Claude smart reports: can your boss see 'off-topic' AI chats? What Anthropic's tool shows
Anthropic's new Claude Enterprise reports count 'personal or off-topic' chats and their cost. What managers see, what stays hidden and what it means for staff.
Companies using Claude Enterprise can now have Claude read a sample of their staff’s AI conversations and produce a report on what each team worked on, what it cost and how much of it counted as “personal or off-topic”. Anthropic made these “smart reports” generally available on Wednesday, 28 October, and the off-topic count has quickly drawn criticism as a new form of workplace monitoring.
The short answer for workers: your manager sees team totals and costs, not your name next to your chats, unless the company switches on a setting that is off by default.
What a smart report shows
An admin, or a team lead the company owner allows, picks a team, a period of up to 28 days and which Claude products to cover. Claude then analyses a sample of that team’s conversations and groups the work into “workstreams”, with the spend on each, the deliverables produced and the common friction points, such as a missing connector or a tool call that failed. Each organisation gets up to 10 reports a month at no extra charge.
Sessions that produced nothing usable, or that Claude classed as personal or off-topic, go into an “inefficiencies” figure with their cost. That is the part that caught attention: a machine, not a person, decides what counts as off-topic.
What stays hidden
Anthropic’s help pages set out several limits that it says cannot be turned off:
- Inefficiencies are shown only as a count and a cost, with no summaries and no way to open individual sessions. Personal conversations also come without names or session details.
- Questions about health, pregnancy, medication or therapy, political or religious views, and union activity (including coworkers discussing pay or working conditions) cannot be asked, and conversations touching them only feed the overall totals.
- Naming individual users is off by default. If a company turns it on, viewers still see generic labels until they click to reveal a name, and every reveal is logged.
Anthropic also says the reports are for understanding adoption and planning spending, “not for evaluating individual performance or making employment decisions”. Companies including Monday, Wix and Deloitte are quoted in the launch post praising the visibility it gives them.
Our take
The safeguards are real and stricter than a lot of workplace monitoring software, some of which logs every keystroke. But two things are worth being clear about. First, “don’t use this for performance reviews” is a policy, not a technical lock: once a manager sees that a team spent a chunk of its AI budget on off-topic chats, nothing stops that shaping how they treat the team. Second, a failed or “inefficient” session can simply be someone experimenting, which is how people learn these tools.
The launch also brought in-app surveys asking staff what Claude saved them time on, and pooled monthly budgets for groups, so the bigger story is companies starting to measure AI the way they measure any other expense.
What it means for Pakistani workers and freelancers
The feature is only on Claude Enterprise, the plan large companies buy, so it does not touch personal Claude accounts. But Pakistani staff at multinationals, banks and IT exporters that roll out company AI accounts should assume their employer can see how the tool is used in aggregate. The simple rule: keep personal chats, CVs and job hunting off a work account, and ask HR or IT whether attribution is switched on. Freelancers working inside a client’s AI workspace should treat it the same way.
