Automation is taking tasks juniors learned on. Here are seven tasks that put the practice back so we keep human element of user experience and build the bench for 2030.
Many teams I know have stopped assigning transcript tagging because now AI does it. Their platform of choice returns sessions tagged, themed, and clipped transcripts before anyone opens them, and does it well enough.
Nobody argued, and nobody wrote down what else that task did.
Tagging transcripts is how researchers and designers learn that people say “it’s fine” when they mean they gave up, which AI never picks up. Marking redlines is how a new designer learns which spacing decisions the system made and which are open.
That work was never worth much as output; it taught the skills needed to repeat in the future so they learned the craft. It was worth a great deal as practice, and the two arrived bundled: cheap judgments under supervision, where being wrong cost nothing.
The numbers of this change are already in.
Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, working from payroll records through June 2026 in Canaries in the Coal Mine, find employment for 22 to 25 year-olds in the most AI-exposed occupations sitting 19% below where it would be had it kept pace with less-exposed peers, and no comparable gap for experienced workers.
The adjustment runs through hiring, not layoffs. And that’s risk that our field doesn’t have the skills to move forward in the future as more than a competency.
Their proposed mechanism is the part design leaders should sit with: AI substitutes for codified knowledge and complements the tacit kind, built by doing, and a lot of designers aren’t doing a lot of doing right now.
I cared about building for that future that I mentored designers continuously during the 2010’s, spoke at University of Washingston and taught part-time as part of General Assembly.
I know, I know, but at least my previous students are doing well.
This is an important moment for all of us and we need to take action. Here are seven moves we can do to put the put the practice back and save the field for the design and research generations to come.
Inventory the tasks automation removed
Start with an inventory, and keep it boring. List every task that left the junior queue in the last 18 months, some of which I actually wrote a book about, uxGPT.
On most teams the list looks like this:
- Transcript tagging and synthesis. Sessions come back tagged, themed, and clipped before anyone opens the recording.
- First-pass wireframes. A prompt returns four layouts in the time it took a junior to sketch one.
- Screener drafts. Recruiting tools generate screening questions straight from a study goal.
- Competitive audits. A model walks the competitor’s flow and summarizes the patterns.
- Redline sheets. Design tools export spacing, type, and token values without a handoff document.
- Meeting notes and status summaries. Written and posted before the call ends.
Then add a second column: what did that task teach the person doing it? A competitive audit is worthless as a deliverable. Building one is how a designer learns that competitors solved onboarding five ways, and that four of them are bad for reasons you only see up close.
Maze’s Future of User Research Report 2026 found 69% of research professionals now use AI in at least some projects, up 19 points in a year, concentrated in transcription, synthesis, and drafting questions — where a junior used to start.
Ethan Mollick makes the case for doing this on purpose in Choosing to Stay Human: the defaults for what work gets handed to AI are being set right now, mostly without planning, and they will be hard to reverse once a generation of workers has built habits around them. The inventory is how you set yours deliberately.
Run it with your seniors in the room.
They can tell you the screener draft was where they learned to write a question that doesn’t leak the answer. Without those stories you have a list of chores, and you will conclude that nothing was lost.
A task can be worthless as output and foundational as training.
That pairing is the bundle: a lesson riding along inside work somebody had to do anyway, invoiced to nobody.
Look hardest at those, because the business case for automating them is airtight and the training case never got written down.
Assign juniors real decisions with guardrails
Surgery ran this experiment first for a pretty long time.
Matt Beane spent two years watching robotic and open operations for Shadow Learning: Building Robotic Surgical Skill When Approved Means Fail. Open surgery needed the resident’s hands; on the robot, the attending works alone while the resident watches a screen.
A national survey of recent United States general surgery graduates, published in 2025, found only 37% reported high autonomy in robotic cases, against 89% in open ones. Same programs, different console.
The analogy has a limit, since design has no credentialing gate. The mechanism is what transfers: when the expert can finish without the novice’s hands, the novice stops learning.
The senior prompts, evaluates, and revises while the junior watches the screen and takes the notes.
Everyone is busy. One person is learning.
So give them the call: which direction goes to the review, which participants come out of the recruit. Julie Zhuo’s The Looking Glass reduces design levels to a line: the more senior the designer, the more abstract the problem.
Pick reversible calls — I call them two way doors, which is a mental model Jeff Bezos has— and let them stand.
Beane’s The Skill Code names challenge as the missing ingredient, and challenge means being able to get it wrong, safely.
Give juniors real work where mistakes are cheap
Researchers already drew this line. User Interviews surveyed 150 researchers in May 2026 for its State of Synthetic Users report: nearly all use AI somewhere, and only 8% regularly use synthetic participants. The field automated everything around the conversation and kept the conversation.
Keep juniors on the near side. A live session is the one rep with no substitute, and the one most quietly reassigned. Teresa Torres argues the team building the product should interview a customer every week, and that the habit only holds when booked.
One session a month with the junior moderating costs a fumbled follow-up, not the recruit, and do it without AI other than the transcription.
Between sessions, juniors need work where being wrong is survivable: internal tools, admin screens, the settings page, bulk actions the operations team will complain about at length. The reflex is to point a tool at them, the most automatable work you have.
That is your training ground, and once it is gone you are asking people to learn on the checkout flow.
Jared M. Spool built Center Centre, a design school, around the same principle. In Interview: Jared Spool on Center Centre, Real Projects, and the Future of UX Education, he describes hiring managers’ near-universal complaint that graduates were not ready to do the work, and a curriculum built on real client projects in response.
With this approach that’s been around for centuries, the Labor Department reports 91% of registered apprentices retain employment after their programs end, on the strength of graduated responsibility on real work.
Make verification of AI output the deliverable
Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real tasks for The Impact of Generative AI on Critical Thinking, presented at CHI 2025.
Two findings run in opposite directions.
- Higher confidence in the tool predicts less critical thinking.
- Higher confidence in your own ability predicts more.
A junior has the least self-confidence on the team and the most exposure to polished machine output, the worst possible combination. Left alone they defer, and deference looks like competence right up until the day it doesn’t.
This is automation bias, an entry I maintain over at UX Guidelines, arriving early in a career instead of late.
Dan Maccarone rebuilt his studio’s process around these tools and came back with a warning that has nothing to do with speed. In Never mind the prompts, here’s the thinking, he names the failure mode: high-speed mediocrity, work that shows up faster precisely because nobody slowed down to think.
His sprints stayed five days long and what changed was how much reasoning the deliverables carried, and that is the part a junior is least equipped to notice is missing.
To do this, make the critique the thing they hand in.
Not “summarize these eight interviews,” but: here is the synthesis the tool produced, find the two claims the transcripts do not support, and bring the timestamps.
Real human in the loop with a judgement call.
Erika Hall ends Let’s Stop Doing Research with a three-word instruction that could headline the assignment: believe and verify. Her larger point is that inquiry is part of design rather than a phase beside it, which is exactly the habit this exercise is trying to build before anyone thinks of it as research. Grade the catch, not the polish, and say so in the review where the team can hear it.
Verification is a skill, and it only builds when somebody (and not some machine) is accountable for the verdict.
Do this weekly for a quarter and you have a researcher who reads output with suspicion. Skip it, and you have someone who forwards it.
Where we have to get to is this: Higher confidence builds taste and judgement, and that takes practice. A lot of it.
Give juniors the AI evaluation work
AI features created a new layer of design work, and nobody has claimed it.
Every AI feature needs someone to write the specification the model works from, assemble the cases it is tested against, categorize what it gets wrong, decide what happens when it fails, and set the ship threshold. Right now that happens in fragments, done by whichever senior has an afternoon.
Hand it over by giving a junior designer the evaluation set for one feature: 30 real inputs from actual users, the output you would accept for each, and the failures sorted into named categories.
Nobody can build that without first deciding what good looks like.
That is judgment, written out as a deliverable.
OpenAI’s GDPval benchmark has expert graders compare model deliverables against work from professionals averaging 14 years of experience, on tasks that took those experts about seven hours each. The best models were rated as good as or better than the human deliverable in just under half of cases. The other half is the job, and knowing which half you are looking at is a skill somebody has to build.
This pairing already has a public example. Josh Clark, with three decades in the field, wrote Sentient Design with Veronika Kindred, a designer much earlier in her career, and in Beyond the Happy Path they describe the gap between their vantage points as the point.
- She questions practices built for static interfaces.
- He brings the history of hype cycles.
- Both halves are needed to judge what an AI feature should do.
Two guardrails. It has to carry a decision, so name who says ship or hold, and it has to attach to a feature real users touch.
Assign each junior a business domain
Generalist screen-making is the most substitutable work in the building. Knowing how the business runs is the least.
MIT’s Project NANDA reviewed more than 300 enterprise AI deployments and found 95% of generative AI pilots producing no measurable impact on profit and loss, as Fortune reported in August 2025. The 95% is contested and the report was never peer reviewed, so take the direction rather than the decimal.
The direction is the useful part.
The researchers blame a learning gap rather than model quality: tools that never adapt to the workflow that were bought to change. Closing it takes somebody who knows that workflow well enough to say what adapting would mean.
Make that somebody your junior. Assign a domain, not a component, and have AI equip them with a starting set of questions to remind them this is a business. Their job is to become the person who can explain how a claim gets denied, why the underwriter overrides the recommendation twice a week, and what happens when customs paperwork lands late.
None of that is written down, and is hidden in the white space of the job.
Nobody automates knowledge that was never written down, it sits with the nine people in operations who do the work.
That means standing time with the business, not a research request filed through a manager. Four hours a month with claims adjudicators, and after six months a junior knows the domain better than the senior who spent those months in review meetings.
An honest conversation I had with a designer recently — in their environment, designers weren’t being hired back because they didn’t understand their domain enough to make a business impact, while we should have been learning it all along. And then I asked them directly if the designers should be hired back, and he said “No” and he taught UX.
We can’t support our own because we haven’t trained them to function in the businesses they work in. This is a way to fix it.
One guardrail: Pick a domain with a real owner on the business side who will take the meetings, or you have assigned homework instead of a beat.
Put training on the calendar and assign a dollar amount to it
Every move above costs senior time. A junior making the call is slower, and the review takes longer than doing it yourself. Pretending otherwise is how these programs die in month three.
So put it in the estimate.
Unbundled training is a purchase, and purchases have prices. Start at 10% more senior time on any project with a junior decision owner, a guess rather than a benchmark, then correct it with your own numbers.
Mentorship is the first thing cut when a date slips.
Peter Merholz, writing in February 2026 that design leaders are failing the next generations, points out that hiring managers prefer senior designers by better than two to one, and that plenty of companies now assume tools will cover the junior tier.
He also notes the pattern predates the current tools.
AI did not invent this problem — we’ve had it for years, quite honestly— and it removed the last excuse for not fixing it.
Notice how quietly it happens. The Stanford decline runs through reduced hiring rather than separations, so no meeting ever decides to stop training the field. A requisition gets written a level higher than last year’s, and again the year after.
Conversely, Nielsen Norman Group’s State of UX 2026 describes the other side: junior supply outpacing junior roles, and openings asking for judgment rather than artifacts.
Nobody stocks a market for judgment you declined to build.
Hire only seniors for a decade and you will bid against every company that made the same call, for a pool nobody grew, including yourself.
Training is now a budget line
None of this requires slowing down the tools, and none of it is nostalgia for busywork. Tagging transcripts by hand was tedious. The tedium was the price of a lesson, and now the lesson needs a different delivery mechanism.
What changed is that the training used to be free. It arrived bundled with work somebody had to do anyway, which meant no leader ever had to argue for it, defend it in a planning meeting, or write it into a number. That bundle is gone. Judgment now has to be bought on purpose, in senior hours, on projects with real stakes, or it does not get built at all.
The decision in front of every design leader is small and immediate. One field in a template. One inverted session a month. One line in an estimate.
None of it shows up in this quarter’s numbers, and all of it shows up in who is qualified to run your practice in 2032.
The next generation of judgment comes from the same place it always did: people making real calls, early, while somebody experienced is close enough to catch the fall. The only difference now is that you have to schedule it.
Where you can start this week
- Run the two-column inventory. List every task that left the junior queue since early 2025, and write what each one taught next to it. An hour of work, and the worry becomes a list.
- Add a decision owner field to your project template. Every project names one non-senior person accountable for a real call, in writing, before work starts. If the field is empty, the project is not staffed for training.
- Put training hours in the estimate. Add a line, defend it in planning, and report against it. A cost you can name survives a schedule slip, and one you absorb quietly does not.
Those three make the rest possible. The sessions, the domain, and the evaluation set all need somebody accountable and hours nobody has to steal.
How we save entry-level UX work that teaches craft in the time of AI. Our field depends on this. was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.
