Claude Enterprise: Smart Reports show what AI costs and delivers at work

What this is about

On October 28, 2026 (US time), Anthropic made three new features generally available for its business offering, Claude Enterprise. The background: many companies now have to show what they spend on AI and what they get for it. The new tools are called Smart Reports, In-app Surveys and Pooled Group Budgets. They're aimed at administrators and IT departments, not at end users who just want to chat.

Smart Reports: what teams do with Claude and what it costs

An administrator picks a team, a time period and the products to include. According to unite.ai, which cites Anthropic's help pages, the period can reach back up to 28 days. You can choose Chat, Claude Code, Claude Cowork, or a combination of Chat and Cowork. Claude then reads a sample of conversation transcripts and sorts the work into workstreams (thematic bundles of work) and by output type, such as analysis, documentation, text drafts or code.

For each workstream, the report shows:

• the cost (at list price, according to unite.ai),
• the outputs produced,
• obstacles, such as missing connectors (links to external services), failed tool calls, approval or sign-in hurdles, and rework loops,
• according to unite.ai, also the average cost per session and a count of sessions with no usable result, given only as a total with its cost.

There's also a list of recommendations. It might suggest, for example, defining recurring processes as Skills (stored standard procedures) or adding rules to code repositories. You can share insights as an Artifact (a shareable output format). Administrators can ask their own questions about model usage and costs and save them as a template. Anyone with billing permissions can raise a spending limit directly from the report. Organization owners can also let team leads without admin rights generate reports for their group.

According to unite.ai, generating a report usually takes a few hours. You can download it as HTML. Each organization can generate up to 10 reports per month for free. The feature is off by default until an owner enables it in settings. Reports can be limited to groups with role-based permissions, including groups synced from an identity provider via SCIM.

Privacy guardrails

Anthropic says explicitly that Smart Reports are meant to help you understand adoption and investment, not to evaluate individuals or make employment decisions. According to unite.ai, two safeguards are always on. Custom questions may not touch sensitive topics. These include protected characteristics, health, union activity, harassment, whistleblower reports, political or religious views, private life, and sentiment about named individuals. Sessions on such topics only feed into aggregated analyses. Attribution to individual users is off by default. If you turn it on, viewers initially see only "User," and every reveal of a name is logged. According to TechCrunch, as cited by note.com, Anthropic doesn't use this data for model training by default.

In-app Surveys: asking users directly

The survey feature captures employees' own view. Four default questions are preset: What was the most valuable work you did with Claude? How has your time spent changed? Which way of using Claude should colleagues adopt? What would stall or be delayed if access went away tomorrow?

Admins can edit or delete questions and set the audience, time period and where the survey appears. It shows up as a short card after a Claude reply. According to unite.ai, it never interrupts a task in progress, takes about 90 seconds and appears only once per user. Responses go to the organization's admins, not to Anthropic. Results appear continuously on the survey page and can be exported as a spreadsheet or CSV, including email address, group and where the response was given. Anthropic treats the responses as customer data and accesses them only to operate the service. Surveys work with customer-managed encryption keys and follow a single, uniform retention period. According to unite.ai, they aren't available with a HIPAA configuration.

Pooled Group Budgets: one pot per group

Until now, limits were mostly set per person. Heavy users hit the cap early in the month, while others leave budget unused. A group budget sets a shared monthly amount across the group. Personal limits stay in place as a brake on outliers. Each request counts against both, and whichever runs out first stops usage. Anyone nearing a limit can request extra budget in Claude, which admins then review. Users can't see the group itself or the pool amount.

During setup, a baseline is suggested based on recent usage. According to unite.ai, the documentation recommends roughly three times the group's monthly spend as the pool, and three times the top individual user's spend as the personal limit. In the example, a group spending $1,450 with a top user at $310 gets a $4,350 pool and a personal limit of about $950. People with billing permissions get emails at 50, 75, 95 and 100 percent. The indicator turns yellow at 75 percent and red at 90 percent. If usage is heavy before the 20th of the month, the docs suggest raising the pool by 50 percent.

According to unite.ai, pools are available on usage-based Enterprise plans, as well as on Standard or Premium seat plans with usage credits enabled. There, only the credits count, not the usage included in the seat. If someone belongs to several groups, the largest pool pays first by default. Budgets can be frozen or paused. They're managed only in organization settings. There's no admin API yet.

Background and reactions

According to note.com, which summarizes TechCrunch, Anthropic reported an operating profit in the second quarter of 2026 and a net revenue retention of 900 percent in its S-1 filing for its IPO. At the same time, many Fortune 500 companies reportedly haven't yet seen big AI wins. The analytics firm Jellyfish reports the following for coding agents (data from 276,000 customers): Claude Code and Anthropic models are the most widely used, spending has grown tenfold since the start of the year and accounts for about 2 percent of engineering costs. Jellyfish's chief product officer Nikolas Albarran sees the reports as a kind of "FDE-lite," a slimmed-down version of the Forward Deployed Engineers that AI labs station with customers. The more deeply embedded a product is, he says, the higher the switching costs.

Anthropic quotes monday, Slalom, Wix, Deloitte, DoorDash and Docusign. Wix says each report draws on tens of thousands of sessions. DoorDash says it saw things it hadn't known about despite being close to the team. These are statements selected by the vendor.

Our take

These features matter for IT leaders who have to explain AI bills. They close a gap that widens with token-based billing. On the critical side, the analysis rests on samples and on Claude's own classification, and costs are shown at list price. The promise not to use the reports for performance reviews is a policy, even though technical safeguards exist. For regulated customers with HIPAA or their own keys, the value is limited. There's no independent evidence yet of any actual savings.