Anthropic has released smart reports, a beta feature for Claude Enterprise that reads a team's Claude sessions and explains how the tool is actually being used. Instead of the session counts and spend figures already shown in the admin console, it produces prose and charts covering what kind of work got done, what it cost, and where people ran into friction. During the beta, each organization can create up to 10 reports per month at no charge.

Reading the transcripts themselves

Reports can be created by Primary Owners, Owners, Admins, or anyone with a custom role that has analytics view access. When setting one up, you choose the team or department to cover, a time range of up to the last 28 days, the products to include (chat, Claude Code, Claude Cowork), the focus areas for the analysis, and any custom questions you want Claude to answer.

Claude then reads a sample of the transcripts in that scope, groups them into workstreams and output types, attaches spend to each group, and writes up what it found. Every chart is interactive. Clicking a workstream opens the sessions inside it ranked by cost, each with a one-paragraph summary, the product used, the date, and the output type. The whole report can be downloaded as HTML with the drilldowns intact.

Generation takes a few hours, and an email arrives when the report is ready.

Cost and friction, broken out side by side

The structure of a report is fixed. Workstreams show what the group used Claude for most, by session count and by spend. Deliverables are grouped by output type, such as analysis, documentation, content drafts, and code. There is an average spend per session for each output type, and a view of what each session actually produced.

What stands out is that the failures get the same level of detail. Frictions are sorted by category: a connector that was never set up, output that did not match the request, approval or sign-in gating, tool failures, and rework loops, each with a count. Sessions that produced nothing usable, and sessions that were personal or off-topic, are counted only as an aggregate figure with no summaries and no drilldown.

A further section looks for repeated patterns worth packaging as a shared skill, so a team stops reinventing the same steps. The examples given include turning call notes into follow-ups, building account briefs, and drafting quarterly review outlines, and each card carries a suggested prompt for setting the skill up. The report also lists where spend concentrates, and the sessions that scored highest on task complexity, time saved, how long Claude worked on its own, and the expertise required. If custom questions were added at setup, the answers appear in their own section alongside the sessions that informed them.

A line drawn against individual performance review

Anthropic states plainly that the feature is not designed for evaluating individual performance or making employment decisions, and the design reflects that. The organization-level toggle for smart reports is off by default. A separate toggle controlling whether individual users can ever be identified is also off by default.

With that second toggle off, every drilldown shows "User" in place of a name and session IDs stay hidden. With it on, drilldowns still default to "User", and real names appear only when a viewer takes an explicit action to reveal them, which is logged. Personal conversations surface only as aggregate data, with no summaries, no individual sessions, and no names.

Some configurations are excluded outright. Organizations using customer-managed encryption keys, HIPAA configurations, or Access Transparency cannot use smart reports, and neither can Enterprise organizations running Claude Code with zero data retention.

From counting usage to reading what happened

On July 2, 2026, Anthropic added usage and cost analytics to the Claude Enterprise admin console, showing usage and spend by group and by user next to the artifacts created and the skills and connectors used, along with Claude Code metrics, an analytics chat that takes natural language questions, and an API for pulling the data into other tools.

Smart reports sit on that foundation but change the object of study. The earlier work counted how often Claude was used and how much was paid; this reads the content of the sessions and describes what people were doing. Plenty of companies have deployed AI and then struggled to explain the return, and one way to read this release is that vendors are now supplying the material that gets carried into a renewal conversation.

A report can be shared with specific people inside the organization, and it opens only through the sharing link, for those recipients and for admins who already have viewing access. Recipients can be chosen while the report is still generating, and reports that are no longer needed are archived first and then deleted.

Summary

Smart reports is a beta feature for Claude Enterprise that has Claude read up to 28 days of transcripts and return them as workstreams, deliverables, cost, friction, and candidate shared skills. Generation takes a few hours, and the beta allows 10 free reports per organization per month. The boundaries drawn from the start, with individual identification off by default and personal conversations reduced to aggregates, are worth weighing alongside the analysis itself.

Note: the thumbnail image is AI-generated and illustrative.