Smart reports: what your team actually uses Claude for, read from the transcripts
In brief
On September 10, 2026 Anthropic released smart reports in beta for Claude Enterprise. An admin picks a team and a window of up to 28 days, and Claude reads a sample of that team's transcripts and reports on workstreams, cost per session, friction, and repeated work worth turning into a shared skill. Here is what is in a report, who can see names, and how to get a useful one out of your 10 free reports a month.
Contents
Usage dashboards tell you how much people use Claude: seats active, messages sent, tokens spent. They do not tell you what the work was, whether it went well, or where it cost money for nothing. That gap is why so many AI rollouts stall at the "we have adoption numbers but can't say what we got" stage we covered in measuring Cowork adoption.
Smart reports, released in beta for Claude Enterprise on September 10, 2026, go after that gap directly. Claude reads a sample of conversation transcripts within a scope you choose and writes a report on what the work was.
Who can run one
Primary Owners, Owners, Admins, and custom roles with Analytics view access. Delegated admins whose access is scoped to specific groups or departments can also run reports, limited to those groups.
It lives at Analytics → Smart reports (beta) → New report.
What goes in
- Scope: groups, departments, or cost centers (from SCIM provisioning), or the whole org.
- Time range: up to 28 days.
- Optional documents: add context such as your rollout goals or team charter, so the analysis knows what you care about.
- Custom questions: ask something specific and get an answer section for it.
What comes out
| Section | What it tells you |
|---|---|
| Workstreams | What the team used Claude for, weighted by sessions and spend |
| Deliverables | Output types — analysis, documentation, code, drafts |
| Cost per session by output type | Which kinds of work are expensive |
| Task outcomes | Whether sessions got where they were going |
| Most common frictions | Blockers, failed tools, rework loops |
| Inefficiencies | Unusable outputs, off-topic sessions |
| Reusable skills and workflows | Repeated patterns worth packaging as a shared skill |
| Most expensive sessions | The outliers driving your bill |
| Complex, autonomous work | Where Claude ran multi-step work on its own |
| Custom question answers | Your own questions |
The two sections most operators will get the most from are frictions and reusable skills. Frictions are a to-do list: a connector that keeps failing, a task people redo three times, a tool with the wrong permissions. Reusable skills are a build list — the report points at work several people keep doing by hand, which is the right input for building skills for your team instead of guessing what to build.
Privacy: what the report shows about individuals
By default, personal conversations appear only in aggregate — no summaries of individual sessions, no names. Admins can turn on attribution to reveal user names and session IDs, and every unveil is logged.
Anthropic's help article says the feature is not designed for, and should not be used for, evaluating individual performance or making employment decisions. Take that seriously in how you announce it. If your team hears "leadership is now reading our Claude chats," you will lose the honest usage that makes the report useful. Tell people before your first report what it covers, that it is aggregate by default, and what attribution would be used for. If you already have an AI usage policy, add a paragraph there.
Beta limits
- 10 free reports per organization per month, reset on the 1st. More requires your account team.
- Not available for organizations using CMEK, HIPAA configurations, or Access Transparency.
- Not available for Claude Code sessions in zero data retention environments.
- Reads a sample of transcripts, not all of them. Treat numbers as directional.
Getting a useful report out of 10 a month
Scope to one functional team, not the org. Anthropic's own guidance: a report on sales or finance produces more actionable findings than an org-wide one, because workstreams and frictions are specific to how a team works. An org-wide report tends to say "people write and summarize things."
Run the same scope twice, four weeks apart. The first report is a baseline. The second tells you whether the connector you fixed or the skill you shipped changed anything. That before/after is the evidence you need for an ROI conversation with leadership.
Use custom questions for your actual decision. "Which recurring tasks in finance could run as a scheduled task?" beats a generic report you then have to interpret.
Pair it with the cost data. Most expensive sessions plus spend controls is how you find the one workflow burning a quarter of the budget.
A sensible first month: one report on your highest-usage team, one on the team you think is stuck, and keep the rest in reserve for follow-ups.
How it fits with the other admin tools
- Analytics dashboard: counts. Smart reports: what the counts were about.
- Compliance API: raw programmatic access to activity for audit and retention. Smart reports: a written read of that activity for decisions.
- Spend limits: controls. Smart reports: the diagnosis of where to apply them.