Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI Is Screening Your Internal Job Applications
Transcript
- Lucas: Luna, I want to start with a number that caught my eye: a Fortune 500 retailer we'll call 'NationalMart' recently told me that before they rolled out an internal talent marketplace, roughly 70 percent of their open roles were filled by external candidates. After one year with the AI matching system, that flipped — now 65 percent of roles are filled internally. Luna: That's a dramatic shift. And I imagine it saves a ton on recruiting fees and onboarding time. Lucas: Absolutely. They estimated they cut external recruiting costs by about $2.3 million annually. But more interesting to me is how the AI actually works. It's not just matching keywords on a résumé. The system ingests performance review data, completion of training modules, even patterns from Slack messages — how often someone contributes to certain channels or answers questions — to build a skills profile that the employee themselves might not articulate. Luna: So it's inferring skills from behavior, not just self-reporting. That's both powerful and a little creepy. Lucas: Yeah, exactly the tension we want to talk about. If today's conversation gave you something usable, whether it's a framework for thinking about AI in hiring or just a stat to sound smart at a meeting, you can support the show by buying me a coffee — literally, the link is at buy me a coffee dot com slash fexingo. It keeps the show ad-free and helps us dig into stories like this. Luna: It's a small gesture that makes a big difference. So back to the system — how does it decide who gets recommended for a role? Lucas: Right. The algorithm assigns each employee a 'fit score' based on three weighted categories: demonstrated skills, which is about 50 percent of the score; growth trajectory, which looks at how quickly they've mastered new competencies, about 30 percent; and cultural alignment, which is derived from survey data and peer recognition, the remaining 20 percent. Luna: And managers see these scores when they're reviewing candidates for an internal position? Lucas: They do. But here's the crucial detail: the system also surfaces candidates who haven't applied yet. If an employee's profile matches an open role above a certain threshold — usually 80 percent fit — the system suggests them to the hiring manager and sends the employee a nudge saying, 'Hey, we think you'd be a great fit for this role.' Luna: That is huge. Because a lot of people don't apply because they don't think they're qualified, or they don't know the role exists. The AI is basically acting as an internal recruiter. Lucas: Exactly. NationalMart saw a 30 percent increase in applications from underrepresented groups for roles they wouldn't have considered before. The system was designed with a 'bias guard' — it explicitly removes demographic data and only uses skill-related signals. But critics argue that if the underlying performance data is biased, the AI just automates that bias at scale. Luna: And we've seen that play out in other contexts, like hiring algorithms that penalized women because historical hiring data reflected past bias. How did NationalMart address that? Lucas: They did two things. First, they audited the model every quarter using a technique called 'counterfactual simulation' — they'd swap the gender or race in the profile and see if the fit score changed. If it did, they'd retrain the model. Second, they gave employees the right to opt out of having their Slack messages and other behavioral data used for matching. About 12 percent of employees opted out. Luna: Twelve percent is low. I'd expect more people to be uncomfortable with their Slack being analyzed. Lucas: I thought so too. But when they surveyed employees, the majority said they'd rather have the AI find them opportunities than rely on their manager to remember them when a role opens up. The perceived benefit outweighs the privacy concern for most. Luna: So the system is popular. But are there cases where it's clearly wrong? Where the algorithm says someone is a bad fit but a human would disagree? Lucas: Yes. One example NationalMart shared: a warehouse supervisor with ten years of experience applied for a corporate logistics role. The system gave her a fit score of 62 percent — below the 70 percent threshold for automatic recommendation. Why? Because she had no formal project management certification, and her Slack activity was mostly operational, not strategic. A human recruiter would have seen her deep knowledge of the supply chain and at least interviewed her. Luna: So the system missed her because it couldn't weight decades of tacit knowledge the same way it weights a certification. That's a fundamental limitation. Lucas: It is. And NationalMart addressed it by adding a 'human override' pathway: any employee whose fit score is below 70 percent can still request a conversation with the hiring manager. About 8 percent of hires now come through that override route. Luna: That seems like a reasonable safety valve. But I wonder if the override is used equally across groups — do women and people of color use it as often as white men? Lucas: Great question. NationalMart's data showed that women were 40 percent less likely to use the override than men. The company's theory is that the confidence gap shows up even when an algorithm says you're not a fit. So they started sending a follow-up message to employees below the threshold that says, 'Your experience is valuable — if you want to discuss this role with the hiring manager, here's a link to schedule.' That increased override usage by 20 percent among women. Luna: That's a smart design intervention. It acknowledges that the algorithm isn't perfect and that human judgment still matters. Lucas: Exactly. And that's the broader lesson I'm taking from this: internal talent marketplaces are powerful because they surface opportunity, but they need guardrails. The companies that do this well treat the AI as a co-pilot, not an autopilot. Luna: I want to zoom out for a second. Do you think this trend will reduce or increase overall employee turnover? Lucas: Early evidence from a few companies suggests it reduces voluntary turnover by 15 to 20 percent, because employees see a path to growth without leaving. But there's a catch: if the AI consistently tells people they're not a fit for roles they want, it could accelerate disengagement. NationalMart saw that effect with about 5 percent of employees who repeatedly got low fit scores — they actually left sooner. Luna: So the AI can become a self-fulfilling prophecy. 'The system says I'm not right for anything here, so I guess I should leave.' Lucas: Right. That's why some companies are now using the same AI to recommend development paths, not just jobs. If you don't match a role, the system suggests courses, projects, or mentors that can close the gap. It turns a rejection into a roadmap. Luna: That's a much healthier framing. I'm curious — what's the one thing you think every listener should take away from this episode? Lucas: That the future of internal mobility isn't just about matching people to open roles. It's about creating a system that continuously surfaces opportunity and development, while building in human oversight to catch what the algorithm misses. If you're an employee, it's worth checking whether your company has something like this — and if not, maybe asking why. Because the data suggests it benefits both the business and the workforce. Luna: And if you're an HR leader, the key design choice is whether your AI is a gatekeeper or a guide. The best ones are guides. Lucas: Exactly. That's the takeaway. And on that note, we'll leave you to think about how your own skills might be invisible to the algorithms your company uses — and what you can do to make them visible.