Your boss is already an algorithm: AI starts reading what you write at the office

THE THESIS. Since Frederick Taylor and his stopwatch, algorithmic management has mostly measured the body: steps, seconds and clicks. What is arriving now is different. AI can read what a knowledge worker writes and judge whether it "adds value." Our thesis has two parts. First, Claude's reports, as designed, are not bossware (employee-surveillance software). They are a cost-telemetry tool with reasonable safeguards. Second, those safeguards are configuration settings and usage guidelines, not legal obligations. The line between measuring adoption and watching people will be drawn by each company's governance, by worker representatives and by regulators. In the short term the risk is real. In the long term, transparent and auditable algorithmic management could be fairer than the arbitrary boss of old. That only happens if we demand it now.

WHAT THE REPORT ACTUALLY DOES. Mashable summed it up in a headline: Claude's new reports "alert bosses when workers are off topic." The fine print qualifies that headline. According to Anthropic's guide and Business Insider's coverage, smart reports are a Claude Enterprise analytics feature. It was in beta from September 10 and has been generally available since October 28, 2026. An admin picks a team and a window of up to 28 days. They also pick which products to include: chat, Claude Code or Cowork. Claude then reads a sample of transcripts and groups them into "workstreams." For each workstream it reports spend, types of deliverables, cost per session and friction points. It can also flag sessions that "produced no work value" or were "personal or off-topic." The feature is off by default and only an organization owner can turn it on. Names are hidden by default. A second toggle, "allow attribution to individual users," shows the email of whoever started each session. Even then, a name appears only if the viewer explicitly reveals it, and every reveal is logged. Personal sessions show up only as an aggregate count and cost, with no summaries and no names. An always-on filter covers sensitive topics: health, race, legal matters, union activity and political or religious views. Those sessions count only in overall statistics. Each organization gets 10 free reports a month. Anthropic's guide says the reports "aren't designed and should not be used for evaluating individual performance or making employment decisions." There are two gaps. Anthropic has not published the error rate of its off-topic classifier. Nor has it explained how the sample is chosen, and on a small team a thin sample can skew results.

FROM STOPWATCH TO CHAT LOG. To see why this launch caused such a stir, it helps to trace the lineage. In 1911, Taylor published "The Principles of Scientific Management." His method was to break each task into motions and time them. A century later, Amazon's warehouses brought that logic to handheld scanners. In France, the CNIL (the data protection authority) fined Amazon France Logistique €32 million for "excessive" monitoring. The regulator cited, for example, alerts triggered after 10 minutes of scanner inactivity. France's Council of State upheld the sanction in December 2025, but cut it to €15 million, according to trade press reports. The next leap came with delivery riders. On platforms like Glovo and Uber, the algorithm assigns orders, rates workers and deactivates accounts. In Italy, the data protection authority fined Glovo's subsidiary Foodinho €5 million in 2024. Among the reasons: tracking riders' locations while off duty and not letting them contest automated deactivations. Then came the pandemic, and with it bossware: screenshots, keystroke logging and active-time tracking. A Chartered Management Institute survey published in 2025 found that a third of UK employers use such software. What changes now is the object of measurement. Before, clicks were counted. Now a language model can read what you asked for, why, and whether it was "work."

THREE THINGS THAT ARE NOT THE SAME. The debate gets muddy because it mixes three different uses. The first is usage telemetry: knowing how much is spent on AI, on what, and with what security risks. That is legitimate and necessary. A company paying millions for licenses has a right to know whether they pay off. The second is performance evaluation: judging a specific person by their productivity. That is legitimate but high-risk, and it requires known criteria, human oversight and a right to respond. The third is behavioral surveillance: spotting who is "wasting time," who talks about what, and in what mood. That is where the technology collides with fundamental rights. Our reading: Anthropic's product is designed for the first use. But the individual-attribution toggle allows a slide toward the second, and the off-topic counter invites thinking about the third. The boundary is not in the code. It lies in what each organization decides.

WHAT THE LAW SAYS. In Europe the framework is dense but has gaps. The EU AI Act has banned emotion recognition in the workplace since February 2025. However, the law defines it based on biometric data, such as faces or voices. Classifying the text of a chat probably falls outside that ban. The same law treats AI used to monitor and evaluate workers' performance and behavior as "high-risk." Those obligations were due in August 2026, though the Commission's Digital Omnibus package proposes delaying them to December 2027. One detail matters a lot. Under Article 25 of the regulation, a user who deploys a system for a high-risk purpose the maker did not intend can take on the obligations of a provider. A company using these reports to evaluate people would carry that burden. The Platform Work Directive, which member states must transpose by December 2, 2026, goes further. It bans processing a worker's emotional or psychological state and their private conversations. But it only protects platform workers. Today the strictest rule protects riders, not office staff. The GDPR (the EU's data protection regulation) lets each country set specific workplace rules under Article 88. Spain did so in Articles 87 to 91 of its data protection law (LOPDGDD), which require informing workers before monitoring digital devices. Spain's 2021 Rider Law added Article 64.4.d to the Workers' Statute. It gives works councils the right to know the parameters and rules of algorithms that affect working conditions, and it is not limited to platforms. On top of that, Royal Decree 723/2026, in force since October 5, requires telling each worker in writing about automated systems that decide on their hours, tasks, pay, promotion or dismissal. A generic clause about "use of AI" is not enough. Our reading: in Spain, using these reports to evaluate someone without informing the works council and the worker is legally risky. In the United States the map is thinner: New York has required written notice of electronic monitoring since 2022, and only a few other states have similar rules.

OUR READING AND IMPLICATIONS. We don't buy the doom headline. Anthropic did several things right: off by default, names hidden, reveals logged, sensitive topics shielded and an explicit warning against using it to evaluate people. That is more than most pandemic-era bossware offered. But we don't buy complacency either. A usage guideline is not a technical restriction. Context matters too: in a "no hire, no fire" labor market, a report that flags "no-value" sessions is a tempting excuse for cuts. Three practical implications. For companies: separate cost telemetry from people evaluation in writing, inform worker representatives, and do not enable individual attribution without a data protection impact assessment. For workers: a corporate AI account is not a private space, and the prudent move is to treat it like work email. For regulators: a ban on reading emotions and private conversations should not depend on whether you work for a platform or in an office. In the long run, our view remains optimistic. If AI takes over administrative and repetitive work, measuring who "chats off-topic" will lose its point. What will matter is judgment, creativity and the work each person is passionate about. And algorithmic management with public, auditable and contestable rules could be fairer than human favoritism. Getting there means setting the basic rule today: AI may measure the work, but it should not read the person.