AI is widening roles before it narrows them

A design leader I’d believe is right now writing four job descriptions. One for the designer who works on AI products. One for the designer who builds the AI itself. Two more for the paths in between. The logic comes straight from a sharp NNGroup piece arguing that AI has split the generalist designer into four distinct jobs, each with its own pipeline, ladder, and pay band. So you hire for the endpoints. You build the ladders now.

I think that’s the wrong move this year, and the reason is hiding in plain sight in what practitioners are actually doing day to day.

Look at three unrelated corners of tech right now and you see the same thing. Not roles narrowing. Roles swelling.

In design, Replit’s Design Mode lets a designer wire up Stripe and stand up backend database logic while still in the prototype. “Convert to App” turns the mockup into something that runs. The designer isn’t retreating into a narrower craft. They’re absorbing infrastructure work that used to belong to an engineer.

In engineering management, the coordination layer is thinning. Gallup’s manager engagement dropped from 27% in 2022 to 22% in 2026, and managers no longer hold any engagement edge over the people they lead. Meta and Amazon made the flattening official. But the management skills didn’t vanish. They migrated down. ICs are now expected to do task decomposition, delegation, and structured feedback, because orchestration capacity, not raw execution, is the scarce resource. IC-to-manager ratios are rising for that reason.

In production engineering, agents brought a whole new layer of work with them. At PostHog, running agents against a codebase of 8,984 files and 1.6 million lines meant building 13-plus guardrails, writing CLAUDE.md context files, and designing subagent delegation patterns. Context engineering is now part of the job. Nobody specialized into it. It got absorbed.

Three domains, no connection to each other, same pattern. Individuals are taking on more adjacent work, not less. Designers do infrastructure. ICs do management. Engineers do context engineering. That’s not fragmentation. That’s compression.

The specialization thesis confused the destination with the road

The four-paths argument might be right about where the labor market ends up. It’s wrong about how anyone gets there.

The NNGroup framework describes a destination. Stratified specialties, separate hiring funnels, the works. Today’s evidence describes the journey. And on the journey, people are moving in the opposite direction from the destination. They’re widening before anything narrows.

The two get conflated because both are real. The endpoints may genuinely be distinct jobs someday. But that’s not an instruction for how to staff a team in 2026. Confusing the map of where the market settles with the route people take to get there produces a specific, expensive mistake: hiring for the endpoints before the road has been walked.

Compression isn’t the opposite of specialization. It’s the on-ramp.

Now the part that flips the whole frame.

Compression may be the prerequisite for specialization, not its enemy.

Think about what it would actually take to specialize into “designing for AI agents,” NNGroup’s third path. You can’t do that work until you understand what agents need from a system. You learn that by absorbing infrastructure and system knowledge first. Which is exactly the absorption Replit’s Design Mode is pushing designers through right now. The compression is the training.

Same shape in management. You can’t have AI-native engineering leadership until ICs have internalized the decomposition and delegation patterns that used to live a layer up. They’re internalizing those patterns now, by doing the work. The specialist leader of the future is being built out of the compressed generalist of the present.

So the path from generalist to specialist doesn’t run in a straight line. It detours through a compression phase. People absorb adjacent layers, build cross-domain intuition, and only then stratify into specialties that are informed by that wider exposure. Specialization emerges from the experience. It doesn’t get assigned by an org chart.

What this changes for how you hire

The specialization thesis is right about the endpoint and wrong about the transition. That distinction has teeth.

If you build narrow ladders now, hire for incommensurable specialties now, split the team into four pipelines now, you’re optimizing for a state nobody has reached yet. You’re staffing the destination and skipping the road. The people who’d fill those narrow roles credibly don’t exist yet, because the experience that produces them is still being lived.

The near-term talent you actually need is the opposite of narrow. It’s compressed generalists. People who can absorb an adjacent layer, the designer who’ll pick up backend logic, the IC who’ll learn to orchestrate, the engineer who’ll own context engineering, and build the cross-domain intuition that real specialization later grows out of. Hire for absorption capacity, not for a specialty that hasn’t crystallized.

This doesn’t mean the four paths are wrong. It means they’re early. Build for the journey your people are on, not the destination a framework drew for them.

The question that decides everything

There’s one assumption underneath all of this, and it’s worth saying out loud. I’ve been calling compression a phase, something with a beginning and an end, after which specialization sets in.

But what if it doesn’t end?

AI capability keeps rising. Every jump pushes another adjacent layer down into the individual role. If new layers arrive faster than people can consolidate the last ones into a stable specialty, compression stops being a transition and becomes the permanent condition of the job. You’d never reach the four paths. You’d just keep absorbing, forever, with the destination receding as fast as you walk toward it.

I don’t know which it is yet. But it’s the right question, and it beats the one most orgs are asking. The question isn’t “which of the four specialties do we hire for.” It’s “is specialization even the state we’re heading toward, or have we mistaken a moving condition for a place we arrive.”

Get that wrong and you’ll build ladders to a floor that was never poured.