What flattens out, and what rises
Once culture and structure have moved, what does the individual person actually need to be good at? That’s what this layer is about. Not roles, not processes, but the craft. What do people need to be able to do, and what no longer differentiates?
Fewer skills than before. Deeper than before. That’s the short version.
What becomes commodity
My thesis is that most of what we call digital craft today becomes commodity within a few years. Not because there’s no quality difference between good and bad execution, but because the difference is no longer what differentiates a company or a career.
Frontend development. UI design. Backlog management. Test writing. Practically all CRUD code and standard integrations. It’s not a question of if. It’s a question of how fast.
Benchmarks give a nuanced picture of how far along we are. AI agents can currently solve 91% of isolated coding tasks in HumanEval (Tufano et al., 2024). But in a simulated enterprise context with real workflows, emails and colleagues, the best agent can only solve 30% of tasks autonomously (Xu et al., 2024). The difference is context and complexity, and it’s closing fast.
My assessment: in Danish enterprise reality, much of what gets built is commodity complexity. Standard flows, integrations, dashboards, forms. AI handles it already or will within a few years. That doesn’t mean the technical roles disappear. It means the technical roles concentrate on what’s actually complex.
What doesn’t become commodity
Some skills rise at the same time:
- Discovery discipline. Investigating before building. Testing assumptions. Continuing to ask “is this the right problem?” No LLM does that.
- Facilitation. Gathering people with different perspectives and moving them from disagreement to decision. Described already.
- The courage to say no. Saying no to 90% so the 10% can live. Knowing the difference between what can be built and what should be.
- Coaching. Helping others make better decisions, not by giving answers, but by asking better questions.
Four more worth adding:
- Writing. When AI executes on text, precise language becomes the new programming language. People who can write clearly, concisely and without ambiguity are going to fly.
- Context curation. The ability to find, gather and structure the right context for the AI to work from. It’s a new skill that barely has a name yet, but it’s the difference between someone who gets three great outputs and someone who gets 30 mediocre ones.
- Systems thinking. When individual decisions get cheaper, the ability to see how they play into a larger system becomes more valuable. What happens to the whole customer journey, not just this feature?
- Output evaluation. AI produces plausible nonsense faster than ever. The ability to read code, design, text and spot where it limps, without having produced it yourself, becomes core craft.
That’s not ten new skills. It’s maybe seven or eight in total. But they’re hard. They take years to build.
AI removes the capacity, not the taste
A developer can make a UI with AI at their side. That’s not the same as knowing what works for the user. A PM can code an MVP in an afternoon. That’s not the same as having the patience for maintenance afterwards. A designer can generate a strategy in Claude. That’s not the same as having a grip on the business numbers.
AI removes the capacity barrier between the domains. It doesn’t remove the taste, the instinct or the judgment that comes from ten years in a field. That’s easy to miss. When AI can fill in what you don’t have yourself, the temptation arises to stop differentiating at all: everyone can just wander around each other’s domains. That freedom exists. It leads to mediocre results in all three domains.
Every person should have a primary area, the place where their judgment is deep and sharp. And a deliberate secondary capacity, where they can contribute without waiting for others. It’s the difference between a T-shape and a comb shape. The T had its time. The comb fits an organization where everyone can touch everything.
Sweet spot: the product leader’s triangle
The comb applies to everyone on the team. The product leader’s triangle is narrower. Three competencies, and it’s the overlap between them that decides whether a product leader is strong in an AI-augmented context.
The first is domain knowledge. The specific understanding of interior designers, hospital pharmacists, pension advisors. Niche knowledge that can’t be acquired through an LLM. It’s also why the startups that win are often vertical: they build for a domain where knowledge is hard to acquire from the outside.
But domain knowledge alone is a trap. Domain experts who don’t understand discovery become the most dangerous people in the organization. They spit out features nobody uses, because their own experience is their only test bench.
Discovery discipline is the second answer. Not enough on its own either. You can investigate forever without making decisions. The build trap gets worse with AI: when the cost of building falls toward zero, the pressure to produce gets even bigger. The only antidote is daring to say no.
The triangle:
- Domain knowledge, you understand the industry, the customers, the problems that aren’t written down.
- Discovery discipline, you learn before you lock. Your own experience is a starting point, not a conclusion. And you know when the problem should stay open, and when it should be closed down.
- The courage to say no, not as an attitude, but as a practice. You say no to 90% so the 10% can live, and you carry the responsibility for the cut.
The research points the same way. A 2026 study of 174 product managers concludes that AI literacy, evaluation discipline and decision support design become core competencies for product managers going forward (Oladeji, 2026), precisely because cognitive overload and decision fatigue are the primary risk as the amount of information grows. Marty Cagan has argued for years that the ability to say no is core to good product management. Melissa Perri calls the build trap the biggest systemic risk in modern product development. John Cutler explicitly points to decision quality as the new bottleneck in AI-augmented teams.
The new product leader lives in the sweet spot. Not in prompting skills, Figma skills or backlog management.
When you hire, the sweet spot is what you’re looking for, not a certification or a title.
And what does an AI product manager actually do, then? The job postings have started appearing. My assessment: the title is new, the craft is this. The triangle is the answer, not a separate role.
Prompting is not an afternoon skill
There’s a widespread assumption that prompting will soon be the same as ordinary text: you just describe what you want, and the AI does the rest. The research says otherwise.
In a much-cited study called “Why Johnny Can’t Prompt”, researchers tested 14 non-experts at designing prompts for an LLM-based chatbot (Zamfirescu-Pereira et al., 2023). They failed systematically, not because prompting is technically hard, but because they overgeneralized from how you talk to humans. They assumed the AI would fill in what they meant but didn’t write. A larger study later confirmed that higher prompt engineering skills directly predict output quality (Knoth et al., 2024).
My assessment: prompting gets more natural over time, the models get better at guessing intent. But it will still separate people. It’s closer to written communication: everyone can write, but some write so much better that it gives them a structural advantage. That’s also why writing is on the list above. It’s the same skill, just applied to a new recipient.
The narrow top and the broad base
Once the commodity layer is big enough, the field splits in two.
At the base: a large group of people who can all prompt and take part in building. Not because they’re developers in the classic sense, but because the boundary between prompting code and writing it collapses. That’s the broad population of AI-competent employees.
At the top: a small group of people who can do the genuinely hard things. Design new AI systems, build technical complexity beyond commodity level, solve problems that haven’t been solved before. That group gets narrower, not broader.
Left over is an old middle category, the classic developer who built CRUD and standard flows for years, who no longer has a place. That’s where the hardest personal challenge sits.
It’s not an A team and a B team. It’s context.
Some will call it an A team and a B team. It’s not about who’s smartest. It’s about who has a growth mindset and who has a fixed mindset.
The nuance matters. A fixed mindset isn’t character. It’s context. People with 15 years of experience have often been rewarded for being the expert. It’s rational to cling to that identity when it has been useful for an entire career. The point is not that a fixed mindset is a personality flaw. The point is that the reward structure that created it is disappearing.
Which means people who have sat in a role for 15 years can have huge potential to move, if they have a growth mindset. But it takes a willingness to change that hurts the identity first.
The habits that are hardest to drop
The force of habit is hard to change, and it varies from role to role. Three patterns recur across all of them:
Hiding behind artifacts. PRD, ticket, mockup, story point. Having produced something has been enough to justify the salary. When AI can produce every artifact, it isn’t enough anymore.
Believing you “own” a discipline. The designer’s “that’s my area”. The developer’s “only I understand the code”. When everyone can touch everything, territorial thinking collapses.
Measuring yourself by busyness. People who have lived on “I have 17 meetings today and a backlog to get through” face an existential challenge when AI takes over both the meetings and the backlog.
It’s identity tied to artifact, territory and ritual. It’s the hardest thing to let go of, and the most important shift for moving on.
The principle for this layer
The principle is short:
Different skills. Deeper judgment.
It’s not everything. But it’s what gets repeated daily, what can’t be automated, and what differentiates. AI can help with the rest.