In partnership with
The GTM bets that shouldn't have worked, and did
One grew revenue 50x after half his team quit over the strategy. One brought in 50K signups in a single day with no paid budget. One generated 100M+ views from a stunt that took 50 hours to conceive. One asked every prospect to demo the product themselves instead of demoing it for them.
None of them followed the safe playbook. They treated GTM like an experiment, moved before they had proof, and made bets most founders would never get approved.
HubSpot for Startups documented all 6 stories in the free Bold Bets Playbook. The risks they took, why it was risky, and what it returned.
Alora knew something was off the moment Ryan asked to “grab fifteen minutes.” He was warm about it, the way good managers are when they’re about to raise something awkward. He’d been looking at the team’s engagement dashboard, he said. Her numbers had dipped. On three client calls that week, the system had registered negative sentiment and low energy, and he wanted to check in — was everything okay?
Alora sat with that for a second. The week he was describing had been a good one. She’d closed two of those three calls successfully. What she had not done was perform enthusiasm for a camera, because on Tuesday she’d had a migraine and squinted through a spreadsheet, and on Thursday she’d been concentrating hard on a client’s complicated tax question. A piece of software had watched her face, decided those expressions meant something, and reported it up the chain as a problem. Ryan, to his credit, was uneasy about the whole thing. But he was acting on the dashboard anyway, because it was there, and because someone above him was looking at it too.
Here is what neither of them knew: the technology that flagged Alora doesn’t actually work, a major government has already banned it outright, and the data it generated about her may have been collected unlawfully.
The Quiet Arrival of the Mood Reader
Emotion AI — sometimes called affect recognition or sentiment analysis — is software that watches a face, listens to a voice, and reads the words being spoken, then infers an internal state: engaged, stressed, frustrated, happy. It is no longer a lab curiosity. The market for real-time sentiment analysis passed six billion dollars in 2025 and is growing at roughly fifteen percent a year (BuildBetter, 2026), and the capability is being folded into the tools people already live inside.
Some of it arrives through the meeting platforms themselves — Zoom’s AI Companion, Microsoft Teams Premium, and a wave of “revenue intelligence” copilots like Gong and Chorus that score buyer and speaker sentiment on sales calls. Some of it comes from specialist vendors selling multimodal emotion engines that fuse voice tone and facial cues, a field whose best-known developer, Hume AI, signed a licensing deal with Google DeepMind in January 2026 (Forasoft, 2026). The pitch to employers is always some version of the same promise: finally, a window into how your people and your customers really feel.
That promise is seductive precisely because it sounds like a superpower. The trouble starts when you ask whether the window actually shows anything real.
It Doesn’t Work, and That’s Not a Close Call
The single most important fact about emotion AI is one its vendors rarely volunteer: the science underneath it is broken. In 2019, a panel of five senior scientists led by the psychologist Lisa Feldman Barrett reviewed more than a thousand studies on whether a person’s emotional state can be read from their facial movements. Their conclusion was blunt. Facial configurations are not reliable, specific signals of particular emotions across people, situations, and cultures (Barrett et al., 2019). You cannot confidently infer happiness from a smile or anger from a scowl, because that is not how human faces actually behave.
The numbers are humbling. People scowl only about thirty percent of the time they are angry, and they scowl plenty of times when they are not — when they’re confused, concentrating, or staring into the sun (Northeastern, 2019). The American Civil Liberties Union, reviewing the same body of work, warned that an entire industry was being built on a foundation the evidence simply does not support (ACLU, 2023). When Alora’s migraine squint got logged as “negative sentiment,” the system wasn’t malfunctioning. Misreading her is what these systems do, because the thing they claim to measure cannot be measured the way they claim to measure it.
Sit with the implication for a moment. A tool that is wrong most of the time is not a neutral inconvenience when it is feeding decisions about someone’s performance. It is a machine for manufacturing confident, official-looking errors about people who have no idea the errors are being recorded.
Europe Has Already Said No
Regulators reached the same conclusion, and one of them has acted decisively. As of February 2, 2025, the European Union’s AI Act prohibits the use of AI systems to infer the emotions of people in the workplace, with narrow exceptions only for medical or safety reasons (EU AI Act, Art. 5(1)(f)). This is not a high-risk-but-allowed category with paperwork attached. It is a flat ban, sitting alongside social scoring and manipulative AI on the list of practices Europe considers fundamentally incompatible with basic rights. The penalty ceiling is the Act’s highest tier: up to thirty-five million euros or seven percent of global annual turnover, whichever is greater (eMonitor, 2026).
The reasoning behind the ban is worth quoting in spirit: the Act’s own recitals cite the “limited reliability” of these systems and their potential for discriminatory and intrusive outcomes (FPF, 2026). In other words, Europe didn’t just decide emotion AI was creepy. It decided the technology was both unreliable and dangerous, and that the workplace — with its built-in power imbalance between employer and employee — was exactly the wrong place for it.
And this is already moving from principle to enforcement. In a case that should make every financial-services executive pay attention, the Hungarian data protection authority ordered a bank to stop analyzing the emotions of callers during voice calls, finding the practice posed fundamental-rights risks under European data protection law (FPF, 2026). A bank. Voice calls. Emotion analysis. Ordered to stop. If that sounds uncomfortably close to the dashboard Ryan was reading, that’s because it is.
The United States has no single equivalent ban, but it is not the open field employers sometimes assume. Illinois’s Biometric Information Privacy Act requires informed, written consent before a private entity captures the facial geometry or voiceprints that emotion engines rely on — the same statute now driving the meeting-AI litigation this newsletter covered in Edition 30. Washington’s My Health My Data Act treats many biometric and health-related inferences as protected consumer health data requiring consent, and it carries a private right of action. The legal exposure in the U.S. is simply more scattered, which makes it easier to miss and no less real.
Meet Alora and Ryan — And Walk Through the Framework
Helen Nissenbaum’s contextual integrity framework asks whether information flows in ways that fit the norms of the context where it originated. Alora’s flagged calls break those norms at every step. Here are the five parameters.
1. The Context
A client video call is a professional context with a clear, shared purpose: do the work, serve the client, represent the firm. Within that context, there are settled norms about what is being observed — what Alora says, the advice she gives, whether the client’s problem gets solved. There is no shared norm that says her face is being mined for inferred emotional states and scored. The emotion layer imports a kind of scrutiny the context was never understood to include.
2. The Actors
Alora is the subject, though she never agreed to be one in this sense. Ryan is the recipient of the inference, acting on a number he didn’t generate and can’t verify. And there is a third actor most people forget: the vendor whose model is making the actual judgment, encoding assumptions about what a face means into a score that travels upward as if it were fact. The client on the call, meanwhile, may have their sentiment scored too — another subject who never consented.
3. The Attributes
The attributes flowing here are not Alora’s words or her work product. They are her facial movements, her vocal tone, her micro-expressions — biometric signals — transformed into claims about her inner life: her engagement, her stress, her supposed frustration. These are among the most intimate attributes a person has, and they are precisely the ones the science says cannot be read accurately. The system is collecting deeply personal data in order to produce an unreliable guess.
4. The Transmission Principles
In a professional context, information about an employee is expected to flow under norms of relevance, accuracy, and notice — you observe what matters to the work, you try to get it right, and the person generally knows the terms of their own evaluation. Emotion AI violates all three. It flows intimate inferences that aren’t relevant to whether the work got done, that are known to be inaccurate, and that Alora was never meaningfully told about. The EU’s response — banning the flow outright in this context — is contextual integrity expressed as law.
5. The Information Flow
Follow the path. Alora’s face and voice, captured for the ordinary purpose of a client meeting, are fed into an emotion model, converted into a sentiment score, surfaced on a manager’s dashboard, and used to question her performance — with no notice, no consent, no accuracy, and no way for her to contest the underlying inference. Information gathered for one purpose, in one context, has been silently repurposed into something intimate, wrong, and consequential. That is the violation, and in Europe it is now also the offense.
Why Financial Services Is the Target — and the Most Exposed
It is not an accident that the cautionary tale involves a bank. Financial services is one of the most heavily marketed-to sectors for this technology, for reasons that are easy to understand and hard to defend. Firms run enormous volumes of client-facing calls, they are obsessed with conversion and retention, and they already operate dense monitoring and recordkeeping infrastructure. Bolting a sentiment layer onto calls that are being recorded anyway feels, to a vendor’s pitch deck, like a small step.
It is not a small step. A sector that already carries some of the strictest obligations around client data and fair treatment is uniquely exposed when it deploys a tool that is scientifically unreliable, freshly banned in a major market, and legally fraught at home. The Hungarian bank order is a preview, not an outlier. Any global financial institution running emotion analysis on calls that touch EU employees or customers is no longer contemplating a future risk. It is contemplating a current violation with a thirty-five-million-euro ceiling.
The competitive irony is sharp. The same firms that would never deploy a credit model they couldn’t validate are being sold an emotion model that, by the weight of scientific evidence, cannot be validated at all.
What Employers Should Actually Do
The guidance here is unusually clear because the underlying facts are unusually one-sided.
Start by finding out whether you already have it. Emotion and sentiment scoring is increasingly bundled into meeting platforms and revenue-intelligence tools as a default feature, which means many organizations are running it without a deliberate decision ever having been made. Inventory the stack, ask every vendor in writing whether their product infers emotion or sentiment from voice or facial data, and turn off what you find unless you can articulate a genuine medical or safety justification.
Resist the relabeling trick. Vendors are already renaming the same capability as “engagement,” “speaker energy,” or “conversational pace” to slip past the bans. If a metric is derived from biometric signals and used to draw conclusions about a person’s state, a new label doesn’t change what it is — and European regulators have signaled they will look through the wording to the substance.
If you operate anywhere near the EU, treat workplace emotion AI as simply off the table. There is no conformity assessment that rescues a prohibited practice; the only compliant posture is not to use it. For U.S.-only operations, assume BIPA-style consent obligations apply to any facial-geometry or voiceprint processing, and document a consent and notice approach before a single call is scored.
And remember the human being on the other end of the dashboard. Even where some narrow version of this technology might one day be lawful, a manager acting on an unreliable emotional inference is making worse decisions, not better ones. The most useful thing Ryan could have done was the oldest thing in management: ask Alora how her week was going, and believe her answer over the software’s.
Trust the Person, Not the Read-Out
Alora’s story ends quietly, the way most of these do. She explained the migraine. Ryan nodded, a little embarrassed, and the conversation moved on. But the dashboard is still running, still watching her face, still translating her concentration into a number that someone, somewhere, treats as truth. Nothing about that conversation changed the system that produced it.
That is the real lesson of emotion AI on video calls. It is not merely that the technology is invasive, though it is. It is that it is invasive in service of something that isn’t even accurate a confident machine-generated guess about the most private thing a person carries, sold as insight and acted on as fact. Europe looked at that combination and called it what it is. The science got there first.
A face is not a confession, and a voice is not a verdict. The firms that remember this — and switch the mood reader off before a regulator or a scientist makes them — will be the ones still trusted by the people they ask to turn their cameras on.
About this newsletter
Remote Work Privacy Insights is a weekly read on workplace privacy, AI governance, and the regulatory ground shifting under both. Written by Dr. Edward Halle, FIP, CIPM, CIPP/US, AIGP, CAIE, LL.M., D.B.A., Privacy & AI Governance Practitioner— author of Rethinking Workplace Privacy, Power, and Productivity in the Age of Remote Work (2025). Each edition applies Helen Nissenbaum's contextual integrity framework to the evolving intersection of workplace privacy, AI governance, and regulatory compliance.
Disclaimer: Remote Work Privacy Insights is a newsletter that looks at privacy issues in the workplace using academic ideas. It's meant to educate and is not legal advice. For advice tailored to your company, talk to a qualified privacy or employment lawyer. The opinions shared are the author's and not those of any employer.
Primary Sources Referenced
— The Science —
Barrett et al., “Emotional Expressions Reconsidered,” Psychological Science in the Public Interest (2019): https://journals.sagepub.com/doi/10.1177/1529100619832930
Northeastern University, summary of the Barrett review (2019): https://news.northeastern.edu/2019/07/19/northeastern-university-professor-says-we-cant-gauge-emotions-from-facial-expressions-alone/
ACLU, “Experts Say Emotion Recognition Lacks Scientific Foundation”: https://www.aclu.org/news/privacy-technology/experts-say-emotion-recognition-lacks-scientific
— The Law —
EU AI Act, Article 5 (Prohibited AI Practices): https://artificialintelligenceact.eu/article/5/
Future of Privacy Forum, “Red Lines Under the EU AI Act” (incl. Hungarian DPA bank case), 2026: https://fpf.org/blog/red-lines-under-eu-ai-act-unpacking-the-prohibition-of-emotion-recognition-in-the-workplace-and-education-institutions/
eMonitor, “EU AI Act Employee Monitoring Guide” (penalties & enforcement), 2026: https://www.employee-monitoring.net/compliance/eu-ai-act-employee-monitoring
Quinn Emanuel, “Initial Prohibitions Under EU AI Act Take Effect,” 2025: https://www.quinnemanuel.com/the-firm/publications/initial-prohibitions-under-eu-ai-act-take-effect/
— The Market —
Forasoft, “Emotion Recognition in Video Conferencing” (vendor landscape, Hume/DeepMind), 2026: https://www.forasoft.com/blog/article/emotion-recognition-video-conferencing
BuildBetter, “Best AI Tools for Real-Time Sentiment Analysis” (market size), 2026: https://blog.buildbetter.ai/best-ai-tools-real-time-sentiment-analysis-2026/