Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI Is Predicting Your Resignation Before You Send It
Transcript
- Lucas: So Luna, here's a stat that stopped me this week. A recent study from a major HR analytics firm found that by analyzing just three months of employee digital exhaust—Slack messages, email metadata, calendar activity—predictive models could flag someone who's going to resign within the next two weeks with about 85 percent accuracy. Luna: Eighty-five percent? That's surprisingly high. I mean, we've heard about people analytics for years, but that kind of precision feels new. Lucas: It is new. And it's catching on fast. Companies like Workday have built modules specifically around 'flight risk' scoring. Microsoft's Workplace Analytics tool now includes a feature called 'Retention Risk' that pulls from your Office 365 data—how often you email people outside your team, whether you're skipping meetings, even the sentiment of your messages. Luna: Wait, they look at sentiment? So they're reading the tone of my emails to guess if I'm unhappy? Lucas: Not reading the full content in most cases—more like analyzing metadata patterns. For example, a sudden drop in positive language, an increase in direct communication with your manager, a spike in LinkedIn activity during work hours. All these are statistically correlated with an upcoming departure. Luna: That's a lot of data points. But isn't there a huge false positive risk? Someone could be job hunting casually, or just having a bad month. Lucas: Absolutely. The models are probabilistic, not deterministic. Workday's documentation says their score is meant to be a 'conversation starter,' not a verdict. But here's where it gets interesting—some companies are using these scores to trigger retention interventions. Increased bonus, new project offer, a chat with HR. Before the employee even says they're leaving. Luna: That feels like a double-edged sword. On one hand, if you're about to lose a star performer, you want to act. On the other, if the model is wrong, you might come across as paranoid or intrusive. Lucas: Exactly. And there's a deeper privacy question. Most employees don't know their digital footprint is being fed into a retention model. It's buried in the terms of service. A 2025 survey from the Electronic Frontier Foundation found that only 12 percent of workers felt they had a clear understanding of how their workplace data was being analyzed. Luna: That's a transparency gap. But let's talk about accuracy—how do these models actually work under the hood? Lucas: The core technique is called 'sequence modeling.' The system looks at your digital behaviors over time—not just snapshots. It learns patterns like: normally you email 15 people a day, but for the last week it's been 5. Or you used to attend 90 percent of team meetings, now it's 60 percent. These deviations get weighted and combined into a risk score. Luna: And they're training on millions of employees across companies, so the pattern library is huge. Lucas: Right. One vendor claims their model is trained on over 10 million employee-years of data. They've seen the exact same behavioral signature play out across industries. It's like building a universal resignation fingerprint. Luna: That's both impressive and creepy. Are there any examples of companies using this that have gone well—or badly? Lucas: Well, one well-known case is a large tech company that used its internal analytics to identify a high-performing engineer flagged as a 90 percent flight risk. They proactively offered her a retention package—a promotion, a bigger project. She later said she hadn't even started looking, but the offer made her feel valued. She stayed. Luna: So that worked. But what about the flip side? Lucas: Another company—a financial services firm—used the tool to flag a whole team as high risk. The manager confronted them collectively, and three people quit within a month, two of whom hadn't been planning to. It created a self-fulfilling prophecy. Luna: Ouch. So the implementation really matters. It's not just the model—it's how you act on it. Lucas: Exactly. The best practices are emerging: use the score as a prompt for a genuine conversation, not a directive. And always couple it with a culture where employees feel they can talk about dissatisfaction without retaliation. Luna: It also raises the question of whether this will change how people behave at work. If I know my Slack patterns are being watched, I might start acting 'normal' even if I'm unhappy. Lucas: That's the Hawthorne effect redux. And it's already happening—some employees are deliberately mixing up their communication patterns to avoid being flagged. There's even a term for it: 'analytics camouflage.' Luna: Analytics camouflage? That's both funny and a little sad. So we're entering an arms race between prediction and evasion. Lucas: We are. But the bigger picture is that resignation prediction is just one piece of a larger shift toward workforce intelligence. The same models can predict performance, team cohesion, even burnout before it happens. Luna: And if the tech is good enough, maybe it can help companies create environments where people don't want to leave in the first place. Lucas: That's the hopeful framing. But it requires a lot of trust—and a lot of transparency. Which brings me to something I wanted to mention. If today's conversation gave you something useful—a new angle on your own work, or just food for thought—we keep this show going because of listeners like you. Luna: Yeah, and it genuinely doesn't take much. A few dollars a month helps us stay ad-free and focused on what we think matters. Lucas: If you've gotten value out of the show, consider throwing a coffee our way at buy me a coffee dot com slash fexingo. It's that simple. Luna: And we really appreciate it. Now, back to the resignation prediction. One thing I'm curious about is whether this technology is more effective in certain industries. Lucas: Great question. The data shows it works best in knowledge industries—tech, finance, consulting—where digital tool use is high and the work is collaborative. In retail or manufacturing, where fewer digital signals exist, the models rely more on things like shift attendance and payroll data. Luna: So the prediction accuracy varies. What about the legal landscape? Are there any regulations around this? Lucas: Not yet in the US, but the EU's AI Act is starting to classify these as 'high-risk' systems, which would require transparency and human oversight. Some states like California are looking at similar bills. So the regulatory environment is catching up. Luna: That feels inevitable. And probably healthy. I think the biggest question for me is: do employees have a right to know their score? Lucas: Some companies are experimenting with that—showing employees their own flight risk score as part of a dashboard. The idea is that if you see you're flagged, you can proactively address your concerns with your manager. Luna: That's radical transparency. I imagine it could backfire if the score is wrong, but at least it's honest. Lucas: Right. And it aligns with a broader trend: treating employees as partners in their own retention. The best prediction tool is still a good conversation. Luna: So where do you see this going in the next two or three years? Lucas: I think we'll see these models get more granular—incorporating voice tone from meetings, biometric data from wearables, maybe even sentiment from facial expressions. The privacy debate will intensify. But the technology is already here, and it's only getting better. Luna: Well, it's definitely something to keep an eye on. Thanks for breaking it down, Lucas. Lucas: Always a pleasure. And if you're listening, we'd love to hear your thoughts—are you okay with your employer predicting your next move? Reach out to us.