Latest / The Diversity Career Podcast with Fexingo: Inclusion, Equity, and Underrepresented Professionals / How One Company Uses AI to Reduce Bias in Performance Reviews
Transcript
- Lucas: There's this moment in every performance review cycle where a manager sits down to write feedback and the words they choose can either open a door or quietly close one. And a lot of that happens below the level of conscious bias. Luna: Yeah, we've all seen those studies — the same behavior gets called 'assertive' in a man and 'aggressive' in a woman, or 'brilliant' for one person and 'hardworking' for another. Lucas: Exactly. And one company decided to actually do something about it at scale. A large tech firm — I'm thinking of a case from about a year ago — took their entire performance review system and layered an AI tool on top that basically flags biased language in real time. Luna: So the manager writes a comment and the tool says, 'Hey, this word has been shown to correlate with gender bias — would you like to rephrase?' I love that. Lucas: Right. And it's not just a flag — it offers alternatives. The company worked with a team of linguists and HR researchers to train the model on a corpus of past reviews that had been anonymized and rated for bias. They looked at things like the ratio of 'praise' to 'criticism' words across demographic groups, and the use of what they call 'hedging language' — like 'she seems to try hard' versus 'he delivers.' Luna: What kind of impact did they see? Because the skeptic in me wonders if managers just click through the warnings. Lucas: Fair question. They ran a pilot with about two thousand employees across engineering and sales. And after six months, they saw a fourteen percent reduction in the gap between average ratings for white men and women of color. That's not nothing. And the tool's adoption rate — meaning managers who actually changed a word after a flag — was around sixty-three percent. Luna: Sixty-three percent is actually pretty high for a voluntary tool. I would've guessed closer to forty. Lucas: I think part of it is the framing. They didn't roll it out as a 'bias detector' — they called it a 'feedback clarity tool.' The pitch was: 'Your review is more likely to be taken seriously if it uses specific, observable language.' And that's true regardless of bias. So managers saw it as helping them be better writers, not as being accused of prejudice. Luna: Clever. And it also protects them from downstream complaints, right? If a review is challenged, the manager can say, 'The tool vetted my language.' Lucas: Exactly. And the tool also gives a kind of 'bias score' for the overall review — not shared with the employee, but the manager sees it. So if you write five comments about 'leadership' for one person and five about 'collaboration' for another, the tool might flag that pattern. Because the research shows that women and people of color are more often described in communal terms — 'helpful,' 'team player' — while men get agentic terms — 'strategic,' 'decisive.' And those distinctions affect promotion rates. Luna: So the tool is looking at the whole review, not just individual words. That's much more powerful. Were there any pushbacks? I imagine some managers felt micromanaged. Lucas: There was definitely some resistance in the first quarter. About twelve percent of managers opted out of the pilot entirely. But the company did something smart: they ran a parallel calibration session where they had a separate committee review the anonymized reviews before and after the tool was used. And the committee found that the tool's suggestions led to more accurate ratings — meaning the ratings aligned better with objective metrics like sales numbers or code commits. So they shared that data with managers, and opt-outs dropped to about four percent by the end of the year. Luna: That's a great example of using data to overcome skepticism. 'The tool doesn't just make you more fair — it makes you more accurate.' That's hard to argue with. Lucas: Right. And I think that's the key to scaling any DEI tool — you have to tie it to performance, not just to values. The company also made one other change that I think is worth noting: they started requiring managers to include at least one 'growth area' comment for every employee. Because one of the subtle biases they found was that underrepresented employees were more likely to get reviews with only praise and no constructive feedback — which sounds nice, but it means they don't get the coaching they need to advance. Luna: Oh, that's huge. 'She's great, no notes' — and then she gets passed over for promotion because she hasn't developed any new skills. Meanwhile, the guy with three pages of detailed feedback knows exactly what to work on. Lucas: Exactly. The tool flagged those reviews as 'incomplete' until the manager added a growth area. And they gave suggested prompts: 'Consider what this person could do to be ready for the next level.' So it's not just about avoiding negative bias — it's about correcting for the 'benevolent bias' of not giving critical feedback to people you're afraid to offend. Luna: That's so nuanced. And I bet it also helps with retention — because employees who get only positive reviews but no promotions feel confused and frustrated. Lucas: Absolutely. The company saw a measurable uptick in retention among underrepresented employees in the pilot groups. About seven percent higher over the next year. And they attributed that to the combination of more equitable ratings and more actionable feedback. Luna: So for someone listening who works at a company that doesn't have this kind of tool — what can they do? Because not everyone can just build an AI. Lucas: A few things. First, you can ask your HR team if they've considered adding a bias-check step to the review process. Some vendors offer lightweight versions of these tools as add-ons to existing HR platforms. Second, you can create your own peer review group — a small group of colleagues who exchange anonymized review drafts and flag language patterns informally. That's low cost and builds awareness. And third, if you're a manager, you can adopt the same principles manually: always use specific examples, avoid labels like 'lazy' or 'brilliant,' and make sure every direct report gets both praise and growth areas. Luna: That's really practical. And it ties into something we've talked about before — that individual actions can create cultural change when enough people do them. Lucas: Right. And I think that's the ultimate takeaway from this case. The tech tool is impressive, but the real innovation was the company's willingness to rethink what a review is for — not just a backward-looking evaluation, but a forward-looking coaching tool that serves everyone equally. Luna: Yeah, and if today's episode gave you a new way to think about your own reviews or a specific step you can take at work — you know, the way we keep these conversations ad-free and independent is listener support. So if this was useful, you can buy us a coffee at buy me a coffee dot com slash fexingo. Lucas: And we really appreciate that. It lets us dig into cases like this without worrying about sponsors. So thank you to anyone who does. Luna: Alright, back to the topic. One thing I want to dig into is how the tool handles intersectionality — because bias isn't just about gender or race alone, but the combination. Lucas: Great question. The company actually found that the tool was least effective for women of color compared to other groups in the first pilot. The language models had been trained mostly on data from white women and men of color, so the patterns for Black women or Asian women were less clear. They had to go back and oversample those reviews to improve accuracy. Luna: So even AI has blind spots. That's a good reminder that these tools aren't a silver bullet — they need constant refinement. Lucas: Exactly. And the company now runs quarterly audits on the tool's performance across demographic subgroups. They publish the results internally, which builds trust. And they've opened up some of their training data to academic researchers to help improve fairness across the board. Luna: That's the kind of transparency we need more of. So what's next for them? Are they expanding to other parts of HR? Lucas: They are. They're piloting a similar tool for promotion packets — the narratives that managers write to recommend someone for a promotion. And they're working on a version for self-assessments, because research shows that women and underrepresented groups tend to undersell themselves. The tool would prompt them to include metrics and accomplishments they might leave out. Luna: Wow. That could be really empowering. Imagine getting a nudge that says, 'You led that project but you didn't mention the budget size — add that.' Lucas: Exactly. And that's the direction I think all this is heading — not just removing bias, but actively correcting for it. And that's where we get from fairness to equity.