TL;DR — Anthropic shipped Smart Reports for Claude Enterprise in 2026. Admins across 28-day windows can read session summaries of employee conversations, drill into workstreams, and rank people by token cost across a 28-day window. Anthropic built 8 protected topic categories, hides names by default, and logs every reveal. Those are reasonable product controls. They are also a poor management strategy. The feature measures activity, not outcomes. I have watched marketing teams chase impressions for 15 years and call it pipeline; this is the same mistake with a bigger invoice. I think the companies that win with AI through 2027 will define what good output looks like and let people get there without a camera in the room.
The feature is real, and so is the temptation
Anthropic’s documentation on Smart Reports is worth reading closely. An admin selects a team, a time window of up to 28 days, a surface such as Claude or Claude Code, and a custom question. Claude samples transcripts, groups activity into workstreams, estimates list-price cost, and returns a structured report. Drill down and the tool serves session-level summaries: date, product, output type, cost. Not just a usage count. Names stay hidden by default, but an admin can reveal them on demand, and every reveal is logged automatically.
That logging detail tells you Anthropic knows what it built. The company also blocked session-level drilldowns on 8 protected categories: personal life, health and medical matters, political and religious views, protected characteristics, union activity, harassment and whistleblower complaints, privileged legal content, and sentiment about a named person. Those protections cannot be disabled. Admins cannot submit custom questions targeting a restricted topic. Because the classification is automated, Anthropic admits there is no guaranteed detection rate.
Credit where it is due: the guardrails are real. The problem is that they are probabilistic, and the governance conversation Anthropic leaves to the customer is the one most customers are not prepared to have. Gartner’s 2025 AI governance survey found that 61% of organizations still had no formal AI usage policy, while 43% had deployed generative tools to 500 or more employees. That gap is where the trouble starts.
The metric is wrong before the ethics are
Strip the privacy argument out for a second and look at what Smart Reports measures. It measures usage. It measures token spend. It measures whether people talk to Claude about the right things. None of those numbers tell you whether the work got better.
This is the oldest mistake in marketing wearing a new interface. Impressions are not pipeline. Sessions are not shipped work. A 40-person team can burn $40,000 in tokens across 6,000 sessions summarizing meetings that should never have happened and produce a beautiful 2,000-word report that proves it. The dashboard shows engagement at 92%. The business shows nothing. LinkedIn’s 2024 workplace data put the average cost of a wasted enterprise meeting at $1,200 in salary time; a 12-person leadership team holding 3 such meetings a week burns $187,200 a year before AI touches it.
The question worth asking about every campaign is the question worth asking about every AI session: what changed because this happened? Did a rep close a deal 3 days faster? Did an analyst catch an error before it reached a $2 million client? Did a writer ship a piece that generated 5 qualified conversations? Those are outcomes. Smart Reports cannot see them, because the outcome lives in the CRM, the repository, or the revenue line, and the report only knows what happened inside the chat window.
Cost visibility is genuinely useful. Spend is real, and enterprise AI budgets are under scrutiny in every Q4 planning cycle from 2026 onward. But cost without context invites the wrong conversation. An admin who sees a $4,000 workstream and no matching output assumes waste. Often the opposite is true: that session replaced 40 hours of manual research at $150 an hour, or $6,000 in analyst time. Measuring the invoice without measuring the return is how finance kills the programs that were working.
Surveillance changes behavior, and not in the direction you want
The cultural problem is bigger than the legal one. People who know their AI conversations are being read do not stop using AI. They get careful. They stop asking the dumb question that would have taught them something. They stop pasting the messy first draft. They stop using the tool for the uncertain, exploratory work where AI is strongest, the work that looks inefficient on a dashboard and produces the insight nobody else has.
Content teams have run through this dynamic. The people who got good at AI were the ones who felt safe being bad at it for a month. The first 30 days of any real skill are clumsy. If every clumsy session becomes a line item in a report to a manager, the learning stops inside 2 weeks. I end up with a workforce that uses AI only for safe, sanctioned, low-value tasks, drafting routine emails and formatting documents, while the high-value thinking goes back to being done alone, unobserved, and unimproved. Microsoft’s 2024 Work Trend Index found that 78% of AI users bring their own tools to work, up from 67% in 2023. Shadow AI is not a discipline problem. It is a surveillance response.
There is a second-order effect that is worse. Surveillance signals distrust. A company that reads your chats tells you, in the most concrete way available, that it expects you to misuse the tool. People reciprocate that expectation. They withhold. They perform. They game the metric. The report says usage is up 40% month over month. The business says the 12 best people are leaving. That is an expensive way to buy a dashboard.
The governance gap is the real story
Anthropic’s controls reflect product decisions. Whether a rollout is lawful depends on local employment law, data-protection rules, works-council obligations, and the contracts a company signed with its own people. Anthropic’s documentation says technical access is not legal permission. Consent, notice, proportionality, purpose limitation, and data minimization sit with the employer, not the model vendor. Some jurisdictions have already moved to restrict how cloud providers handle sensitive workplace data. A product toggle does not satisfy a regulator.
This is where most enterprises stumble. The feature ships turned off. Somebody turns it on because it is there and the license covers it. Nobody checks whether the employee handbook discloses session-level monitoring, whether the works council was consulted, whether the retention policy matches what the tool stores, or whether the local regulator treats AI chat logs as employee monitoring under a stricter regime than the vendor’s terms assume. The vendor’s compliance page is not legal review. A lawyer’s sign-off is legal review. Under GDPR alone, unlawful workplace monitoring can draw fines of €20 million or 4% of global annual revenue, whichever is higher; California’s CCPA adds statutory damages of $2,500 to $7,500 per violation.
The reveal logging is good design. It creates accountability. It does not create consent. Those are different things, and conflating them is how a company ends up explaining itself to a regulator 6 months later.
What to do instead
The strategic move is not to ban the feature. It is to change what you ask it to answer.
I start with outcomes. Before turning Smart Reports on for a team, I write down the 3 business results AI usage should produce this quarter. For a sales team, that might be faster proposal turnaround or better account research. For a content team, it might be fewer review cycles or higher engagement quality. Then I measure those. Use the usage report only to find friction, workstreams where people are burning time without producing anything, and Fix the workflow, not the person.
I treat session summaries as a coaching input, not an audit trail. The useful signal is pattern. A team asking the same question 30 times in 28 days means the knowledge base is failing them. A workstream with $4,000 in cost and 2 outputs means the prompt or the process is wrong. Those are systems problems. Surveillance framed as performance management turns systems problems into personnel problems, which is how organizations lose good people over bad tooling.
I am explicit with the workforce. I tell people what is collected, who can see it, what is protected, and what is never used against them. I publish the retention window. I name the reviewer. I make the reveal log auditable by someone outside the management chain. Trust is cheaper to build than to rebuild. The cost of getting this wrong is not a fine. It is the slow erosion of the one asset that makes AI adoption work: people willing to think out loud.
I measure signal, not volume. The dashboard will always count sessions. My job is to know which sessions mattered. I pick 5 outcome metrics that connect AI usage to revenue, cycle time, or quality. I review them monthly. I kill the workstreams that produce cost without output. I protect the ones that produce output without looking busy. That second category is where the advantage lives, and it is the first thing a surveillance culture destroys.
What I actually think
Anthropic is not the villain. The company shipped a capable tool with genuine protections and disclosed the limits in public documentation. The villain is the management reflex that reaches for a dashboard whenever it feels uncertain, and I have watched it take over every time a new metric becomes available. Marketing leaders did this with vanity metrics from 2010 to 2025. HR and operations leaders are doing it with AI usage now, across 3,000-seat Claude Enterprise deployments. The reflex says: if it can be measured, it can be managed. The reality says the things worth measuring in knowledge work, judgment, initiative, and willingness to experiment, do not show up in a session summary. They show up 6 months later in the quality of the work and the retention of the people doing it. A tool that makes people afraid to think out loud is not an oversight tool. It is a tax on the exact capability the company hired for. Salesforce learned this in 2019, when Einstein Activity Capture monitored rep email across 40,000 seats and the rollout stalled for 18 months.
The bottom line
Anthropic built a capable monitoring product and shipped it with real guardrails. That deserves respect. It is still not a management strategy. The companies that turn AI into an advantage will use these reports to understand their systems, not to supervise their people. They will measure what changed in the business, not what happened in the chat window. They will accept that the messy, exploratory, sometimes expensive work is where the learning happens, and they will defend the conditions that make it possible.
The alternative is a workplace where everyone is watched and nobody is thinking. That is not a productivity problem. It is a leadership problem, and no dashboard will fix it.
