The CPO Summit ranged from boardroom finance to enterprise pilots, but talk after talk converged on the same word: judgment. Not the AI feature. Not the velocity. The unglamorous, hard-to-replicate human work of deciding what's worth doing, what to measure, and when to stop.
I came away with a tighter version of an argument I've been circling in my own writing: AI raises the price of judgment rather than replacing it. The summit gave me the evidence, and a few frameworks I didn't have before. Here's what stuck.
Your judgment isn't yours: it's your cap table's
Elena Leonova opened with the most genuinely uncomfortable talk of the day, and the one I keep thinking about. Her claim: most CPOs believe their product judgment is a portable skill they built over a career. It isn't. Product strategy is shaped less by customer insight than by a document most CPOs never read closely: the cap table.
She walked through four terms: liquidation preference, participating vs. non-participating preferred, the preference stack, and the "effective minimum exit" (the price at which founders and employees actually make money). Then a rule of thumb: the exit needs to clear roughly 2–3x the preference stack before common shareholders see anything real. The punchline lands hard: that preference stack quietly sets the minimum ambition of your roadmap. Then she made it concrete with her own history: one company, seven years, four investor events, four complete strategy rewrites. A strategic payments investor pushed one direction; an acquisition turned the product into a surfacing platform; a PE buyout flipped the goal to EBITDA; further PE investment demanded geographic expansion to juice the resale number. Same company, same market: investors rewrote the roadmap each time. "Executives evaluate exposure, not ideas."
What this reframed for me: in The CPO's New Reality I wrote that AI compresses decision cycles, pushing the bottleneck from engineering to judgment. Elena's talk added a layer beneath that. AI compresses how fast you can decide. The cap table governs which decisions you're even allowed to make. If your roadmap keeps getting overruled, it may not be a communication problem or a stakeholder problem: it may be an investor-exposure problem, and no amount of better storytelling will fix it. As someone actively evaluating my next role, her Monday-morning diagnostic (read the cap table, research each investor's vintage and exit aspiration, map the roadmap against that reality) is now part of how I'll vet any opportunity. The job description tells you the title. The cap table tells you the strategy you'll actually be hired to execute.
When everyone can build, quality is the product
Michael Krug (unitQ) made the cleanest statement of the day's core thesis: "When everyone can build, quality is the product." His framing is that building is now commoditized but not worthless: the danger is shipping the wrong thing faster, which he bluntly called "slop." Speed without quality just produces churn at a higher rate.
The data he brought made it sting. A 2025 PwC survey: 90% of executives believe customer loyalty is growing; only 40% of customers agree, a 50-point perception gap. And unitQ's own cross-industry quality score sat at about 60 out of 100, a D-minus, unchanged from 2024 even as release velocity rose sharply. "Building faster, but not better." We accelerated output and held quality flat, which is arguably worse than standing still.
His reframe (from velocity of shipping to velocity of learning) is the part I'll use. The closed loop he described (signal → connection → metric → action) puts the hard problem in the right place: not collecting feedback, but connecting it to business impact. Moving from "users are annoyed about payments" to "this specific payments bug in France is tied to $4M of revenue." That's not a support function or a quarterly dashboard; it's prioritization infrastructure. It maps directly onto the Impact and Confidence inputs in my SU-RICE framework: the scores are only as good as the customer-friction data feeding them, and most teams are guessing where they could be measuring.
The moat is what compounds, not what you bolt on
Ivan Galea (Databook) gave the keynote version of the build/buy argument I've been making, and sharpened it into a test I'll keep. His thesis: "AI didn't create differentiation — it compressed it." Foundational models and engineering scale are available to everyone now, so throwing resources at a problem buys you nothing durable. His line: "If your roadmap can be replicated in six months, it's not a moat."
He told a story about an algorithm he built on Wall Street that improved trading by 30%+ and then sat unused for months because traders feared it, then adopted in weeks only once leadership reframed it and aligned incentives. The capability was never the issue; the product judgment was. From that he drew the distinction I find most useful: a data lake is a stranded asset; a data flywheel is a living advantage. Capturing data isn't a moat. A loop where insight attracts users, users generate behavior, and behavior sharpens the insight. That compounds, and competitors can't copy it by shipping a feature.
His three Monday-morning questions are going on my wall: Could a competitor ship this in six months? Are we making a task faster, or rethinking the work from first principles? Does this roadmap item feed the data flywheel or just consume it? That last one is the kind of question my "buy the commodity, build the differentiation" rule was always reaching for and never quite named.
Pilots die in the room you weren't in
Arvita Tripathi's (Vahana Labs) closing keynote was the most operationally useful talk for anyone selling AI into the enterprise. Her statistic: only about 4 of every 33 AI proofs-of-concept reach production. And her diagnosis is that they die not as sales failures but as product design failures. The pilot is a product surface, the first time your product meets organizational reality: governance, IT constraints, budget cycles, veto-holding stakeholders.
Two deals, same product and team and year, opposite outcomes. In one, procurement wrote the spec and the evaluation measured the relationship, not whether the product worked. In the other, the customer put the patient first, and clinical/regulatory metrics survived to renewal: the platform got filed as "required infrastructure." The only difference, she said, was who was in the room when the evaluation criteria were written. If product isn't there, the organization wins by default and you spend six weeks answering a 630-question security review you never designed for.
Her practical fix is to define kill criteria up front, in three classes (evidence, adoption, and governance) and to treat governance as a product decision, not an accuracy test: "The question isn't how does it work — it's can you be accountable when it doesn't." This connects directly to the agentic-retail talk earlier in the day, which noted that pilots fail on integration, data, and commercial misalignment rather than on technology. Tripathi named the mechanism: buyers can't articulate their own kill criteria, so it's the product leader's job to write the spec nobody asked for.
The thread: discipline over capability
The panels filled in the rest. "AI Isn't Your Strategy" had panelists converging on ideas I've already published: AI is an accelerant, not the strategy; MCP is becoming table-stakes; advantage lives in niche, data, and trust. The framing I'm taking from it is "token economics as headcount": treat tokens like a resource you allocate to your highest-value work, because one company in the room burned a year's token budget in six months demoing without any ROI model. And "Vision to Velocity" reframed the strategy-execution gap as process debt (the same way we talk about technical debt) with a guardrail philosophy I like: give teams context, not control.
Step back and every talk reduces to the same move. Elena: judgment is governed by financial structure, so understand it. Krug: judgment about quality is the moat once building is free. Galea: judgment about what not to build is the only thing that compounds. Tripathi: judgment about how you'll be evaluated has to be designed, not discovered. None of them were really talking about AI. They were talking about the human work AI makes more valuable, not less.
That's the throughline of everything I've been writing this year, and the summit pushed me to name it more directly. I think my next piece is the pairing essay, what AI won't decide for you, built on these four talks. And I'm increasingly convinced that the "Mountaineer's Mindset" framing I use isn't just a personal brand flourish. The mountain doesn't reward the fastest or the best-equipped. It rewards the one with the judgment to read the conditions, allocate energy, and know when to turn around. That's the CPO job now, too.
Notes are mine; any misattributed quotes are the fault of auto-generated transcripts, and I'll correct them as I verify against the program.