Claude smart reports: what they show, how to turn them on
2026-10-29 · 4 min read
Claude smart reports are an admin tool in Claude Enterprise. They read a sample of one team's Claude sessions from the last 28 days, group the work into workstreams, and show what each workstream produced, what it cost and where people got stuck. Anthropic made them generally available on October 28, together with in-app surveys and pooled group budgets, and an Owner has to switch them on under Organization settings before anyone can run one.
On X, the launch got read as an HR story first. @HedgieMarkets summed it up as Claude reading a team's transcripts and flagging "inefficiencies," including sessions it tags as personal or off-topic. The post also points out that the report "comes with a button to raise a worker's spending limit." A few minutes later, Irish Independent tech editor @adrianweckler quote-posted it with a sharper framing: "Anthropic's new tools for ordinary managers to weed out 'inefficient' staff."
Anthropic's announcement tells it another way. It says the reports are there to show what AI spend buys, and that they "aren't for evaluating individual performance or making employment decisions." I think both readings are fair, and the help center settles most of the argument.
What a smart report shows
The breakdown below comes from Anthropic's smart reports guide in the help center.
| Item | What you get |
|---|---|
| Workstreams | Sampled sessions grouped by type of work, each with deliverables, spend at list price and friction points (a missing connector or a failed tool call, for example) |
| Recommendations | One consolidated list per report |
| Inefficiencies | Personal or off-topic sessions, shown only as a count and a cost, with no summaries and no drilldown |
| Names | Hidden by default. Every drilldown says "User" |
| Limits | 10 reports per organization per month at no charge, each covering up to the last 28 days |
Two privacy details matter more than the headline. If an Owner enables "Allow attribution to individual users," a viewer can reveal names and session IDs, and each reveal is logged. And some topics are always off limits: health, union activity, political or religious views, personal life and opinions about named people stay out of reports, and that guardrail can't be turned off. Organizations with customer-managed encryption keys, a HIPAA configuration or hybrid cloud inference can't use smart reports at all.
How to turn smart reports on
- As a Primary Owner or Owner, go to Organization settings, then Capabilities, and find the Analytics section.
- Turn on the "Smart reports" toggle. Leave "Allow attribution to individual users" off unless you have a written reason to turn it on.
- Go to Analytics, then Smart reports, and click "New report."
- Pick a team, a date range of up to 28 days and the products to include. Check the preview of how many sessions will be sampled.
- Add questions or a saved template, such as "Which repeated tasks could become a shared skill?"
- Click "Create report," then share it from the Share button.
The other two features
Pooled budgets live under Organization settings, Usage, Spend limits, on the "By group/tier" tab. You pick "Edit limit and budget" on a group's row and set one monthly pool alongside each member's limit. In Anthropic's words, "a member stops at whichever runs out first." Billing admins get email alerts at 50%, 75%, 95% and 100%. The timing makes sense: business customers are a big part of what the Anthropic S-1 says about Claude, and they're the ones asking what their token bill buys.
Surveys show up as a short card after a Claude reply and take about 90 seconds to answer. One detail to know before you launch one: the help center says the CSV export lists each respondent's email. If you tell staff the survey is anonymous, that won't be true.
My read
The Hedgie and Weckler posts aren't wrong about the risk. Anything that counts "off-topic" chats and pairs them with dollar figures will tempt some manager to start counting heads. Still, the defaults are tighter than the X framing suggests: names stay hidden, sensitive topics are blocked, and revealing a name leaves a log entry. Where this goes wrong is an Owner flipping attribution on without telling anyone.
The useful part is the friction column. If one team keeps failing on the same missing connector, you know what to fix. That's more practical than another usage chart, and it pairs well with the live views in Claude Dashboards.
What this means for a small business
Most small offices aren't on Claude Enterprise, but the lesson carries over. Before you watch how staff use AI, write down what you'll measure, who can see names, and what you'll never use it for. Then tell the team before you turn anything on. Our post on ChatGPT reviewers and data retention covers the other side of the question: who outside your company can read those chats. If you'd rather start by finding which of your processes are worth automating at all, New Face Design offers a free process audit.