Fireside: hiring for AI product teams

Farah Siddiqui, who leads talent at Globex Payments, explained why they scrapped their old PM interview loop. The take-home case study rewarded polish: candidates who were good at slides did well. But the job now is mostly judgement on messy, probabilistic output, and the case study never tested that.

Kabir Shah asked her to walk through the new process. There's no whiteboard. Candidates sit down with a laptop, a real (anonymised) feature, and the last twenty things the model produced for customers. They get forty-five minutes and one question: which of these would you ship, and why.

What Farah listens for is whether people slow down. The strongest candidates get suspicious of the outputs that look fine and dig into them. The weaker ones skim, approve the confident-sounding ones and move on.

Globex looked back at who did well after joining. Years of experience predicted much less than they expected. Curiosity, clear writing and a habit of building small things on the side predicted more. Farah's line was that ten years of experience can be one year, ten times; she'd rather see what you built last month.

Globex dropped the "AI PM" title entirely. Farah's view is that there's no such thing: there are PMs, and some of them have spent real time with the models. The separate title had created a two-tier team, with AI PMs treated as specialists everyone else could ignore.

They now level PMs who work on model features on the same ladder as everyone else, with one extra expectation: you can explain how your feature fails.

New PMs at Globex spend their first month on three things before touching the roadmap: pairing with an engineer to build or extend an eval set, shadowing the support team for a week, and shipping one small fix to production. Kabir noted that all three are about seeing real failures up close.

The question that got the biggest reaction was: if agents do the junior work, how do juniors ever learn? Farah said it's the question she loses sleep over. Globex's answer for now is that juniors review the agent's work before seniors do. They learn by catching its mistakes, and seniors review both.

Asked how to interview designers for AI features, she suggested the same approach: show them real outputs, including bad ones, and ask how they'd design the moment the model is wrong.