Anthropic has launched smart reports for Claude Enterprise. Claude reads a sample of a team’s conversations and turns them into a report: what people worked on, what they produced, what it cost, where they got stuck, and how many sessions were „personal or off-topic”. Headlines have focused on bosses reading workers’ AI chats. The reality is more nuanced, and worth understanding whether you run a Claude rollout or just use Claude at work. Here’s what the reports show, what they deliberately don’t, and what to do about it.
What Anthropic announced
On 28 October 2026, Anthropic made three admin features generally available for Claude Enterprise:
- Smart reports: AI-generated analysis of how a team uses Claude and what that work costs.
- In-app surveys: short questions shown to employees after a Claude reply, such as „What would break first if the team lost access?”
- Pooled group budgets: one shared monthly budget per group, so heavy users don’t run out early while others leave budget unused. Admins get email alerts at 50%, 75%, 95% and 100%.
Smart reports get the attention, because they are the first Claude feature in which the AI itself reads employees’ conversations and summarises them for management. They first appeared as a beta in September and are now generally available.
How a smart report works
An admin picks a team, a date range of up to the last 28 days, and the products to include: Claude chat, Claude Code, Claude Cowork, or the unified chat and Cowork experience. Before starting, they see a preview of how many sessions will be sampled. Claude then reads a sample of the transcripts in scope, not every conversation, and groups the work into workstreams. A report usually takes a few hours and arrives by email.
| Report section | What it shows |
|---|---|
| Workstreams | Sessions and spend grouped by what the team used Claude for |
| Deliverables produced | Output types: code, analysis, documentation, content drafts |
| Cost per session by output type | Average spend per session for each type of work |
| Most common pain points | Friction such as a connector that wasn’t set up, tool failures, rework loops |
| Inefficiencies | Sessions that produced nothing usable, and personal or off-topic sessions, as a count and a cost only |
| Complex, autonomous work | Sessions that scored highest on complexity, time saved and run time |
| Most expensive sessions | Where the spend concentrates |
| Custom questions | Answers to the organisation’s own questions about usage and cost, saved as reusable templates |
| Recommendations | Workflows to turn into skills, rules to add to code repositories, and people held back by their spend limit |
Charts are interactive. Clicking a workstream lists its sessions ranked by cost, each with a one-paragraph summary, the product, the date and the output type. Reports can be shared as a link inside the organisation or downloaded as HTML.
The „off-topic” part, and what managers can’t see
Mashable’s headline, „Claude’s new smart reports tell bosses when workers are 'off-topic'”, is accurate, but the details matter. According to Anthropic’s documentation:
- Off-topic sessions are totals only. The inefficiencies section shows how many personal or off-topic sessions there were and what they cost, with no summaries and no drilldown into individual chats.
- Names are hidden by default. Drilldowns show „User” instead of a name. Seeing names requires a separate „Allow attribution to individual users” setting that only an Owner can switch on. Even then, a viewer has to reveal names in the report, and every reveal is logged.
- Restricted topics are always protected. Custom questions can’t ask about protected characteristics, health, union activity, harassment or whistleblowing, political or religious views, personal life, or sentiment about named people. Sessions on those topics count only in aggregates, with no summaries. Admins can’t switch this off.
- It’s off by default. An Owner has to enable it in Organization settings → Capabilities → Analytics.
- Not every organisation can use it. It isn’t available with customer-managed encryption keys, a HIPAA configuration, or a hybrid deployment that runs inference in your own cloud. With zero data retention for Claude Code, reports cover chat and Cowork but not Claude Code.
Anthropic is explicit about the intended use:
„Smart reports help you understand adoption and plan your investment in Claude. They aren’t designed and should not be used for evaluating individual performance or making employment decisions.”
Anthropic, Claude Help Center
That is a policy statement, not a technical barrier. With attribution enabled, a manager can still see which named person ran an expensive or unproductive session. That is why the concern is real: Mashable points to a survey in which one in four US managers said they use AI „often or all the time” for layoff decisions, and to former Meta employees’ allegations, which Meta denies, that AI tools and token usage were used to rank staff for layoffs.
What it means for companies
For an organisation paying for hundreds of Claude seats, the useful part isn’t the off-topic counter. It’s finally seeing where AI spend turns into work. Which teams get real value, which workflows are repeated often enough to become a shared skill, and which missing connector costs everyone time. Wix, quoted by Anthropic, says its reports are based on „tens of thousands of sessions”. No one reads that many by hand.
If you plan to switch it on, especially in the EU, treat it like any other tool that processes employee communications:
- Tell people first. Explain what is analysed, who sees the reports and why. Surprise monitoring destroys trust faster than any report builds it. Under GDPR, employees have a right to know how their data is processed.
- Involve your DPO, legal team and, where you have one, the works council or union before enabling it. Local labour law on monitoring employee communications varies by country.
- Keep attribution off unless you have a clear, documented reason to turn it on. Aggregate data answers almost every adoption question.
- Use the recommendations, not the rankings. Turning repeated workflows into skills and fixing pain points is where the return is.
- Write down that reports won’t be used for performance reviews, matching Anthropic’s own guidance, and stick to it.
What it means for employees
If your company uses Claude Enterprise, assume your work conversations may be read by Claude for reporting, even if no human reads them one by one. The simple rule: keep personal use on a personal account. Planning a holiday or asking about a health issue on a company seat was never private from your employer’s admins. Now it can also be counted. If you use Claude for both work and private projects, our guide to running separate work and personal Claude accounts shows how to keep them apart on one machine.
The upside for employees is real too. The report can flag people „held back by their spend limit”, and admins with billing permission can raise that limit straight from the report. Showing that your work needs more Claude is now visible data, not a request in a queue.
Availability and price
- Plans: Claude Enterprise only. Not available on Pro, Max or Team.
- Price: up to 10 reports per month per organisation at no extra charge. The count resets on the first of each month. For more, Anthropic says to contact your account team.
- Who can run them: Owners, Admins and roles with Analytics view access. Owners can also let team leads run reports for their own groups without making them admins.
Rolling out Claude across your company and want to measure value without losing your team’s trust? Talk to the Fireup team. We help set up AI tools, skills and policies that developers actually want to use.
