Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI-Generated Avatars Are Turning Meetings Into Training Data
Transcript
- Lucas: So there's this mid-size SaaS company — about 400 employees — that decided to record every single internal team meeting for six months. Stand-ups, sprint retros, one-on-ones, even the all-hands. That's about 2,000 hours of video and transcript data. Luna: Two thousand hours. And they did what with it? Lucas: They fed it into an AI model to build their own internal productivity assistant. The idea was that the AI would learn the company's specific language — project names, team acronyms, the particular way their engineering manager says 'we need to unblock this' — and then surface summaries, decisions, and action items for every meeting going forward. Luna: So basically they turned their employees' conversations into training data without necessarily asking for permission. Lucas: Exactly. And that's the part that's starting to get attention. The tool worked — it now generates pretty accurate summaries and flags decisions that people miss. But the employees were never explicitly asked if they were okay with their voices and faces being used to train an internal model. Luna: I mean, was it in the employee handbook somewhere? Buried in the fine print of the IT policy? Lucas: It was in the standard recording consent clause that most companies have — the one that says meetings may be recorded for quality purposes. But 'quality purposes' originally meant training new hires, not feeding a machine learning model. And that's the gap. The technology moved faster than the policy language. Luna: Right, so the consent was technically there but it wasn't informed consent. Nobody signed up for their daily stand-up to become part of an AI training set. Lucas: And that's the real tension here. On one hand, the output is genuinely useful. The assistant now catches action items that get mentioned in passing and never written down. It can tell a manager which topics keep coming up without resolution. It basically gives a company a second brain for its meeting data. Luna: But on the other hand, you've got employees whose every hesitation, every awkward pause, every offhand comment becomes part of a permanent record that's being analyzed by an algorithm. Lucas: Yeah. And the company in question didn't do anything malicious. They're not selling the data. They're not using it for performance reviews — at least not yet. But they also didn't give employees a way to opt out. If you're in that meeting, you're in the training set. Luna: And that's where this gets sticky. Because once the model is trained on that data, you can't really untrain it. Even if they later let people opt out, the model already learned from those conversations. Lucas: Right. So the question becomes: should companies be building custom AI models on internal meeting data at all? And if so, what are the guardrails? Luna: Well, what does the law say? Is there any regulation that covers this? Lucas: It depends where you are. In the EU, GDPR has pretty strict rules about consent and purpose limitation. Using meeting recordings for AI training probably counts as a new purpose. In California, the CCPA gives employees some rights over their data, but it's not airtight. And in most other US states, there's basically no specific regulation covering this use case. Luna: So it's the Wild West. Companies are building these tools because they can, not because there's a clear legal framework. Lucas: Exactly. And the vendors who sell meeting recording and analysis platforms are starting to offer AI training as an add-on feature. They'll say things like 'train a custom model on your company's unique communication patterns.' It sounds great in a sales demo. But nobody on the sales call asks what the employees think. Luna: So what should an employee do if they're worried about this? Can they actually find out if their company is doing it? Lucas: First step: check the company's data privacy policy. Look for language about 'machine learning' or 'AI training' or 'model development' in the context of meeting recordings. If it's vague, ask your HR or IT department directly. Some companies are starting to create opt-out mechanisms, but you have to know to ask. Luna: And if you're a manager or a team lead considering this kind of tool? What's the responsible way to do it? Lucas: Transparency first. Tell the team exactly what you're planning to do with the recordings. Explain what the AI will be trained on and what it won't be used for. Give people a real opt-out — not just a 'you can leave the meeting' option, because that defeats the purpose of having the meeting. Maybe allow audio-only participation for those who opt out, or let them contribute via text chat. Luna: And then after training, give people access to the model's outputs so they can see what it's capturing about their own contributions. Lucas: That's a great point. If the tool is supposed to help everyone, then everyone should be able to see what it's saying about them. Some companies are doing that — they have dashboards where employees can review the AI's summaries of their meetings and flag inaccuracies. Luna: But that still doesn't address the training data issue. Once it's in the model, it's in there. Lucas: Yeah. And that's why some privacy advocates argue that companies should only use synthetic or anonymized data for training — not actual recordings of real employees. But the problem is, synthetic data might not capture the specific language and context that makes these custom models useful. Luna: So there's a trade-off. Accuracy versus privacy. And right now, most companies are prioritizing accuracy. Lucas: I think it's more that they haven't fully thought through the privacy implications. The tool is new, it's shiny, it promises productivity gains, and the legal risks feel abstract. But they're not abstract. We're already seeing lawsuits around biometric data and voice recordings without consent. Luna: Right. And this is exactly the kind of issue where a class-action lawyer's eyes light up. 'Your company recorded and analyzed 2,000 hours of employee conversations without clear consent.' That's a headline nobody wants. Lucas: Exactly. And the irony is that the companies building these tools genuinely think they're helping. They see it as a way to reduce meeting fatigue — you don't have to take notes anymore, the AI does it for you. But they're not connecting the dots between helpful summarization and permanent data capture. Luna: So what's the best-case scenario here? How do we get the productivity benefit without the surveillance creep? Lucas: I think the best-case scenario involves a few things. One: clear, upfront consent with a meaningful opt-out. Two: data retention limits — maybe the raw recordings get deleted after the model is trained, so there's no permanent library of everyone's conversations. Three: employee access to and control over their own data within the model. And four: an independent audit process to make sure the model isn't being used for performance evaluation or other purposes it wasn't intended for. Luna: That sounds reasonable. But it also sounds expensive and time-consuming. Most companies are going to look at that list and say 'we'll just add a line to the privacy policy and move on.' Lucas: You're probably right. But here's the thing — the companies that do this right will have a competitive advantage in hiring. As these stories start to surface, job candidates will start asking about meeting recording policies the way they ask about remote work or diversity initiatives. 'Do you train AI on my meetings?' That could become a real differentiator. Luna: That's a good point. And it ties into something we've talked about before — trust as a currency in the future of work. If employees don't trust their tools, they won't use them effectively. Lucas: Exactly. And if today's conversation gave you something useful to think about — whether it's a question to ask your HR department or just a new lens on your own meeting culture — you know, the reason we can keep doing these deep dives without ads is listener support. Luna: Yeah, it's a small thing that makes a big difference. If you want to help keep the show going, you can find us at buy me a coffee dot com slash fexingo. Lucas: And that's it — no pressure, just a genuine thank you to anyone who has supported the show before. It lets us spend time on stories like this one instead of reading ad copy. Luna: Alright, back to the data. So we've got the company that did this well — what about the ones that didn't? Any cautionary tales? Lucas: There's one that's been making the rounds internally at a few tech companies. A larger enterprise rolled out a meeting analysis tool without telling anyone. Employees started noticing that their meeting summaries included references to things they said off the record — like personal anecdotes or complaints about a project. The AI was capturing everything. When people complained, the company's response was basically 'it's in the fine print.' That didn't go over well. Luna: And I bet morale took a hit. Suddenly every meeting feels like you're being recorded for a performance review. Lucas: Exactly. The tool was meant to increase productivity, but it ended up decreasing trust. And once that trust is gone, it's really hard to get back. Some employees started having sidebar conversations in text channels instead of speaking up in meetings. The whole point of the tool — to capture decisions and action items — was undermined because people didn't want to be captured. Luna: So in the end, the company actually got less productive. Which is the exact opposite of what they wanted. Lucas: Right. And that's the lesson. If you're going to build or buy a tool that trains on employee conversations, you have to bring the employees along with you. Transparency isn't just a legal checkbox — it's a prerequisite for the tool to actually work. Luna: So the takeaway for our listeners: if you're an employee, ask the question. If you're a decision-maker, think hard about consent and trust. And if you're building one of these tools, build it with opt-in from the start. Lucas: Yeah. Because the technology is only going to get cheaper and more capable. The question isn't whether companies will use it — they will. The question is whether they'll use it in a way that respects the people whose data makes it work.