Part 2: AI and the People You Already Employ

Part 2 of AI and HR covers the AI tools that evaluate, schedule and monitor current staff, the EU ban on emotion recognition at work, two enforcement cases (Amazon's warehouse scanners and a delivery platform's algorithm) and what employers must tell employees and their representatives. Researched and hosted by Mia and Adam, two AI voices.

Tue, 03 Nov 2026 · 16 min

Made for you, by AI, from the sources below.

Listen to it in AI and HR

Transcript

The script the two voices were given, with the sound cues taken out. A voice may say a word a little differently, so the audio is the episode.

Mia: Welcome to Walking Points. One flag in an Amazon warehouse scanner went off whenever a worker scanned an item less than 1.25 seconds after the previous one. The French regulator nicknamed it the stow machine gun indicator, and it ended up before the highest administrative court in France.

Adam: That's a very specific number to be judged on. So did the court say it was unlawful?

Mia: No, the court let that indicator stand. The fine around it was cut by more than half, though, and the reason why is the heart of this episode.

Mia: This episode is about AI and the people you already employ. This part was made in October 2026. It covers the tools that score people's performance, schedule their shifts, count their keystrokes and, in some products, claim to read their mood. We'll walk through four steps. First, what these tools actually do, and why a performance score is a different thing from a hiring score. Second, the one outright ban in this area, on emotion recognition at work. Third, two enforcement cases, the Amazon one and an Italian one about a delivery app's algorithm. And fourth, what staff, works councils and unions are entitled to hear before a system helps decide about them.

Adam: And this builds on last Tuesday, right?

Mia: Yes. Last Tuesday we covered AI in hiring: how screening tools rank candidates, the bias cases, the New York audit law, and the EU rules that classify hiring systems as high-risk. This episode is about what happens after the person is hired.

Adam: So why is that a different case? A score is a score.

Mia: Think about the position of the person. A candidate meets an AI system once, from the outside, and can walk away. An employee lives inside one. The data is richer, because it's every day of someone's working life. The person can't easily leave. And the decision is continuous, not a one-off.

Adam: Okay, but what do these tools actually do? Give me the families.

Mia: Five of them. Performance tools turn work data into scores: sales calls transcribed and rated, tickets closed per hour, scanner activity. People analytics platforms combine pay, tenure, absence and survey data to predict things like who is likely to leave. Scheduling and task allocation tools decide who works which shift and which job comes next, which for delivery riders is most of what a manager used to do. Monitoring tools record logins, application use, location, sometimes webcam or voice. And a smaller, controversial family claims to infer emotions, like stress or engagement, from faces, voices or typing.

Adam: That's a wide net. Is there one word for all of it?

Mia: Algorithmic management. A study for the European Commission documented it in recruitment, scheduling, monitoring, evaluation, compensation and even dismissal. The European Parliament's employment committee defined it in a draft report as using automated systems, including AI, to monitor, assess or make decisions affecting workers. So the legal question is rarely whether something is AI. It's whether a system watches, rates or directs a person who depends on this employer for a living.

Adam: Fair. But does watching people more actually get more out of them? That's the sales pitch.

Mia: The research is not flattering to that pitch. Large meta-analyses of electronic performance monitoring find modest drops in job satisfaction, slight increases in stress and negligible gains in performance. With a caveat: those are averages across many workplaces, and it depends on whether people understand the monitoring and whether it's used to help or to punish. Still, more measurement meaning more output is a weaker claim than vendors suggest.

Adam: Hm. So the average says it mostly doesn't pay off, and the rulebook says what, exactly?

Mia: Under the EU AI Act, the Annex III list of high-risk uses covers employment, and that reaches beyond hiring to promotion, termination, task allocation and monitoring performance. A company using such a system is called the deployer. But those deployer duties were postponed, by what's called the digital omnibus amendment, to 2 December 2027. So the high-risk duties for workplace tools are not in force yet.

Adam: So employers have a free pass until then?

Mia: No, and that's the point. Two things are in force right now. The ban on emotion recognition at work, and the GDPR, which has applied since 2018 and does most of the real enforcement. Its Article 22, the right not to be subject to a solely automated decision with significant effect, applies to firing or demoting by algorithm just as it does to hiring.

Adam: Let's take the ban first, then.

Mia: Since 2 February 2025, Article 5(1)(f) of the AI Act prohibits using AI to infer the emotions of a person in the workplace or in education, except for medical or safety reasons. The fines are the Act's highest: up to 35 million euros or 7 percent of worldwide annual turnover, whichever is higher. And since this is law, one light note: this is a briefing, not legal advice.

Adam: Define emotion recognition for me. Because a lot of software says it understands people.

Mia: The Act's definition is a system that identifies or infers emotions or intentions from biometric data, meaning a face, a voice, a gait. The Commission's guidelines list what counts as an emotion, and it's longer than you'd guess: happiness, anger, anxiety, shame, and, importantly for workplaces, attentiveness, focus, boredom and impatience.

Adam: Wait, focus counts? So a tool that scores whether someone looks engaged in a meeting is covered.

Mia: Exactly. The guidelines' own banned examples are ones HR will recognise: webcam and voice tools tracking employees' emotions in a call centre, monitoring facial expressions in meetings to rate engagement, and using emotion recognition on candidates or on staff during a probationary period.

Adam: And the exceptions? There must be some, or a pilot's fatigue alarm would be illegal.

Mia: Right, and they're narrow. Safety can cover something like detecting fatigue in a driver or pilot, in high-risk sectors, and only where nothing less invasive would do. The medical exception does not stretch to general wellbeing. A system built to detect burnout or depression at work would not fall under it, and stays prohibited.

Adam: Mia, that sounds like it catches half the wellbeing dashboards on the market. Is that too broad? Surely a lot of this is harmless.

Mia: It's a fair challenge, and the ban has real edges. Detecting a physical state like pain or fatigue isn't inferring an emotion. Noting that someone is smiling isn't covered, as long as the system doesn't infer a feeling from it. And analysis of written text, like sentiment analysis of survey comments, isn't biometric, so it's outside this ban. I'll concede that commentators flag two grey zones: stress, which sits between a physical state and an emotion, and vendors relabelling an inference as a mere detection.

Adam: So what does an HR team do with that?

Mia: Ask the vendor one direct question about any engagement or wellbeing feature: does it infer emotional states from the faces or voices of our staff? If yes, it's likely prohibited in the EU whatever the marketing says, and the user, not only the vendor, can be liable.

Adam: Okay. Now, you promised me what happened to that stow machine gun flag.

Mia: On 27 December 2023, France's data protection authority, the CNIL, fined Amazon France Logistique 32 million euros for excessively intrusive monitoring of warehouse staff. Every worker carries a scanner, and every scan feeds indicators on quality, productivity and inactivity. The CNIL objected to three: that quick-scan flag, an idle-time measure, and a latency measure around breaks. It also said keeping all this detailed data for 31 days wasn't justified.

Adam: And Amazon appealed.

Mia: Yes, and the outcome is the lesson. On 23 December 2025 the Conseil d'État cut the fine from 32 million to 15 million euros. The court accepted that all three contested indicators were lawful as designed. The quick-scan flag only triggers on a very short gap and serves to catch scanning errors, not to demand speed. The idle-time indicator only triggers after more than ten consecutive minutes of inactivity. And the latency indicator's scope around breaks was limited.

Adam: Hold on. So a court said second-by-second scanner monitoring is fine? That sounds like Amazon won.

Mia: Partly, and I'd be careful about the word won. The court upheld the finding that keeping every indicator for that long, without showing it was necessary, was unlawful. And violations nobody contested stayed in place: inadequate information to workers, weak video-surveillance notices and weak security over who could access the monitoring software.

Adam: So the metrics survived, but the paperwork didn't.

Mia: Right, the dull, checkable, HR-owned parts failed. Monitoring isn't banned, and a court has now said some very granular productivity measurement can be proportionate. The test the case teaches is to ask, for every metric, what it's for, how coarse it can be, how long the data is kept and who may see it. Which brings us to the second case, and to scheduling.

Adam: Where an algorithm decides who gets the next job.

Mia: Yes. Italy's data protection authority, the Garante, fined Foodinho, the delivery company owned by Glovo, 5 million euros, after an earlier fine of 2.6 million euros. The regulator found the company hadn't explained clearly to riders how the system assigned orders and shifts, and hadn't made sure the automated assessment of rider performance was accurate. It ordered changes to the algorithms managing and controlling riders.

Adam: What's the takeaway for a company that isn't a delivery app?

Mia: Two ideas. Transparency about how work is allocated is a legal duty under GDPR, not a courtesy. And a rating that decides who gets work has to be accurate, and someone has to be able to check it. Those same ideas are now written into EU labour law for platforms, and the open question is what about everyone who isn't a platform.

Adam: Start with the platforms, then.

Mia: The Platform Work Directive has a transposition deadline of 2 December 2026, less than two months away. It's the first EU text that regulates algorithmic management in employment-like relationships directly. It bans platforms from using automated systems to process a worker's emotional or psychological state, private conversations, or data collected when the worker isn't working. It requires human oversight by trained staff who can override automated decisions. And it requires platforms to inform workers about their monitoring and decision systems and review their impact at least every two years.

Adam: Good. But it doesn't cover a normal employer?

Mia: It doesn't, and that's the gap. The European trade union confederation, ETUC, argues the AI Act gives workers no specific rights and wants a dedicated directive on algorithmic management. The Parliament's employment committee proposed one in its draft report: information duties, consulting workers before systems are introduced or updated, human oversight and explanations for decisions. I haven't found a Commission proposal yet, so what an employer must do beyond the AI Act and GDPR comes from national law.

Adam: Which means works councils.

Mia: In Germany, a works council has a right of co-determination under section 87(1) number 6 of the Works Constitution Act over technical systems that are suitable for monitoring behaviour or performance. Suitable, not used for it. That's why a performance tool nearly always needs the council's agreement before it goes live. The Hamburg labour court showed the edge: it rejected a works council's request for an injunction against staff using ChatGPT through their own private accounts, because the employer didn't provide the tool and couldn't monitor those accounts.

Adam: So an employer-provided tool that logs usage is a different thing.

Mia: Very different, legally. And other countries vary, so a team should know which route applies where it operates.

Adam: Okay, so what does an employer actually have to tell staff? In plain order.

Mia: First, GDPR transparency, now: what data is collected, why, for how long, who sees it, and for automated decisions, the logic involved and its consequences. The Amazon case shows weak notices counted as violations even when the metrics survived. Second, under Article 26(7) of the AI Act, once Annex III applies on that December date, an employer must inform workers' representatives and the affected workers before putting a high-risk system into service. That concerns the fact of exposure, not the technical details.

Adam: And beyond that?

Mia: Alongside it sit duties to assign trained human oversight and keep the system's logs for at least six months, and an affected person can ask for a clear explanation of the system's role. Third is the platform directive we just covered. Fourth, where emotion recognition is lawful, say a medical or safety application, people exposed to it must be informed.

Adam: So for someone who works in or around HR, or just buys software at a company, what's the order of business? The dates are scattered.

Mia: The date that matters isn't the far one. The ban is live, GDPR is live, and the platform directive lands in December. Do an audit first: list every system that scores, schedules, ranks or watches staff, including features inside tools already bought for something else, like a meeting platform or a contact-centre suite. For each, ask the vendor what it infers and what it retains and for how long, and check whether the works council must agree to it. Training programmes can wait.

Adam: And if you're just a curious employee?

Mia: Then the same logic works from your side. You're the person being scored, and you're entitled to know what's collected, why and for how long. Here's one thing to try this week: pick one tool your company already uses that has an engagement or wellbeing feature, and ask whoever owns it whether it infers emotions from staff's faces or voices. The reason is that the ban is live now, and a vague answer tells you the audit hasn't happened.

Adam: And it comes back to the scanner. The metric survived, but the notices and the retention didn't.

Mia: That is what the Amazon case shows. Adam, you asked at the start whether a score is just a score, and here the answer was that the metrics survived, and what failed was the dull, checkable parts around them: the retention, the notices and the security.

Adam: So the score itself can stay. What gets judged is everything around it, and that part is HR's to get right.

Mia: That was part two of three. Part three is about HR's own work with AI, from onboarding and learning to service chatbots and policy drafting.

Heard something wrong? Tell us