Last year, 54% of designers used AI weekly. This year, 91%. Half have shipped AI-generated code to production: product designers, brand designers, and the design engineers who used to own that work. Working prototypes are now an expected output for 43% of teams.
These numbers come from the *AI in Design Report*, published this month by Designer Fund and Foundation Capital. 906 designers, twenty-five-plus interviews, fifty-plus public sources.
The data makes the abstract concrete. The interesting parts are the contradictions inside the data, and what those contradictions tell us about where this all ends up.
That's the piece I want to write.
The middle is going
The shift across every function (engineering, PM, design) has been the same: AI is taking the middle, the body of execution that sits between framing the problem and judging whether the form fits.
In design, the middle is everything from the pixel-correct screen to the wireframe to the design system component to the redlined spec. Until recently, we drew lines inside that body of work. Wireframing felt different from front-end production. Design systems felt different from QA. AI is treating the whole middle as one continuous body of work.
Look at where AI use moved most year over year. Code generation and front-end: +31 points. Wireframing: +27. Design systems and components: +24. UI copy: +18. Design QA and accessibility: +16. Developer handoff: +12. Different kinds of work that used to feel different. All of it climbing at roughly the same rate.
Underneath the numbers is a redistribution of work. 65% of designers say they're doing more PM and engineering work. 40% say PMs and engineers are doing more design work. The lanes are no longer holding. The industry calls this the "block-shaped designer," replacing the older T-shape framing. AirOps calls the new role "Agent Captain." Offsite calls it "systems architect." Different vocabularies, same observation.
What's happening underneath is the bifurcation of the design role. The work that used to take up most of a designer's week is becoming a shared utility that PMs, engineers, and agents can operate together. The most important work of a designer is moving to either end of that middle.
This is where the next eighteen months get interesting.
The two ends that remain: the question and the judgment
When the middle gets compressed, what survives is the work AI can't do. There are two pieces.
The question. Before any artifact gets made, someone has to decide what should exist. That's the work of framing a customer problem precisely enough that an answer is possible. It's the work Karri Saarinen describes when he writes, borrowing Christopher Alexander's framing, that design is the search for a good fit between form and context. AI accelerates the production of form. It doesn't help you understand the underlying problem, and sometimes it actively obstructs that understanding by generating plausible answers before the question is sharp enough to be answered. The question lives upstream of the PRD. It's the work of getting close enough to the customer, the business, and the material to know what the right form even is.
The judgment. Once form gets generated, someone has to decide whether it fits. AI gets a team to roughly 80% of any output, fast. The last 20% is the work that decides whether the result meets the bar, fits the brand, and actually serves the customer in front of you. This is the eye for the gap between competent and exceptional. Spacing that's still wrong. Tone that rings cheap. An animation curve that lands a half-second late. The smoothness of a transition. The weight of a haptic. When the cost of producing design approaches zero, only quality matters. And quality is a taste problem.
The question and the judgment are different skills. They live at opposite ends of the design process. They share one property. Neither one can be automated, because both depend on a human being holding context that no model has access to. The customer in their actual life. The business in its actual strategy. The material in its actual capability.
The middle, the part that used to fill the calendar, becomes shared infrastructure. Designers still curate it, still set the bar for what passes through it, still defend the standard. The execution burden no longer belongs to design alone.
Research is the conduit to the customer
The report mostly treats research as a chapter of the design story. I think the research transition is sharper, and the stakes are higher.
The synthesis work that has traditionally filled a researcher's week, cleaning transcripts, clustering quotes, summarizing themes across a stack of interviews, is exactly the kind of work AI is best at. AI gets faster and cheaper at synthesis every quarter. The function should absorb the speed-up rather than fight it. Synthesis stays part of the job. The part where a researcher earns their seat at the table moves elsewhere.
What gets more important is the work synthesis can't do. Question framing. Foundational research. Live contact with real customers.
The risk underneath this whole wave is something nobody is pricing yet. As synthetic personas, AI-generated user simulations, and model-driven research get cheaper and better, companies will be tempted to do more of their research without ever meeting a customer. The output will look fine. The slides will be plausible. The team will feel productive.
The slow consequence is that the company drifts from the people it serves without noticing. The personas become the customer. The team starts answering questions the customer never asked. The product gets polished, and the gap between the product and the customer's actual life quietly widens. By the time anyone notices, the drift has already become a strategy problem.
Research is the conduit. Real customer research, conducted by real researchers in real contact with real people, is the thing that keeps a company connected to the world it claims to serve. As synthetic alternatives get better, that conduit gets more important.
The companies that treat research as a synthesis cost center will lose that conduit. The companies that treat research as the connection itself will get the strategic upside. Foundational research becomes one of the most valuable seats in the org, because it's the only defense against generated form that solves a problem nobody actually has.
The new shape of the work
If you take the question and the judgment as the two anchors, and treat the middle as shared infrastructure, the operational implications start to write themselves. The report gives us four patterns worth naming.
Higher altitude. The companies furthest along invest in practitioners working at higher strategic altitude. The work moves upstream and downstream from the middle. Both ends get more senior, more opinionated, and more connected to the business.
Internal tool stacks as institutional infrastructure. The Anthropic kit (ideation sandbox, design-system picker, research index, looping PRs, content guardrails). Stripe's ProtoDash and Dante. Ramp's Inspect and Glass. Notion's prototype playground. AirOps' Claude skill library. The most interesting investment patterns in the report are companies treating their internal AI workflows as shared infrastructure, packaged and re-used across the function. External tool choice stays fluid. Internal scaffolding becomes a competitive moat.
No lanes. Cross-disciplinary fluency is the bar across the AI-native cluster (Ramp, Shopify, Anthropic). Designers ship code. PMs run usability tests. Engineers shape product. Research findings show up in PRs. The role descriptions are blurring, and design's job inside that blur is to set the quality standard for work that everyone now helps produce. That makes the design function more important. Someone has to hold the bar when everyone has the capability to ship.
A new hiring profile. The leaders in the report screen for three things in order: AI tool fluency, systems thinking and strategic judgment, and technical fluency. Only 5% deprioritize execution craft. The hire that wins right now is someone with a real opinion about what should exist, the technical literacy to argue with engineers about how to build it, and the AI fluency to operate at the new tempo. That's a different person than the production-layer designer most orgs were hiring two years ago.
None of this is theoretical. The report names the companies doing it, by name. The patterns are visible.
The risk nobody is pricing: the apprenticeship problem
There's one piece of this story that the report touches but doesn't fully draw out, and I think it's the most important thing on the horizon.
For most of the history of product design, the way a junior designer became a senior designer was by doing a lot of middle work. The pixel-correct screen. The seventeenth wireframe variant. The redlined spec. The flow that needed three rounds of usability testing. Most of it was unglamorous. Nobody designed that path as a curriculum. It was what the job happened to require, and the years of repetition built the eye and the instinct that eventually became judgment.
The same data points that prove the middle is being compressed describe the disappearance of the work that used to make juniors into seniors.
Junior UX postings have been below a thousand globally since early 2025. Only 7% of teams have added AI-focused roles. Peer learning more than doubled (24% to 70%) while formal leadership recommendations halved (32% to 16%). The learning system is bottom-up, ad hoc, and tilted toward whoever can teach themselves fastest.
Today this looks like a curriculum problem. In 2028 or 2029 it shows up as a hiring crisis.
The judgment layer that the rest of this piece celebrates is a stock. Stocks get depleted unless something replenishes them. Every senior practitioner alive today came through some version of the apprenticeship that is now being automated away. Without deliberate investment in producing the next generation of judgment-layer practitioners, the pool runs down. The first sign will be senior roles that go unfilled for a year. The second sign will be the quality of work at companies that didn't invest in growing their own.
I don't have a clean answer for this one. Nobody does yet. The orgs that get it right will probably look like a return to deliberate craft training, paired apprenticeship, time-boxed learning environments where the junior is allowed to do work that AI could do faster, because the work matters for what it produces in the practitioner.
Naming the problem is the first step. Pretending we have ten years to solve it would be a mistake.
Where this leaves us in the next 18 months
Four forecasts for design and research between now and the end of 2027. Each is something I'd put real money on.
1. Internal AI tool stacks become a competitive moat. External tool churn continues. Figma, Cursor, Claude, Lovable, Paper, Replit, Bolt. The market sorts itself out, and no winner takes everything. The orgs that pull ahead are the ones investing in internal scaffolding the way the report's leaders already are. Skill libraries, design-system-aware prototyping tools, research indexes, content guardrails. Treated as infrastructure. Owned by the function. Compounding over time.
2. Foundational research becomes one of the most valuable seats in the room. Synthesis stays cheap and gets cheaper. The practitioner who can frame the right question, get in front of real customers, and bring back a finding the model couldn't have produced becomes scarce and valuable. This person is a specialist with deep customer access, a real point of view, and the discipline to keep the company connected to the human being on the other side of the product.
3. The Agent Captain role solidifies, and design becomes the function that owns the quality bar. Right now it's a phrase a couple of companies use. In eighteen months it's a hiring profile. The function that sets the standard for work that everyone now helps produce becomes formally responsible for the quality bar at the org level. Design as the standard-setter moves from cultural preference to organizational mandate.
4. The apprenticeship problem starts showing up as a hiring crisis in 2028-2029. Medium confidence on timing. High confidence on direction. The companies that didn't invest in growing seniors from juniors during this window will find themselves bidding against everyone else for the same shrinking pool of judgment-layer practitioners. The companies that did invest will have an unfair advantage that compounds for the next decade.
The practitioners and orgs that win the next two years will be the ones who reorganized around the two ends of the work that AI can't do, and protected the conditions under which both can still be learned. Tool adoption alone has already become table stakes. The question, asked with real customer contact and real strategic judgment. The judgment, exercised at the last twenty percent where taste decides everything. Everything else is shared infrastructure now.
That's where design and research are headed. The companies that see it clearly have an opportunity that will look obvious in retrospect, and is invisible to most teams right now.