AI News
5 min read

Claude Smart Reports: What Admins Learn From Employee Chats and When Works Councils Get a Say

Colorful building block towers on a light background, a large block magnifying glass above, two toggle switches beside them
Image: AI-generated
In short

Claude Smart Reports read a sample of your workforce's real chats and condense them into workstreams, costs, and problems. Names start out hidden, but a second toggle lets readers reveal them. If your company has a works council in Germany, involve it before you turn this on: under German case law, what counts is whether a tool is capable of monitoring people, not what you intend to do with it.

What exactly are Claude Smart Reports?

Smart Reports are an analytics feature for Claude Enterprise that reads content instead of just counting it. According to the release notes, Anthropic launched them in beta on September 10, then made them generally available on October 28. In Germany, the tech news site heise covered them on November 4.

Here is how it works, per Anthropic's help page: you pick a team, a date range of up to 28 days, and the products (chat, Claude Code, Cowork). Claude then reads a sample of the session transcripts. Anthropic gives no fixed sample size, but you see the number of sessions before you start. In Anthropic's announcement, Wix describes reports based on tens of thousands of sessions.

Ten reports per month are included at no extra charge. Anthropic names only Claude Enterprise, and the Team plan does not appear on the help page. Enterprise organizations with customer-managed encryption keys (CMEK), a HIPAA configuration, or a hybrid deployment in their own cloud are excluded too. If you agreed to Zero Data Retention for Claude Code, you get reports for chat and Cowork only.

What do admins see in a report?

Admins see what a team works on, what it costs, and what goes wrong. That goes well beyond a usage statistic:

Section Content Why it can be sensitive
Workstreams, outcomes What Claude was used for, what each session produced Shows the work, not just the volume
Most expensive sessions Where spend concentrates Individual sessions stand out
Inefficiencies Sessions without a usable result, personal or off-topic sessions as a total with cost Exposes a team's share of personal use
Complex work Sessions scoring highest on complexity, time saved, and expertise Reads like a leaderboard
Custom questions Answers to freely worded questions The admin decides what gets asked

Anthropic built in guardrails that cannot be turned off. Sessions about health, union activity, harassment, whistleblowing, political or religious views, private life, and sentiment about named individuals feed only into the aggregate analysis and get no summary. Custom questions on these topics are blocked. Whether a session falls into one of these categories, however, is up to Claude.

Do employees stay anonymous in a report?

Only as long as both toggles allow it, and even then not in every case. The first toggle enables Smart Reports, the second allows attribution to individuals. Both are off by default. Without attribution, detail views show only "User" and session IDs are hidden. With attribution, "User" remains the default display, but whoever reads the report can reveal names and session IDs. Every reveal is logged.

Anthropic says the reports should not be used to evaluate individual performance or make employment decisions. That is a usage policy, not a technical lock.

Our take: in a team of four, "User" protects nobody. The department knows who built the newsletter on Monday. Neither the help page nor the page on team leads with their own access mentions a minimum group size. That access lets team leads create reports for their own groups without an admin role.

Does a German works council have to approve Smart Reports?

If your company has a Betriebsrat (a works council, the elected employee body in German companies), assume yes. Under Section 87(1) no. 6 of the Works Constitution Act, it has a say on technical systems that can monitor behavior or performance. In a 2022 case about Microsoft Office 365, Germany's Federal Labor Court made clear that the employer's intent is irrelevant: objective capability is what counts. It was undisputed in that case that the data generated was suitable for monitoring performance or behavior. With multiple sites and central administration, the central works council was responsible.

So Anthropic's note that the reports are not meant for performance reviews does little for you. The law firm Covington treats an agreement before productive use as mandatory for AI that can monitor people. It lists injunctions, a conciliation committee, and possible suspension of the tool as consequences. It also points out that the governing coalition is planning reforms that could change these rights.

Without a works council, codetermination falls away, but data protection does not. The help page says nothing about retention, EU processing, or informing employees, so you need to clarify those points with Anthropic and with whoever handles data protection at your company. This section is not legal advice.

What does this mean for you?

Leave the toggle off until these five points are settled:

  1. Check existing agreements. Does a works agreement on Claude cover only usage counts, or reading content as well? Reading content is something new, even if Claude has been running for a long time.
  2. Put the purpose in writing. For example: only identifying needs for training, templates, and integrations.
  3. Skip attribution. You do not need names to find out which work can be automated.
  4. Keep groups large. Our guideline: no reports on fewer than ten people, and no team lead access for small groups. That is our own rule, not an Anthropic requirement.
  5. Limit custom questions. A fixed template instead of free-form questions makes clear what actually gets analyzed. Anthropic offers saved question templates.

The upside is real: recurring processes, missing connections, and failed tool calls show you what is worth automating next.

Want to roll out Claude in your company and set up permissions, rules, and costs properly from day one? That is what we do in our AI consulting.

What is enough if you only care about costs?

Then you do not need transcripts. Questions and tools at a glance:

Your question Tool Reads chat content?
What does each project cost? Project tags (beta since November 6) No, spend per tag
Who uses how much? Enterprise analytics: the dashboard shows individuals (Top spenders, Leaderboard), the spend export shows each person's email Per the help page, usage and cost figures; content access not described
Is Claude helping the team? In-app surveys (anonymity not described there) No
Where does the work fail? Smart Reports Yes

So even standard analytics can fall under codetermination if it shows individuals. If your cost problem comes down to model choice, take a look at our math on Claude Haiku 5.5. For how OpenAI is restructuring its limits, see our article on the ChatGPT usage limit.

Frequently asked questions

Does a Smart Report cost extra?

Up to ten reports per month and organization are free, and the quota resets on the first of the month. Failed reports do not count, queued ones do. For more, you need to contact your Anthropic account team.

Does Anthropic use the analyzed chats for training?

According to its privacy page, Anthropic does not use inputs and outputs from commercial products for training by default. The Smart Reports help page does not address the topic separately.

Who can turn on Smart Reports and who can create them?

Only Primary Owners and Owners can turn them on. Admins, roles with analytics permission, and team leads with access to their own groups can also create and view reports.

Can a report be deleted?

Yes. You archive it first, turn on the archived view, and delete it there. Reports can also be downloaded as HTML, though, and you have to keep track of those copies separately.

Sources

  1. Anthropic: Get started with smart reports
  2. Anthropic: Smart reports, surveys, and pooled group budgets
  3. German Federal Labor Court: Order 1 ABR 20/21 of March 8, 2022
  4. German Works Constitution Act (official English translation), Section 87
  5. Covington: Works Council Rights in AI Implementation Projects
This article was created with AI.
Got a project in mind?

Tell us what you're planning. You'll get an honest assessment within 24 hours.

Got it! We will get back to you asap!
Whoopsie! Something went totally wrong here... Maybe just send an email to hi@blocksandcolors.de? Sorry!
Or email me directly: hi@blocksandcolors.de