Claude Enterprise can now tell you what your AI spend produced, broken out by workstream, and it gets there by having Claude read a sample of your employees' transcripts. Anthropic made smart reports, in-app surveys and pooled group budgets generally available on October 28. Smart reports are off by default, and only an Owner or Primary Owner can enable them.
Before anyone does, settle three questions: who may reveal an employee's name, what you will do with the count of personal chats, and how far you trust dollar figures calculated at list price. Budgets and surveys need less preparation, although each has a trap of its own.
What Smart Reports Show That Usage Dashboards Do Not
Smart reports connect spend to the work it paid for, which token and seat counts cannot do. An admin picks a team, a date range and the products to include, then Claude groups a sample of transcripts into workstreams, each with its spend, deliverables and friction points such as a missing connector or a failed tool call. The existing analytics dashboard and spend export give you per-user, per-model tokens and estimated spend. They do not tell you whether the money went to contract review or to reformatting slides.
The case for wanting this is strong. Jellyfish, which sells an engineering-management platform, drew its October report from more than 1,300 companies and 276,000 developers. It put median Claude Code spend at $63 per active developer per week, or $252 on a four-week month, with a 90th-percentile monthly equivalent of $1,666. R&D World's account does not say whether those spend figures cover the full sample, and Anthropic's own docs run lower: an average of $150-250 per developer per month, with 90% of users under $30 per active day. An earlier Jellyfish analysis, as reported by Digital Today, found its top 10% of developers used about 10 times the tokens of mid-level users for about twice the output, although that report does not define output and other accounts describe the comparison group differently. On those numbers, consumption alone looks like a weak signal of value, a problem M&T's Copilot seat counts showed from the licensing side.
Anthropic's launch post carries endorsements from Monday, Slalom, Wix, Deloitte, DoorDash and Docusign. A Wix engineer says each report draws on "tens of thousands of sessions". Those are vendor-selected testimonials with no published outcome data, so read them as evidence of use.
Who Can See a Name Once Attribution Is On?
Report viewers can reveal names once your organization enables attribution, and each reveal is logged. The org-level setting, "Allow attribution to individual users," is off by default. With it off, drilldowns show "User" and hide session IDs. With it on, drilldowns still start at "User," but viewers get a control that reveals names and session IDs. A per-report setting can also extend the attributed view to people the report is shared with.
There is a pull toward switching it on. The Recommendations section identifies people whose work is held back by their spend limit, and billing users raise those limits from the attributed view. An admin who wants that convenience needs the same toggle that exposes names across every drilldown.
Delegation widens the audience further. Owners can let non-admin team leads run reports for assigned groups, departments or cost centers. Delegates cannot run whole-organization reports or see other people's reports unless shared, and privacy guardrails apply to them too. Two details deserve attention: delegates share the organization's limit of 10 no-charge reports a month, and when a delegate's scope changes, reports they already ran stay available to them. A department head reassigned to another team keeps the old team's report unless their delegate access is removed.
Anthropic says plainly that reports aren't for evaluating individual performance or making employment decisions. That sentence is a statement of intended use. Nothing in the two help articles describes a technical control that stops a manager from drawing conclusions from a logged reveal, so the restriction has to live in your own policy.
How Personal Chats Land in the Inefficiencies Count
Personal chats appear as a cost line in the report, with protections that are real and incomplete. The help page says "Inefficiencies counts sessions that produced nothing usable and sessions that were personal or off-topic," and shows them only as an aggregate count and cost, with no summaries or drilldown.
Anthropic has also put in always-on guardrails. Custom questions cannot address restricted topics, which include protected characteristics, health, union activity, harassment, whistleblower or legal matters, political or religious views, personal life and sentiment about people. Sessions Claude identifies as touching those topics are used only for aggregate analysis.
Three gaps remain. Identification is Claude's own judgment, and Born City notes that Anthropic has published no guaranteed detection or error rates for these filters. A session that discusses a medical leave and gets misread as ordinary work would not get the aggregate-only treatment.
The second gap is team size. The help page states no minimum group size, and reports can be filtered to an RBAC group, department or cost center. On a team of four, an aggregate figure for personal chats comes close to describing a person. Microsoft's Copilot Dashboard, by comparison, shows group metrics only for groups at or above a threshold that defaults to 10 employees and cannot go below five. Copy that discipline by policy, since the Anthropic pages describe no such setting.
The third gap is interpretation. The label puts a failed agent run and a question about a child's homework under one heading. If your acceptable-use policy allows incidental personal use, the number is close to meaningless as a cost target. If it forbids personal use, the report gives a manager a dollar figure to chase. Decide which applies before the first report circulates.
In Germany the toggle may be the smaller question. Section 87(1) no. 6 of the Works Constitution Act gives works councils co-determination over "technical devices designed to monitor the behaviour or performance of the employees." One German compliance guide reads that threshold as broad, reaching analytics and dashboards, and advises considering a data protection impact assessment before rollout. GDPR Article 35 requires that assessment before processing likely to create high risk. Whether a default-anonymous report crosses either line is a question for your counsel and your works council, and it should come before the toggle.
Why Report Spend Will Not Match Your Invoice
Treat workstream dollars as a way to compare work, and keep the invoice as your system of record. The help page describes spend as "what it costs at list price." If your contract carries a discount or commitment, every workstream total will overstate what you paid.
Sampling adds a second layer. Claude "reads a sample of transcripts in that scope" and "attaches spend to each group," and the creator can preview how many sessions will be sampled. The page gives no fixed sample size and no statement on whether group totals are scaled up from the sample, nor any accuracy range. The spend export measures something different again: it reports net and gross spend per person and model, looks back 90 days against the report's 28, and on seat-based Enterprise plans covers only spend above the seat allotment.
Run one report, then reconcile it against the export and the invoice for the same period. Write the ratio down, and put it on every slide that quotes a workstream number. Anyone still sorting out how Claude Enterprise billing works should start with what usage-based billing really costs.
Which Organizations Cannot Turn Smart Reports On?
Organizations with the strictest data setups are excluded. Smart reports are unavailable with customer-managed encryption keys, a HIPAA configuration or a hybrid deployment that runs inference on your own cloud provider. If you have zero data retention for Claude Code, reports cover chat and Cowork and leave Claude Code out.
That exclusion can distort the picture. A report that omits Claude Code ranks workstreams only on the spend it can see, so an engineering organization should label it as partial.
What Pooled Budgets Fix, and the Shared Stop They Add
Pooled budgets reduce stranded allowance, in exchange for the chance that a whole group stops at once. Every request counts against both the user's own limit and the group's pool, and a user stops when either runs out. When a pool empties, usage pauses for members with no other pool until an admin raises it or the month resets.
The individual limit does the protective work. Using Jellyfish's numbers as an illustration, a 100-developer group funded at the $252 median gets a $25,200 pool. Set each person's limit at the $1,666 90th-percentile figure, and 15 people at that limit consume $24,990 of it. Your own distribution will differ, so size the limit from your export.
Five operational details will shape a rollout:
- The announcement calls the feature generally available, yet Anthropic's consumption guide still labels it beta. Ask your account team which terms apply.
- Pools aren't available through the Admin API yet, so any automation you built on earlier spend controls will not manage them.
- On Standard and Premium seat plans, the pool applies only to usage credits.
- Members never see the group's name or the pool amount. They get a message and a "Request more" button, so route those requests to someone who is watching.
- Limits reset at 00:00 UTC on the 1st, and a member with no applicable limit has uncapped spend.
Alerts go to billing users at 50%, 75%, 95% and 100%. Work in progress may pause at the limit, which Anthropic says loses nothing already produced, though a deadline may still slip. Uber's AI budget overrun is the case for having a cap at all.
Surveys Carry Less Risk, but They Are Not Anonymous
Surveys are the easiest of the three to adopt, provided nobody promises anonymity. Four default questions ask about valuable recent work, time changed, a practice worth copying and what would break without access, and admins can edit them and choose the audience. The card tells members that answers go to admins, and the CSV export includes each respondent's email.
That changes how you read the results. Someone asked by name what would break without Claude has an incentive to say a lot would, especially with budget decisions in the air. Use the answers to find workflows worth copying and leave them out of the ROI model. Surveys now work with customer-managed keys but not with a HIPAA configuration, and Anthropic warns regulated customers to keep regulated data such as PHI out of responses.
What to Do Before an Owner Flips the Toggle
This Week:
- Check whether you are eligible: CMEK, HIPAA, hybrid deployment and Claude Code zero data retention each change the answer.
- Confirm who holds Owner and Primary Owner roles, and ask them not to enable smart reports until legal and HR have signed off.
- Check your spend-limit settings and find members with no limit. Their spend is uncapped today, whatever you decide about pools.
This Month:
- Write a one-page rule covering who may enable attribution, who may reveal a name and why, and who reviews the reveal log. Anthropic logs each reveal; the help page does not assign anyone to read it.
- Set a minimum team size for reports. Microsoft's default of 10 is a reasonable starting point.
- Decide what the personal-chat count means under your acceptable-use policy, and tell employees before the first report runs. If you have a works council, take it to them first.
- Run one whole-organization report with attribution off. Reconcile it against the export and invoice, then decide how to ration the 10 monthly reports before delegating any.
- Pilot one pool on a group with uneven usage, with individual limits sized so a handful of people cannot empty it.
Before Renewal:
- Ask Anthropic for three things in writing: a minimum group size setting, published error rates for restricted-topic detection, and workstream spend at your contract rates.
- Check how portable your skills and reports are. Channel Insider's analysis warns that better visibility may deepen dependence once integrations and skills are built, and buyers can test that claim before signing.
The Bottom Line
This release moves AI cost management a step past capping consumption and toward asking what the consumption bought, the question finance has been asking of Copilot seats too. Reading transcripts is how Anthropic answers it, and the product's defaults are careful: reports off and attribution off, with an always-on guardrail holding sensitive topics to aggregates. Either default lasts until one Owner changes it. Write the attribution and team-size rules first, then run the first report with attribution off and reconcile it against your invoice.
Continue Reading
- Claude Enterprise Spend Controls: 5 Things to Configure Now
- Claude Pricing 2026: Usage-Based Billing Real Cost Breakdown
- Claude Enterprise Billing Goes Usage-Based: ROI Impact 2026
- True Cost of a Copilot Seat: The Licence Is the Floor
- M&T Published 3 Copilot Counts. None Was Weekly Actives.
- Agentic AI Pricing: Don't Buy Consumption Without a Cap
