; by default its telemetry is operational data. In current versions of Claude Desktop, Cowork’s export carries only metadata: prompt text, Claude’s responses, and tool inputs are off by default, and the otlpContentCapture device policy turns them on. Some older versions send that content unless the policy is set to metadata only (see Cowork monitoring for how), and the metadata always includes members’ email addresses and the paths of folders they connect, so filter those in your collector if your organization requires it. Treat OpenTelemetry as “how it’s running,” complementary to the Compliance API’s “what was said,” and optional: the Compliance API, not telemetry, is the record to rely on for content. If your platform team already runs an observability stack, this is a routing task for them.

Retention: how long content is kept

Retention is how long conversation content is kept before deletion. Left at the default, content is kept indefinitely until you or your members delete it. If you set a custom retention window instead, content is deleted automatically on your schedule, and for a chat kept inside a project, the project’s window applies instead of the chat’s. For when the retention clock starts and what a custom window doesn’t cover, see custom data retention and, if your organization uses the unified Claude experience, its setup article. The shortest custom window is 30 days. This setting has no zero-retention option.

Activity feed events and session transcripts reached through the Compliance API follow their own retention model:

  • Activity feed: six years regardless of the window you set.
  • Remote session transcripts (Cowork sessions started on claude.ai on the web or mobile, which run in the cloud in Anthropic-managed environments, and each run of a Cowork scheduled task in the cloud): no default retention period. A session Claude starts from a chat is deleted with that chat, including when the window you set deletes it.
  • Local session transcripts from Cowork and Claude Code: your custom window when you set a finite one, or six years when you don’t; in organizations with HIPAA readiness enabled, 30 days from capture, even if your custom window is longer.

Chat, file, and project content follows the window you set, so a short window limits what members keep and what the Compliance API can return.

Retention is an organization-wide setting, so choose it against the data-governance policy your company already has for this class of content, and involve whoever owns that policy, so the number is approved before you set it. If a regulated function needs a shorter window than the rest of your organization would like, the shorter window is what everyone gets, so name that trade-off to the affected teams so they’re aware.

One proposal to bring to that conversation is inference hooks: every prompt from Claude chat, Cowork, and Claude Code goes to a server your security team or their vendor operates for an allow-or-deny decision before Claude answers, so content that shouldn’t reach Claude at all is blocked at the door rather than governed by retention after the fact. The help center’s Inference hooks overview covers what they are and who can turn them on.

The interactive widget below shows how each of these decisions affects your ability to answer questions about your data after the fact.

Loading

Pluto’s visibility

Pluto’s Primary Owner enabled the Compliance API before general rollout and routed it into the security review Pluto already runs, so a named reviewer reads it as part of an existing process. Cowork and Claude Code sessions fall under the Compliance API’s coverage, and their operational telemetry exports to the observability stack that Platform already runs. And because Payments & Trust required a shorter window than the other units would have chosen, the risk owner set the organization-wide retention to that shorter number. The reviewer pulls whatever Pluto must keep past that window into Pluto’s own archive before it’s deleted, because the Compliance API can’t return chats, files, or local session transcripts that retention has removed.

1 · Your decision

What retention window do you set, and is the Compliance API on and routed to a reviewer?

2 · The choices you can make

The choiceWhen you’d choose itWhat it means and the impact it has
Match your existing data policy, Compliance API onThe default. You have a data-governance policy for this class of content and a security/compliance review process a named reviewer already runs.Retention set to whatever your data-governance policy already says for this class of content. Compliance API on and routed to the review you already run.
Regulated posture, Compliance API expectedYou handle regulated data.Compliance API moves from optional to expected. The retention window is typically set organization-wide against your regulatory and contractual obligations rather than your general data policy — that call sits with your risk owner. The whole organization gets the regulated group’s number.
Compliance API off, condition documentedNo review process exists and none is planned.Leave the Compliance API off for now and write down the condition that turns it on. Note that until it’s on, questions about that period have no activity feed events or local session transcripts to draw on.

If part of your company’s data must be fully separated from the rest, that isn’t a retention setting: it’s the separate-organization question from Lesson 4, so route it back there rather than trying to solve it here.

3 · If you change this later

The Compliance API and OpenTelemetry are toggles and destinations you can change any time, though any stretch with the Compliance API off leaves the same gaps as turning it on late. Retention is different in kind: the setting moves freely, but the content it deletes doesn’t come back. Shortening the window later erases history your members had built up under the old one. So agree on the window once, before members accumulate work under a number you’ll want to change.

Set up resources

When you’ve made the call, retention and the Compliance API are set in Organization settings, and the Primary Owner turns the Compliance API on.

Lesson activity

This is one of the two decisions that usually escalates, so the companion section doubles as the brief you hand the visibility owner from Lesson 2. The risk lead makes the call for this one. Assemble it in four parts:

  1. The three pieces (conversation content, operational telemetry, retention), the process at your company it plugs into, and the person who owns that process. Any piece with no existing process is one where you’d be building a review process from scratch. Flag it before rollout, because those take the longest to stand up.
  2. Your recommended retention window and the existing policy it matches.
  3. Your Compliance API recommendation, with the name of whoever would read it.
  4. Where telemetry lands.

What to bring your visibility owner: the three pieces, your retention recommendation, your Compliance API recommendation and who would read it, and where telemetry lands. Hand it to the visibility owner you named in Lesson 2, and record their call when it comes back.

→ Record this in the work-along companion, section Lesson 11.

Downloads

  • Work-along companion (Word)

    The editable Word version of the rollout plan you fill in lesson by lesson.

    Download
  • Work-along companion (PDF)

    The same rollout plan as a fillable PDF.

    Download

What’s next

You can now prove what happened. Whether the rollout is working is a different question with a different answer: the adoption signals your analytics report, read against the rollout objective you set in Lesson 1. The next lesson shows you how to read them, where they can mislead you, and how to set the pace goal that gives your objective a timeline.