Craig Unsworth: “If your value is writing the document, AI is coming for the document

Let me catch you up. The job you hired for is changing shape while your team is still doing it. Nobody sent a memo. But the work underneath the titles, PM, designer, engineer, has shifted, and the people doing it feel it even if they can't name it.

The loud version of this story is "AI kills product management." That's not what's happening. What's happening is messier and more useful to understand. The parts of these jobs that moved information around are losing value fast. The parts that require judgment are gaining it. Your real decision is where to draw the line between what your people own and what the agents do. Here's where we are.

The deep cut

• Judgment is the job now, not the handoff. Craig Unsworth says if your value is writing the PRD, "AI is coming for the document."
• One broad role is not an operating model. Unsworth warns the all-in-one Product Engineer is "a Marvel character," not a plan.
• Review is a different muscle than build. Amplitude's AI engineers now turn unverified agent code into something customers trust.

The machinery is losing its value, not the judgment

The version of product management built on coordination is in trouble. Craig Unsworth puts it plainly: "If your value is writing the document, AI is coming for the document." Requirements, tickets, status reports, translating between smart people who should just talk directly. That work is getting cheap fast.

But the need to pick the right problem is going up, not down. When you can explore ten ideas in the time one used to take, choosing which one deserves attention matters more than documenting the choice. Robby Stein from Google Search said it at the Lenny Summit: "If almost anything you can think of can get built," the real value of PM "is around judging. It's around taste." Claire Vo's line from the same stage lands harder: "now we can ship all our bad ideas, like congratulations to us."

The one-person-does-everything trap

The tempting fix is to collapse the roles. Put engineers close to customers, hand them AI tools, let them discover, design, build, and ship. Fewer handoffs, more ownership. There's real good in that. But watch where it goes.

Unsworth tracks the wishlist for this new "Product Engineer" and it gets absurd fast: excellent engineer, deep customer judgment, design taste, pricing instinct, market knowledge, plus B2B procurement and security. "At some point, we have stopped designing a role and started writing a Marvel character." AI makes a talented person broader. It does not make everyone interchangeable. A coding tool does not hand a PM years of engineering judgment, and it does not hand an engineer commercial instinct.

The summit crowd agreed from the other side. Tamar Yehoshua of Atlassian framed it as roles overlapping and expanding, not merging into one builder. Same evidence, opposite conclusion from the hype.

Review became the job when nobody asked

Here is the shift your engineers are already living. Amplitude's Adam Bonefeste describes it well: engineers stopped writing code and started overseeing agents that write it. Shopify, Airbnb, and Google now report half or more of their code is AI-generated. The engineer assigns work, waits, reviews, ships, and owns the result. "If the agents did bad work, the engineer did bad work."

That sounds like a promotion nobody applied for, and it is. Writing code let you catch mistakes as you went. Now review is a separate step bolted on at the end, a different skill entirely. Faros AI found engineers using AI touch 67% more work per day but leave 26% more tasks idle for a week. Easier to start, harder to finish. One Amplitude engineer calls himself "the janitor" because human attention, not code, is now the scarce thing.

Draw the line on purpose, or it draws itself

The leaders doing this well are explicit about who owns what. Zohre Shirazi's rule is the cleanest I've seen: normal case, AI handles it; uncertain case, AI escalates; high-impact case, a human approves; unsupported case, the workflow stops. The failure mode is telling someone to "just review everything." That doesn't remove work. It moves it from writing every reply to reading every reply.

Mirza Besirovic at Zendesk, whose teams scaled AI agents past $100M in ARR, is blunt about where he stops: "I limit AI's use when it comes to people." Team development, career growth, relationships stay human. And he warns against mandates without support: do that and "you're effectively tilting at windmills." The French CPOs back this up. At Cegid, a separate AI team failed; a team that tooled up the squads worked. At YouTrust, the biggest blocker wasn't tech, it was people gripping their old job while it changed.

Three questions for your team

• Which activities on your team are coordination, and which are judgment? Name them, because AI is pricing the first group down fast and you don't want to defend headcount on the losing side.
• Who owns the outcome when an agent ships bad work? If the answer is "the AI did it," you have a gap. Pick the human owner before the failure, not after.
• Where does the line sit between what agents do and what people decide, and is it written down? If your team can't repeat it in a review, it isn't real yet.