Why the best product leaders are becoming ICs again | Tom Verrilli (CPO of Whatnot)

Why does it matter? Because the skills that earned PM promotions for a decade are the ones that no longer transfer

The CPO of Whatnot spent seven years at Twitch and started his career at Twitter — and his sharpest conclusion is that most product teams are organized around the wrong idea. Tom Verrilli's argument isn't that product management is worthless. It's a precise diagnosis of why a role that mattered so much produced so few practitioners who can actually do the work.

• The 6-engineers-to-1-PM ratio was a delegation shortcut that scaled into a specialist class — and atrophied everyone else's product judgment in the process
• 31,832 people applied to be a PM at Whatnot in two years; one was hired — because candidates had mastered alignment theater, not the actual craft of scoping, understanding customers, and defending a position with data
• Whatnot's VPs spend 90%+ of their time as individual contributors; even the CPO is 50% IC — because seniority without ground truth produces bad approvals, not better ones
• AI collapsed a week of data science work into hours; fluency with tools like Hex threads is now the real PM edge, not stakeholder management

Hiring a PM for every team robs your engineers of the exact reps that would make them great

The 6-engineers-to-1-PM ratio was never architectural necessity — it was scale anxiety dressed as process. Verrilli traces the lineage directly: internet businesses grew faster than any other industry in history, founders delegated specifics to a new specialist class, and somewhere that delegation calcified into a hiring formula no one questioned.

The damage is specific: "hiring so many PMs infantilizes the engineers and the designers who are perfectly capable of making good decisions but just never had to because there was always a PM to babysit them." PM work is a trade, not a qualification — built through reps. Map a PM to every team and no one else ever gets them.

At Whatnot, PMs are mapped to problems, not teams. A 20-person PM org serves the fastest-growing US marketplace business of all time. The semi-annual planning cycle produces a list of what needs to be true, assigns DRIs, and regularly surfaces gaps — "there's this thing that's like second or third priority in a bunch of different teams' roadmaps — who owns that?" That's when a PM gets assigned. Not as a default.

Before adding a PM to a team: diagnose whether the gap is context or capability. If it's context, fix it directly. The PM hire treats the symptom; fixing context fixes the root cause.

The industry promoted its A-players straight out of doing anything useful — and the fix is putting them back on the keyboard

"We took all of our A players and then promoted them out of doing things." Verrilli says it flatly. The career ladder became a mechanism for extracting experienced people from the work they were best at, replacing it with coaching sessions, reviews, and alignment meetings.

At Whatnot, the model is inverted: managers "would spend 90 plus% of their time doing IC work." Verrilli himself is 50% IC — pulling data personally, querying the codebase directly, writing specs, sitting in support tickets. The leverage case is concrete: a VP with 15 years of reps can cover the workload of multiple junior PMs and make faster decisions without the alignment tax.

He cites a Twitch fix: the discovery team and the ads team were in perpetual war over feed impressions until he made one PM accountable for both. Months of cross-team politics dissolved. "When you put the same person across multiple things, they tend to organically align those things and you just cut out months and months of back and forth."

The Messi argument is the one that lands: "why wouldn't you want Messi playing for your team rather than trying to have the academy coming along all the time?" The industry answered that question wrong for a decade.

31,832 applicants, one hire — what the case study revealed was brutal

Give someone a prompt and some data. Ask them to come back with a point of view. Make them defend it verbally. "How quickly the thinking decays from folks who are good at the theater but not the specifics" — Verrilli frames this not as gatekeeping but as an accidental audit of what the PM market actually produced.

Two years of applications, one hire, and the signal was consistent: candidates had gotten exceptional at presenting frameworks to leadership and narrating alignment wins. The actual craft — scoping correctly, understanding the real customer problem, defending a position under pressure with data — had atrophied. Candidates presented incredibly well, then the thinking collapsed the moment a real data prompt arrived.

Verrilli doesn't spare himself: "I'm guilty of this. We rewarded it for so long." What's trending down in his interviews: candidates who lead with stakeholder management and driving alignment. What's trending up: people who hold macro and micro simultaneously — articulate a full system belief, then immediately pivot to the smallest possible test. That switch, executed live under a case study prompt, is where the theater falls apart and "there's not a lot of place to hide in that anymore."

The accordion forces the discipline product teams keep skipping — compressing between full vision and V1

Ship without a hypothesis: spaghetti. Write a three-year roadmap and ignore what A/B tests teach you: you've surrendered software's one structural advantage over every other industry. Both failure modes are endemic, and the accordion is Verrilli's memory device for the discipline that sits between them.

Before you play a note on a piano accordion, you stretch it all the way out — pull in the air. "But you don't make music until you press the key and push it all the way back into V1." Then, before the next move, stretch it out again: given what we just learned, what do we actually believe now?

Verrilli grounds it in a live Whatnot dilemma: sellers in live commerce don't historically need listings — just hold up the product and describe it. Fast, low-friction. Zoom out: search can't work if the platform doesn't know what's being sold before it sells. Mandate listings? Zoom out again: three minutes per listing dramatically reduces how many items a seller can move per hour, hurting their business. The accordion makes you run the full loop before committing to any of it.

The operational version: state the belief, find the minimum test, read the result against that belief, update the plan — then re-expand to full strategic view before the next step.

AI's real unlock for PMs isn't prototyping — it demolished the data science bottleneck

Ten times more time in data. Less time talking to data scientists than at any point in a 15-year career. Verrilli calls it "the first one by a country mile" when ranking what AI has actually changed about how PMs operate.

The specific comparison: a Hex thread now delivers what an Amazon L7 data scientist needed "a week, two weeks" to produce in 2017 — nuanced cohort reports, individual user logs, regression models, sensitivity forecasts. The structural bottleneck that governed PM velocity for years dissolved. "I've spent less time in the last year talking to a data scientist than I ever have in my career, even though I've probably spent 10 times more time in data."

The caveat is the key one: the tool lowers access costs; it does nothing for interpretation quality. This is precisely why Verrilli argues for fewer, more senior PMs — the leverage from AI-assisted data work accrues to people who already have enough reps to recognize when an answer looks wrong. He also cites a second unlock: talking to Claude to understand codebase architecture directly, sidestepping a class of engineering interruptions entirely. Self-service data without judgment is just faster noise.

A 3% adoption rate can be 100% of someone's business — and averages will never show you that

"They just lie to you all the time." The failure mode Verrilli traces back as the thread through his worst professional decisions: relying on averages without segmenting who lives beneath them.

The scenario is routine: a feature sits at 3% usage. Low adoption, maintenance burden, deprecation candidate. On averages, the call looks clean. But beneath the average: a group for whom that feature is their entire workflow, their primary use case, their livelihood. "This is somebody's business right — if we're just not reliable, it's kind of like a Westfield mall just turning off the power in the lead-up to Christmas without thinking about it."

In e-commerce especially, the downstream spiral is enormous. Seller trust is the foundation; reliability is the asset. Deprecate a feature used by 3% of sellers without understanding that those sellers run specific business models that depend on it, and the blast radius shows up months later in churn data you'll struggle to explain.

The fix: segment before any deprecation or prioritization decision. The 97% who don't use a feature are irrelevant to that specific call. The 3% who can't operate without it are the only number that matters. Bezos captured the same logic from the opposite direction: when you have data and an anecdote, trust the anecdote.

'Know then go' turns systems thinking from a caution reflex into a speed advantage

PMs who skip the second-order effects mental exercise don't move faster — they get ambushed later, when legal, finance, or a downstream team surfaces the thing that wasn't thought through. The alignment meetings they avoided find them anyway, at the worst possible moment.

"Know then go": think through everything that could go wrong before you start, map the knock-on effects, understand where scale will break — then move. "If you've thought through all the things that could happen at scale, you're probably going to preempt a bunch of them. You don't have to solve all of them. You've just got to think through all of them and then you end up solving more than you think."

The companion question he asks in every product review: "What do we do if it's green? What do we do if it's red?" If the answer to either is "I'm not sure how my strategy would change" — stop. The hypothesis isn't testable yet. The mental exercise before pen on paper: what if adoption is 1000x expected, which teams get affected downstream, what would legal or finance flag? You don't build a remediation plan for all of it. You just have to have run the simulation.

Product judgment is being redistributed — and organizations still building layers are running a model the world no longer requires

The direction all of this points is less about headcount and more about where judgment lives. When senior PMs do IC work, engineers get product reps, and AI handles data access — the justification for the dedicated PM as a structural layer dissolves in more and more places. The role doesn't disappear. It narrows to where genuine craft is irreplaceable. What fills the space left behind isn't chaos; it's engineers and designers who finally got the reps. The organizations still stacking reviews on top of reviews are running a delegation architecture built for a slower, more opaque world. The question was never whether to hire fewer PMs. It was always whether the best judgment in the room was anywhere near the actual decision.

---

Topics: product management, team structure, IC vs management, AI tools for PMs, hiring, systems thinking, product leadership, organizational design, career advice, consumer marketplaces, data-driven product