Latest / Hiring & Firing with Fexingo: Recruitment, Performance, and Workforce Decisions / How Workforce Analytics Predicts Your Best Hires
Transcript
- Lucas: So there's a number I keep coming back to this week: a mid-sized tech firm in Austin cut its first-year turnover by eighteen percent, simply by letting a model do part of the screening. Luna: Eighteen percent — that's not a rounding error. What kind of model are we talking about? Lucas: It's a predictive hiring model built on their own internal data. They had five years of performance reviews, engagement surveys, and tenure records. They fed all of that into a system that basically learns what a successful hire looks like — for their specific roles, their specific managers. Luna: So it's not a generic 'grit scale' or something. It's literally their own history talking. Lucas: Exactly. And that's the key difference. Most companies buy off-the-shelf psychometric tests. This firm built what's essentially a custom prediction engine. They didn't need a huge data science team either. They used standard machine learning libraries, a couple of analysts, and a clear question to answer. Luna: What was the question? Lucas: Will this candidate still be employed, performing at or above expectations, at the eighteen-month mark? That's it. Binary outcome. They trained the model on historical hires who either fit that description or didn't. Luna: And then they just started scoring candidates on that? Before the interview? Lucas: Yeah, it became a pre-screen, alongside the résumé review and the first phone call. Candidates got a score from zero to a hundred, and anything below a threshold was basically deprioritized. They didn't automate rejection — a human still looked at every candidate — but the model informed where recruiters spent their time. Luna: I like that it's not a hard wall. But I've got to ask — what data actually went into this? Because performance reviews can be pretty biased themselves. Lucas: That's a fair challenge, and it's exactly the kind of thing that keeps a project like this from working. They started with a wide net — everything from salary growth and promotion speed to how many times an employee changed teams. But they explicitly excluded demographics, and they ran a bunch of fairness checks. Luna: But if your historical promotion data is biased, the model learns that bias. I've seen that happen. Lucas: Right. And that's why in the first year, they ran the model in shadow mode. It scored candidates, but those scores didn't influence decisions. They compared the predictions against what actually happened — who stayed, who performed. That's how they found the eighteen percent improvement was real, and they also caught a few biases early. Luna: So shadow mode is like a testing phase. How long did that last? Lucas: About six months. They had a relatively high volume of engineering hires, so they got enough data to validate the model. And the validation step is crucial. You can't just trust a model because it's built on data — data can be messy and historical patterns can shift. Luna: What were some of the strongest predictors? I'm curious what actually mattered. Lucas: The surprising one was the number of times a candidate had changed jobs in the first five years of their career. Not as a red flag, but the pattern mattered. Candidates with two or three short stints early on, but then a longer stable job after that, actually performed better. It suggested they were learning what they wanted, not that they were flaky. Luna: That's a nuance you'd never get from a human résumé screen. You'd just see 'job hopper' and maybe pass. Lucas: Exactly. And that's the promise of this approach — it finds patterns that are counterintuitive, but real for your specific organization. Another strong predictor was how a candidate scored on a work sample test that they'd been using for years. The model essentially said, 'this test is actually more predictive than you thought.' Luna: So it validated a tool they already had. That's a nice bonus. Lucas: It is. And it helped get buy-in from hiring managers who were skeptical. They could see, okay, the model isn't replacing my judgment — it's amplifying the signals I already use. Luna: I want to talk about the ethical side more. When you're feeding a model data like performance reviews, what about privacy? And what about candidates who don't make the cut because of a score they'll never see or understand? Lucas: That's a big question, and it's one that any company should grapple with before starting. In this case, they made a point of not using any data from outside the application — no social media scraping, no public records. Everything came from the résumé, the work sample, and the interview notes. And candidates were informed that a scoring system was being used, though not how it weighted everything. Luna: Is that enough? I mean, 'we use an algorithm' can be a black box. Lucas: It's a start. They did publish a transparency report, which is rare. But you're right, there's a real tension here. The more accurate the model, the more data you're often tempted to feed it, and the more opaque it becomes. That's a genuine problem that doesn't have a perfect answer yet. Luna: I think that's part of why this story is interesting — it's not a perfect utopia. It's a company that made trade-offs, and they're being transparent about it. Lucas: And that's exactly the mindset we'd want to see. They didn't just bolt on a tool and call it a day. They iterated, they validated, and they kept humans in the loop. That's the difference between a fad and a real improvement. Luna: So for a smaller company listening, is this something they can actually do? Or is it only for firms with a data science budget? Lucas: I think it's more accessible than people assume. You don't need a Ph.D. You need clean data, a clear question, and a willingness to test. There are tools like H2O.ai or even spreadsheet machine learning add-ons that can handle this. The hard part isn't the tech — it's the data hygiene and the discipline to validate. Luna: What's the one thing a company should start with if they're curious? Lucas: I'd say start with attrition. It's the most common pain point, and you probably already have the data — hire dates, end dates, maybe performance scores. Build a simple model that predicts who stays past twelve months. That's a manageable first project. Luna: And if they don't have great data? Like, they've been sloppy with HR records? Lucas: Then fix that first. Garbage in, garbage out. You can't build a model on missing or inconsistent data. But cleaning up your HR data is valuable on its own, even if you never touch machine learning. Luna: That's a good point. It's not just about the fancy model — it's about knowing your workforce. Lucas: Exactly. And that brings me to something that's been on my mind about this episode. These conversations we have here are a bit like that — they're about making an impact with the resources you have. And I'd be remiss if I didn't mention that the reason we can keep doing this, ad-free, is because of listeners who chip in a couple of dollars a month at buy me a coffee dot com slash fexingo. Luna: Yeah, it genuinely makes a difference. Even a small contribution helps keep the lights on and lets us keep digging into topics like this. Lucas: So if you've gotten something out of these conversations, that's where you can support the show. Just a thought. Back to the data — the other part of this Austin firm's success was that they didn't just look at the model score in isolation. Luna: How so? Lucas: They combined the model score with a structured interview rubric that was already in place. So the final decision was a blend of the algorithmic score and a human interview score, but each was weighted. Over time, they found that when the two disagreed, the human usually won in the short term, but the model was better at predicting long-term retention. Luna: So there's a humility built in. The model isn't perfect, but it adds a perspective humans miss. Lucas: Right. And that's the sweet spot we're moving toward — human plus machine, not either/or. The companies that get the most out of analytics are the ones that treat them as a conversation partner, not an oracle. Luna: I think that's a great note to end on. Workforce analytics isn't about removing human judgment — it's about sharpening it. Lucas: Yeah. And the most successful implementations I've seen are the ones where the hiring team actually feels empowered, not replaced. That's the outcome worth aiming for.