Latest / Future of Work Tech with Fexingo: Remote Tools, AI Productivity, and Workplace Software / How AI Is Automating Your Internal Candidate Screening
Transcript
- Lucas: Luna, I want to talk about something that's quietly reshaping how companies fill roles from within — ai powered candidate screening for internal talent pools. Luna: You mean like when a job opens up and an algorithm decides which current employees get a look? That feels like it could get messy. Lucas: It can, but companies are sprinting toward it anyway. Unilever's internal mobility platform now uses machine learning to scan employee profiles — skills, past projects, performance reviews — and match them to open positions before they're even posted externally. Luna: Before they're posted? So the algorithm is basically a gatekeeper. Lucas: Exactly. And Unilever says it's cut their time to fill for internal roles by about 40 percent. They're also seeing higher retention rates among employees who move via the system — around 15 percent better after two years. Luna: Those are big numbers. But I wonder — if the algorithm is trained on historical data, doesn't it risk just recycling the same patterns? Like, promoting people who look like the people who already got promoted? Lucas: That's the tension. Gartner put out a report in late 2025 that said 68 percent of large enterprises now use some form of AI for internal candidate screening, up from 42 percent just two years earlier. And about a third of those companies have caught their algorithm surfacing a biased result. Luna: Caught it after the fact, or during testing? Lucas: During testing in most cases, thankfully. But there was a well-documented example at a Fortune 500 financial firm — I won't name them — where their screening model was disproportionately flagging employees from the finance and operations departments while overlooking people in marketing and HR, even when those people had relevant skills. Luna: Because the training data had more finance people in leadership roles, so the algorithm learned to associate 'finance' with 'promotable.' Lucas: Exactly. They had to restructure the model to weight actual skill tags more heavily than department history. It's a classic case of garbage in, garbage out — but the stakes are higher because you're messing with people's careers. Luna: And their sense of fairness. If I'm in marketing and I never get surfaced for a role I'm qualified for, I'm going to feel the system is rigged. Lucas: That's why companies like IBM and Deloitte have been pretty vocal about the guardrails they're building. IBM, for instance, requires that any internal screening algorithm be audited quarterly by a team that includes HR, legal, and a data ethics officer. They also publish a transparency report showing the top factors the model used in the previous quarter. Luna: So you can see, like, 'the model weighed project experience 40 percent, manager ratings 30 percent, and years of tenure 10 percent'? Lucas: Something like that. They stop short of full disclosure on individual scores, but they're trying to build trust. Deloitte goes a step further — they allow employees to opt out of the screening entirely and apply the old-fashioned way, through a manager referral. Luna: Opt out — that's interesting. It acknowledges that the algorithm might not be right for everyone. Lucas: And that's rare. Most companies using these tools make the AI screening mandatory if you want to be considered for an internal role. But Deloitte's approach seems to be working — they report that 80 percent of employees still choose the AI path, and satisfaction scores are higher among those who do. Luna: Because it's faster and maybe feels more objective? Lucas: Right. The human process can be slow and political. An algorithm doesn't know that you had a disagreement with the hiring manager three years ago. It just sees that you led a cross-functional project that matches the role requirements. Luna: But it also doesn't know context — like, maybe that project was a disaster but you learned from it. Lucas: That's where the next frontier comes in. Some vendors are now building 'skills inference' models that don't just rely on keywords in your resume or LinkedIn profile. They analyze your actual work output — emails, documents, code commits — to infer what you're capable of. Luna: That feels invasive. My emails being scanned to decide if I'm ready for a promotion? Lucas: It's opt-in so far, mostly at tech-forward companies. But the pitch is that it catches skills you might not even list. For example, an engineer who writes really clear documentation might be great at product management, but they'd never put that on their internal profile. Luna: So the AI surfaces them as a candidate for a PM role. That's actually kind of cool — if it works. Lucas: If it works and if it's fair. The fear is that inference models pick up on subtle signals that correlate with demographics — like email writing style or time of day you code — and encode bias in ways that are harder to spot. Luna: Yeah, that's a whole new layer of opacity. At least with keyword matching you can argue about it. With inference, you're fighting a black box. Lucas: If today's conversation gave you something useful to think about, I want to mention — the reason we can do these deep dives without ads is listener support. People who find value here can help keep it going at buy me a coffee dot com slash fexingo. Luna: And it genuinely makes a difference. Every contribution goes straight into research and production. Lucas: So back to inference models — one startup I've been watching, let's call them SkillSight, worked with a mid-size retailer to infer skills from Slack messages and project management tickets. They found that the AI identified 30 percent more relevant candidates for each role than the traditional self-nomination process. Luna: Thirty percent more? That's huge. But did those candidates actually perform better? Lucas: Early data suggests yes — their six-month retention rate was 10 percent higher than candidates who applied through the standard process. The theory is that the AI was catching people who were already doing the work informally but hadn't thought to raise their hand. Luna: So it's democratizing access, in a way. The quiet high-performers get a shot. Lucas: That's the optimistic framing. The pessimistic one is that it's another layer of surveillance, and that people will start gaming the system — writing emails in a way that signals 'leadership' or 'strategic thinking' to the algorithm. Luna: Which is already happening with resume keywords. You'd think we'd learn. Lucas: We're slow learners. But some companies are trying to stay ahead by making their models explainable. IBM publishes the feature importance list I mentioned. Others are using what's called counterfactual explanations — telling a candidate, 'You would have been flagged if you had led a project of this size.' Luna: That's actionable. 'Do this one thing and you'll be surfaced next time.' Lucas: Exactly. It turns the algorithm from a gatekeeper into a career coach, at least in theory. The challenge is scaling that kind of feedback without overwhelming HR teams. Luna: So where do you see this going in the next year or two? Lucas: I think real-time skills inference will become more common, but only for roles where the data is rich — like software engineering or consulting. For roles with less digital footprint, we'll still rely on self-nomination and manager referrals, augmented by lighter AI matching. Luna: And the bias question — will regulation catch up? Lucas: The EU's AI Act is starting to bite, and New York City's Local Law 144 already requires bias audits for hiring algorithms. I expect more jurisdictions to follow. Companies that are proactive about transparency now will have a smoother path. Luna: So the smart ones are investing in explainable AI for internal mobility, not just hoping the black box works. Lucas: Right. The ones that treat it as a black box are the ones that will end up in headlines for all the wrong reasons.