On 14 May 2026, Italy's data protection authority, the Garante, issued provision no. 342 (doc-web 10255494) — a formal warning against Myndoor S.r.l., a Milan-area start-up whose AI plug-in for Slack and Microsoft Teams performs semantic analysis of workplace chat to estimate employees' psychological stress. The authority opened the file in June 2025 after press reports suggested the tool was being used inside public administrations, and made its decision public in a press release on 28 May 2026. No fine was imposed; the Garante issued a warning under GDPR Article 58(2)(a) and directed Myndoor to ensure that no stress data — "including via the aggregated report" — reaches employers about workers using the plug-in.
The decision matters because it is one of the first European enforcement actions to cite the GDPR and the EU AI Act in the same breath. It also illustrates a recurring problem with how data authorities are reaching for the AI Act: invoking its most dramatic prohibition even where the prohibition's own text and the Commission's guidance say it does not apply.
The strong part of the ruling
Start with the steelman, because the Garante's core privacy reasoning is sound. Myndoor positions itself as sole controller, lets employees voluntarily activate the tool, and only shares aggregated reports with employers once at least ten weekly active users are present. The Garante acknowledged these safeguards but concluded it "cannot absolutely exclude" that an employer could re-identify individuals from aggregates, given how much organizational context — team size, department structure, who was visibly stressed that week — an employer already possesses. That is a real risk, not a hypothetical. A ten-person threshold inside a six-person team is no anonymity at all.
The Garante also leaned on Italian labor law and GDPR Article 9. Inferred emotional and stress states sit close to health data, a special category. And under the Italian Workers' Statute, employers have long been barred from monitoring the personality and opinions of their staff. "Voluntary" activation is doing heavy lifting here: in a workplace, consent given under an employer's gaze is rarely free. On these grounds — Articles 5, 6, 9, 25, and 88 of the GDPR, plus Articles 2-ter and 113 of the Italian Code — the warning is a proportionate, evidence-based intervention. It blocks a specific data flow without banning the product or fining a small company out of existence.
The part that overreaches
The weaker move is the Garante's reliance on Article 5(1)(f) of the AI Act, which prohibits AI systems that infer emotions in the workplace. The provision has been in force since 2 February 2025 and carries penalties of up to €35 million or 7% of global turnover. It sounds tailor-made for Myndoor. It isn't.
The prohibition applies only to emotion inference drawn from biometric data. The European Commission's February 2025 guidelines on prohibited practices are explicit on this point. As the Future of Privacy Forum summarizes, the guidance states that "the prohibition in Article 5(1)(f) refers only to emotions deriving strictly from biometric data," and — directly on point — that "inferring emotions from a written text does not fall within the scope of the prohibition." Myndoor analyzes written text. It reads no faces, voices, or physiological signals. By the Commission's own reading, it is outside the ban.
This is not a pedantic distinction. The AI Act drew a deliberate line: biometric emotion-reading — facial-expression and voice-stress systems that operate on the body itself — is a red line, while text-based sentiment analysis is permitted but pushed into the high-risk tier when used for employment decisions, where it triggers documentation, oversight, and conformity obligations. The Garante's decision blurs that line, citing a near-absolute prohibition against conduct the legislature chose to regulate, not forbid.
Why the distinction is worth defending
There is an instinct to treat the biometric-versus-text boundary as a loophole. It isn't. Text content is something the employee writes and can see; a face scanned for micro-expressions is not. The legislature's choice to treat them differently was a calibrated judgment about intrusiveness, and regulators should not erase it by enforcement.
The practical cost of blurring it is legal uncertainty for exactly the kind of product that deserves room to exist: an opt-in, individual-facing wellbeing tool that, in the Garante's own telling, never handed raw data to employers. As one legal analysis of the decision notes, the warning left controllership, the anonymization threshold, and concrete corrective measures unresolved — guidance builders need, replaced by a prohibition invoked in error.
The Garante reached the right outcome on the wrong statute. The aggregation-to-employer flow was a genuine GDPR and labor-law problem, and stopping it is defensible. But proportionate regulation means citing the rule that actually applies. Stretching the AI Act's headline ban to cover a text tool the Commission expressly excluded substitutes a 7%-of-turnover sledgehammer for the scalpel the GDPR already provided — and tells every European builder that the categories in the AI Act mean whatever an enforcer later decides they mean.