Anthropic made smart reports generally available for Claude Enterprise on October 28, after a beta that began September 10. An admin picks a team, a period of up to 28 days and the Claude products to cover. Claude then reads a sample of that team's transcripts, groups the sessions into workstreams, puts a list-price cost on each and lists where work got stuck. The feature is off by default, and each organization gets 10 reports a month at no charge.
Mashable's headline says the reports tell bosses when workers are "off-topic". Anthropic's help center is more specific: the reports do count personal and off-topic sessions, but only as a total and a cost, with no summaries, sessions or names behind the number.
What admins and team leads can and cannot see
Anthropic's setup guide and its page on delegated access set out the limits.
Item | What the report shows |
|---|---|
Work sessions | A one-paragraph summary each, with product, date and output type, ranked by cost |
Who ran a session | "User", unless an Owner allows attribution; each reveal is logged |
Personal or off-topic | A count and a cost only, with no summary, session or name |
Restricted topics | Aggregate analysis only, with no summary |
Questions on them | Blocked; the guardrail cannot be turned off |
Team leads | Only the groups, departments or cost centers assigned, never the whole organization |
Source: Anthropic's help center, checked November 1, 2026.
The seven restricted topics include health, protected characteristics, personal life, political or religious views and union activity, which covers "coworkers discussing pay or working conditions". Claude decides which sessions fall into them, and the pages do not say how accurately.
Names stay hidden unless an Owner turns on "Allow attribution to individual users", also off by default; even then, a viewer must click to reveal them. Both pages carry one caveat.
They aren't designed and should not be used for evaluating individual performance or making employment decisions.Anthropic's help center, "Get started with smart reports"
That sentence is guidance; the settings do not enforce it. The Recommendations section identifies "people whose work is being held back by their spend limit", and an admin with billing permission who is using the attributed view can raise those limits from the report.
Who can run one, and where it does not work
Only a Primary Owner or Owner can switch the feature on, under Organization settings, Capabilities. Owners, Admins and custom roles with analytics access then create reports, which usually take a few hours. Owners can also let team leads who are not admins report on up to 50 groups, departments or cost centers of each kind; those delegates see only the reports they ran or were shared and cannot delete any. More than 10 reports a month means contacting Anthropic's account team, which publishes no price.
Organizations that use customer-managed encryption keys, a HIPAA configuration or a hybrid deployment cannot use smart reports at all. Where zero data retention is on for Claude Code, reports cover chat and Claude Cowork and leave Claude Code out.
How it compares with Microsoft, OpenAI and Google
SaasBuzz compared the usage dashboards each vendor documents. None is a compliance archive: OpenAI sends admins to its Compliance API for raw logs.
Dashboard | Reads chats? | Privacy rule |
|---|---|---|
Claude smart reports | Yes; a summary per session | Topic blocks; names hidden by default |
ChatGPT Enterprise | Classifies them; totals only | Small groups left off the heat map |
Microsoft Copilot | No; counts actions | Minimum group size |
Google Gemini | No; counts use | None given; usage level shown per user |
Source: vendors' documentation, checked November 1, 2026: Anthropic's help center, OpenAI's workspace analytics, Microsoft's Copilot Dashboard and Google's Gemini usage reports.
OpenAI's Task insights, on by default, sorts ChatGPT conversations into task types "without exposing individual prompts, conversations, or user-level activity", and leaves SCIM groups under 10 members off its heat map. Microsoft's dashboard hides metrics "for groups that are smaller than the minimum group size". Google's gives each user a usage level and a count of active days.
Claude's reports go furthest into the work itself. Of the four, only Anthropic's puts a written summary of a single session in front of an admin, and its privacy controls run on topics and names, where OpenAI's and Microsoft's run on group size.
What it means for buyers and employees
Anthropic is selling the reports on cost. Its help center suggests customers "bring evidence to renewals". In its announcement, John Jannone, AI enablement lead for engineering enablement at DoorDash, said the reports "still surfaced things I didn't know" about teams he already works closely with.
Finance and IT buyers face the same question across every AI seat they pay for: whether usage is turning into work. SaasBuzz reported that tension in Amazon's memo telling engineers that outcomes matter more than AI usage. Microsoft and Google answer it with counts, and OpenAI with task categories totaled across the workspace. Anthropic answers it with a model's reading of the transcripts, which tells a buyer more and exposes more of the work of the people whose sessions are sampled.
For employees, the defaults decide that exposure: work sessions in summary, personal sessions as a number, and no names unless an Owner chooses otherwise.
What Anthropic's pages do not say
Whether members are told when a report samples their sessions.
How many sessions a sample holds; the help center says only that the count is previewed before a report runs.
How often Claude misfiles a personal or restricted session as work, which would put its summary in front of an admin.
What reports beyond the free 10 a month cost.
