Anthropic's new 'smart reports' let managers see when staff use Claude off-topic, blurring analytics and surveillance

Anthropic has launched "smart reports" for Claude Enterprise, Mashable reports. They are available now on paid plans and must be switched on in the organisation's settings. A manager picks a team, a period of up to the last 28 days, and which products to include: chat, Claude Code, Claude Cowork or the unified chat-and-Cowork experience. The system then reads the team's conversation transcripts, groups them into workstreams and output types (for example "vibe code" or technical documents), and calculates spend.

The result is a set of interactive charts. Managers can drill into individual sessions and rank or filter them by cost, product, date or output type. They can save custom questions about usage and cost and reuse them. The reports also make recommendations, such as new rules to add to a code repository. Admins with billing permissions can raise a spend limit directly from the report if it is "holding someone back," according to Anthropic. Organisations get 10 free reports a month and pay for more.

The most sensitive feature is what Anthropic calls "inefficiencies." These include sessions that produced no usable output and sessions classified as "personal or off-topic." Mashable puts this in context. One recent survey found that one in four US managers use AI for layoff decisions "often or all the time." In July, former Meta employees alleged that the company used poorly tested AI tools to assess productivity and track token use before a mass layoff. Meta denied this and said workforce decisions were made by people. The launch also follows Anthropic's IPO filing, in which it said existing customers had increased their spending by 900% in a year. Mashable reports that the company still posted an $80bn loss in the first half of 2026.

**Our reading.** On its face this is a governance tool. It speaks directly to the gap we have described between deploying agents and controlling them: companies need to know what they spend, what gets produced and where workflows get stuck. In that sense it is mature, useful software. The problem is the framing. Grouping usage by team to optimise cost is one thing. Labelling individual conversations "off-topic" is something else. That turns a productivity assistant into a supervision tool. Employees will have to assume their chat history can be read and judged, and that pushes them toward hiding things rather than exploring. Yet trial and error is exactly how people get the most out of AI.

The commercial incentive is obvious too. A report that suggests raising spend limits suits a company heading to the stock market with large losses. That does not make the tool bad. It does mean its findings should be read with that interest in mind.

Over the long run, AI should make work more autonomous, not more watched. For that to happen, these tools need clear rules: aggregated data by default, transparency toward employees, and an explicit ban on using them for individual disciplinary or dismissal decisions. In jurisdictions with stricter worker-privacy rules, we expect the "off-topic" label to come under scrutiny. The company that best balances visibility with trust will win in enterprise, not the one that monitors the most.