Most Human Wins

What gets cheap gets ignored

When a capability becomes cheap enough, people stop paying for it and start paying for whatever sits next to it. That is the pattern behind every wave of automation, and it is the pattern reshaping product management right now. The tasks AI compresses were never the point of the job. They were the toll you paid to do the point of the job. AI is now covering that toll, and the work that remains is the whole job.

Look at what actually got automated first. Writing the PRD from a rough set of notes. Summarizing 40 user interviews into themes. Drafting the launch email, the release notes, the competitive teardown, the first pass at a metrics dashboard query. Generating six variants of a feature spec so you can react to something instead of staring at a blank page. Every one of these was real work. None of them was ever the reason a product succeeded or failed.

This is the commoditization pattern, and it is older than software. When a capability gets abundant, its price falls toward zero and its strategic value falls with it. Compute got cheap and the interesting question stopped being “can we afford to run this” and became “what do we point it at.” Manufacturing got cheap and the value migrated to design and brand. The thing that becomes common stops being the thing you compete on. It becomes table stakes, then it becomes invisible, and eventually nobody remembers it was ever hard.

PM craft is going through this now. The parts of the job that were teachable, repeatable, and documentable are exactly the parts a model trained on a million documents does well. That is not a coincidence. Repeatable is what training data captures. The mechanical middle of the job is collapsing, and it is collapsing fastest for the tasks that used to signal competence: the clean doc, the tidy summary, the well-formatted brief.

What stays scarce gets valuable

Packy McCormick’s frame for this is that as AI makes more things abundant, the scarce human things become the whole game — most human wins. In product work, that abstraction gets specific fast.

The mirror image of commoditization is scarcity, and it moves in the opposite direction. As the abundant thing loses value, whatever it was sitting next to gains value. The cost of producing a spec drops to near zero, so the cost of choosing the right spec to produce becomes the thing that separates good from bad.

What is scarce in product work? Judgment about which problem is actually worth solving. Taste about what “good” feels like when the metrics are ambiguous. The ability to sit with a customer and hear the thing they didn’t say. Knowing when the data is lying to you because you understand how it was collected. Deciding what to cut when everything looks important. Building the trust inside an organization that lets a hard call stick. Holding a coherent point of view about where a market goes next when the current surface is stable and everyone assumes it will stay that way.

None of these compress well, because none of them are repeatable in the way a document is. They depend on context that never repeats: this customer, this org, this market, this moment. A model can draft the PRD, but it cannot decide that the PRD is for the wrong feature. It can summarize the interviews, but it cannot notice that the three most valuable customers all hesitated on the same word. The synthesis it produces is fluent and average, because average is what a model trained on everything regresses toward. And average is exactly what scarcity is not.

So the value in the job migrates upward, from producing artifacts to deciding which artifacts should exist and whether the ones you have are any good. This has always been the senior part of the work. What changed is that it is no longer the extra part, the thing you graduate into after you have proven you can grind out specs. It is now the entry price. The grind got automated. The judgment is the job.

Two PMs, one résumé, two futures

Picture two product managers with identical résumés. Same schools, same companies, same shipped features on paper. Hand both of them the same AI tooling.

The first PM uses the tooling to do the old job faster. More docs, more summaries, more dashboards, more variants of the spec. The output volume goes up and the quality of the output stays exactly where it was, because the constraint was never speed of production. This PM feels productive and looks busy. The problem is that everything they are now faster at is the thing that just became free. They have optimized the toll booth. Their comparative advantage is evaporating in real time, and the faster they run, the more it looks like they are running.

The second PM treats the tooling as a way to buy back time from the mechanical work and spends that time on the scarce work. They talk to more customers, not to generate more notes, but to sharpen a point of view about which customer even matters. They pressure-test the metrics instead of just querying them faster. They spend the recovered hours on the argument nobody else in the room can make: this is the problem, this is why now, this is what we cut. Their output volume might not change at all. Their judgment compounds.

Same starting point. In 18 months these two are not competing for the same role. The first has become a faster version of a commodity. The second has become the person the commodity works for. The tooling did not decide the outcome. What each of them chose to point it at did.

The advice that follows

If the scarce work is now the whole job, the career move is to get deliberately worse at the things that are getting cheap and deliberately better at the things that stay expensive.

Concretely: stop treating volume of artifacts as evidence of value. A PM who ships 30 clean documents a quarter is demonstrating a skill the market is done paying a premium for. Redirect that energy toward the calls that require being in the room, holding the context, and owning the outcome. Get closer to customers than any model can, because the model only knows what was written down and the important things rarely are. Build the internal trust that lets your judgment carry weight, because a correct call nobody believes is worth nothing. Develop an actual point of view about where your market’s next distribution surface goes, and be early enough that it is a bet and not a consensus.

The strongest objection here is that the mechanical work was how junior PMs learned judgment in the first place. You wrote 100 bad specs to develop taste about good ones. Remove the reps and you starve the pipeline that produces senior judgment. That is real, and it is the genuine risk in this transition. But the answer is not to protect the busywork as a training program. The answer is to teach judgment directly, earlier, by putting junior people on the scarce work sooner and letting the tooling carry the artifacts they used to grind through. The reps were a means to develop taste. They were never the taste itself.

The uncomfortable part is that this was always true. The judgment was always the job, and the artifacts were always the toll. AI did not change what mattered. It just stopped letting anyone hide inside the part that didn’t.