Which is a different problem, because tasks are not spread evenly across a career. Production fills 46% of a junior designer’s week and 4% of a lead’s. This page models that gap rather than arguing about it — twenty-nine tasks, five levers, and whatever conclusion the numbers support once you set them yourself.
11%
of a junior week the model puts on AI today
57
entry-level work left under the compression (100 = today)
128
designer headcount under design everywhere
Almost every claim about AI and design cites real evidence. The disagreement is not usually about facts; it is about which facts get left out. Here are two sets that are both well-sourced and point in opposite directions.
Any account of design work that cannot hold both at once is incomplete.
UX/UI postings in India grew roughly 40% between 2024 and 2026, led by Bengaluru, Hyderabad and Delhi NCR.
India runs about a 4:1 developer-to-designer ratio. Most software here ships without a designer anywhere near it.
Global capability centres employ 2.36 million people across 2,117 centres, with $98.4B in FY2026 exports and roughly 510,000 hires projected for 2026. Bengaluru holds about 36% of that talent.
82% of design leaders say their need for designers has increased or stayed the same.
US projections still show growth for digital design: +7% for web developers and digital designers, 2024–2034.
Indian IT fresher intake fell from about 600,000 in FY2022 to roughly 120,000 in FY2025 — an 80% drop in three years.
EY estimates entry-level IT roles have already declined 20–25% from automation.
About 55% of Indian IT companies reduced entry-level hiring, against 14% for senior roles.
US graphic design postings fell 33% in 2025 after 12% in 2024, and the WEF moved the role from moderately growing to one of the fastest declining.
Senior and generalist roles are recovering faster than entry-level ones, which remain scarce and heavily contested.
Note what the second column mostly describes: software engineering, not design. That is deliberate. Indian IT services has already run a version of this experiment at a scale design has not, and it is the closest thing to evidence we have about what happens next.
“Will AI replace designers?” cannot be answered because “designer” is not a unit of work. It is a bundle of tasks that changes shape as a career progresses, and the bundle is what AI acts on.
Two designers with the same title can hold almost disjoint task mixes. One spends most of a week producing screens; the other spends it deciding which screens are worth producing. A forecast at the level of the job title averages those two people together and tells you nothing about either.
So this model works one level down. It scores each task on three things: how automatable it is, how expensive it is for a human to verify, and how much of it is irreducibly judgment. Then it asks how much of a week each seniority band spends on each task.
The middle term matters more than it first appears. A task that a machine can do but nobody can cheaply check does not get automated — it gets produced by a machine and then re-done by a person, which is worse than either. Verification cost, not capability, is what has held adoption at roughly a third of core design work while 72% of designers use generative AI somewhere and only 32% trust it enough to ship.
Twenty-nine tasks that make up product design work, each scored on how automatable it is and how expensive it is for a human to check the result. Verification is the brake: a task can be trivial for a machine and still stay human if nobody can cheaply tell whether the output is right.
Switch seniority to resize the bubbles. The scores do not change — only which tasks fill the week.
2–5 years. Still making most of the artefacts.
Hover or tab through a bubble to see why it scores the way it does. Bubble area is that task’s share of the selected role’s week.
| Task | Automatability | Verification cost | Share of week | Absorbed by AI |
|---|---|---|---|---|
| Participant recruiting & scheduling | 85% | 20% | 3% | 15% |
| Running user interviews | 35% | 70% | 4% | 4% |
| Research synthesis & tagging | 75% | 65% | 4% | 9% |
| Competitive & heuristic audit | 80% | 35% | 2% | 12% |
| Survey design & analysis | 70% | 45% | 1% | 10% |
| Problem framing | 25% | 85% | 4% | 3% |
| Scoping & prioritisation | 30% | 75% | 3% | 3% |
| Defining success metrics | 45% | 70% | 2% | 5% |
| Stakeholder alignment | 10% | 60% | 4% | 1% |
| Concept sketching | 60% | 40% | 5% | 9% |
| Divergent variant generation | 90% | 35% | 4% | 14% |
| Visual & art direction | 55% | 75% | 4% | 6% |
| Wireframing | 85% | 25% | 6% | 14% |
| High-fidelity UI | 80% | 50% | 9% | 11% |
| Responsive states & edge cases | 85% | 30% | 5% | 14% |
| Icons & illustration | 88% | 20% | 2% | 15% |
| UX copy & microcopy | 85% | 50% | 3% | 12% |
| Interactive prototyping | 75% | 35% | 4% | 12% |
| Redlines, specs & handoff docs | 90% | 20% | 2% | 15% |
| Asset preparation & export | 95% | 10% | 1% | 17% |
| Component building | 80% | 55% | 4% | 11% |
| Token & library maintenance | 75% | 50% | 3% | 11% |
| System documentation | 85% | 30% | 3% | 14% |
| Consistency & QA audits | 88% | 25% | 2% | 15% |
| Dev handoff & pairing | 40% | 50% | 5% | 6% |
| Design review & critique | 30% | 80% | 4% | 3% |
| Usability testing | 55% | 60% | 4% | 7% |
| Roadmap negotiation | 15% | 65% | 2% | 2% |
| Mentoring & growing designers | 20% | 70% | 1% | 2% |
Percentages are the share of each task AI absorbs at the current settings. Bars are the share of the week the task fills.
Five levers decide how much of that atlas moves. Start from a named scenario or set them yourself; the outputs update as you go. None of this is a forecast — it is a way of checking whether a belief about design work is internally consistent.
The levers where the best available evidence puts them in late 2026.
How far along the capability curve design tools get. Scales how much of each task’s intrinsic susceptibility is actually reachable.
Share of design work actually done with AI in the loop — not the share of designers who have tried it.
How much AI output teams ship without a human pass. Interacts with each task’s verification cost — this is the model’s main brake.
Does cheaper design mean more design gets made, or the same amount with fewer people? Expressed as a multiplier on total design work.
How deliberately teams rebuild an entry path once the tasks juniors learned on are automated.
The tick under each track marks where the evidence puts that lever today. Sources and reasoning are in the method note at the end.
Week absorbed by AI
9%
of a mid designer's week, at these settings
Entry-level work left
100
index vs today (100 = today) · no change
Designer headcount
104
index vs today · +4
Demand needed to hold flat
1.10×
design work must grow this much to keep headcount steady
At these settings, AI absorbs 9% of a mid designer’s week. Design headcount grows to 104 against today’s 100, because demand rises faster than output per designer. The pool of work a newcomer could learn on sits at 100, and juniors make up 28% of a team.
Both bars use the same scale. The dashed remainder is freed capacity — what happens to it is what the demand lever decides.
| Phase | Today | This scenario |
|---|---|---|
| Discovery | 14 | 13 |
| Definition | 13 | 13 |
| Exploration | 13 | 12 |
| Production | 32 | 28 |
| Systems | 12 | 11 |
| Delivery | 16 | 15 |
Identical levers across all four bands. The gradient comes entirely from task mix — juniors hold the work that automates first.
| Seniority | Absorbed |
|---|---|
| Junior | 11% |
| Mid | 9% |
| Senior | 7% |
| Lead | 5% |
Each of these is a coherent set of lever positions rather than a prediction. What separates them is not arithmetic but belief — about whether design was rationed by cost, and about how much unchecked output a team will tolerate. Load one into the explorer above to see it run.
The levers where the best available evidence puts them in late 2026.
Tools keep improving but stay assistive. Designers work faster; teams keep reviewing everything.
Capability and adoption both rise sharply while demand for design stays roughly flat. Teams get smaller and more senior.
The cost of a designed artefact collapses, so far more things get designed. Demand expands faster than automation absorbs capacity.
Production is almost entirely automated, but verification stays expensive. A small number of highly paid judgment roles; the middle hollows out.
Indian IT services was built as a pyramid. Hire freshers in the tens of thousands, train them for three to six months, deploy them on routine maintenance and development, and promote the ones who stick. The routine work at the base was not a side effect of the model — it was the model. It paid for the training and it was where judgment got built.
Coding assistants automated that base layer specifically. Fresher intake fell roughly 80% in three years. Entry-level roles are down 20–25% by EY’s estimate. The asymmetry is the tell: about 55% of firms cut entry-level hiring against 14% at senior level, and the stated preference shifted to smaller teams of people who can contribute immediately.
Design in India has the same shape. Production-heavy delivery, a large junior tier, and a career ladder whose bottom rungs are exactly the tasks the atlas scores as most exposed — 16 of the twenty-nine.
What makes design different is the demand side. Software engineering was well-supplied when automation arrived; design is not. A 4:1 developer-to-designer ratio means there is a large backlog of work that never got a designer at all. That is the case for growth, and the model takes it seriously — the design everywhere scenario puts headcount at 128 against today’s 100.
But run that scenario and look at the entry-level figure, not the headcount one. It still lands at 71. Growth in the profession and contraction in its entry rung are not alternatives. The same automation produces both, because the work that expands is judgment work and the work that disappears is what juniors were doing to earn the right to do it.
This is the part of the model that generalises least gracefully and matters most. There is a circularity in the atlas worth sitting with: mentoring is among the least automatable tasks in design, and it is the one that depends on there being juniors to mentor.
These are not a list of comforting things to tell designers. They are the tasks that stay below 15% absorbed even at maximum capability, adoption and trust — because judgment load caps them outright, or because verifying the output costs as much as producing it.
Organisational politics is not a text-generation problem. Someone has to be accountable in the room.
judgment 90% · verification 60%
Trading scope against time with people who have competing incentives. Almost entirely social.
judgment 90% · verification 65%
Note the circularity: this is among the least automatable tasks, and it depends on there being juniors to mentor.
judgment 90% · verification 70%
The highest-stakes verification in design: a wrong frame produces confident, coherent, well-crafted work that solves nothing, and it stays invisible until launch.
judgment 85% · verification 85%
A model will tell you the contrast ratio is wrong. It will not tell you the whole direction is a dead end and be willing to own that call.
judgment 85% · verification 80%
Taste is expensive to verify — you cannot unit-test whether a direction is right for a brand.
judgment 80% · verification 75%
Deciding what not to build depends on context that mostly lives in people’s heads.
judgment 75% · verification 75%
One structural observation from the atlas: there is almost no such thing as a design task that is hard to automate but cheap to verify. The two scores are strongly anti-correlated, which is why the “assisted” corner of the matrix is nearly empty. Design work tends to be either mechanical and checkable, or judgment-loaded and expensive to check. Very little sits in between, and that is what makes the middle of the career ladder feel thin.
The model is deliberately simple enough to argue with. For each task it computes how far capability reaches into that task’s intrinsic susceptibility, how much of that reach adoption puts to work, and then applies a gate: trust raised to a power set by verification cost. A task that is cheap to check passes through at close to the trust level; a task that is expensive to check is crushed. Judgment load then caps the result. Freed capacity meets the demand lever to give headcount.
The employment figures are published and cited below. The lever defaults for adoption and trust come from Figma’s 2026 survey work. The demand default reflects India’s measured undersupply. The ladder default reflects what Indian IT services actually did, which was not rebuild.
The task scores are not measured. No public dataset scores design tasks for automatability, verification cost and judgment load, so all three are my estimates, applied consistently across the twenty-nine tasks rather than derived from data. The share-of-week figures per seniority band are also estimates, built from a decade in product design and sanity-checked against the shape of Indian services delivery. Each role’s column sums to exactly 100 points, which is a modelling convenience, not a finding.
The adoption default of 40% deserves a specific caveat. The two figures it sits between — 72% of designers using generative AI somewhere, 31% using it for core design work — are both shares of designers, not shares of work. The model needs the latter. 40% is an interpolation and nothing firmer.
Figures were compiled in September 2026 from published reports and reporting on them, and they will age. If you think a task score is wrong, it probably is — the useful version of that disagreement is a specific score and a reason. Tell me.
Ten years of product design, currently in Bengaluru. I mentor designers earlier in their careers, which is the part of this page I have the most stake in.
hello@asifasharaf.com →