AI in the design process: how artificial intelligence transforms my work.

AI as the designer’s ally (and as a source of new responsibilities).

A few months ago I reviewed a flow I’d generated with an AI tool. The screen looked fine. The hierarchy seemed right, the copy was convincing, the components fitted together. I accepted it and moved it into the sprint. Three days later, the developer asked me why the primary button had a 3.2:1 contrast ratio (below WCAG’s AA minimum), why the form didn’t associate labels with inputs, and why the error message assumed a language that wasn’t the user’s.

I hadn’t designed badly. I had accepted without evaluating. And that is, in 2026, the most common and most dangerous trap of designing with AI.

Artificial intelligence doesn’t replace the designer’s judgement: it amplifies it, accelerates it, and above all, puts it constantly to the test. AI produces plausible screens in seconds, and the speed at which they land on your desk can overwhelm the critical muscle that makes them acceptable.

Over the last few years, AI has moved from being an experimental tool to being a structural layer in the design process. It doesn’t just change how we work; it changes what we design, under which legal constraints, and who answers when something goes wrong. That last point is the one that gets discussed least, and it’s the one that has changed most.

Integrating it isn’t a matter of efficiency. It’s a shift in the nature of the work and in the responsibility assumed when doing it.

The thesis: from the designer who makes to the designer who answers.

For decades, the designer’s job was mainly to produce: screens, flows, components, deliverables. In 2026, when a model can produce 40 variants of a screen in two minutes, the designer’s value has shifted to three territories AI cannot occupy:

Framing the right problem. Deciding what should be built before building it. AI accelerates execution; it doesn’t know what to execute.
Evaluating what AI produces. Judgement, taste, contextual sensitivity. AI generates the plausible; the designer distinguishes the appropriate.
Answering for the result. AI is neither a legal nor a moral agent. When a generated screen fails WCAG, discriminates against a group of users, or misleads, the responsible party is still human. In most cases, it’s you.

This third point has changed the most and gets mentioned the least. AI displaces work but doesn’t displace responsibility. And in 2026, with the EU AI Act fully applicable from August, that responsibility has names, articles and concrete fines.

Redefining the design process.

AI doesn’t optimise the existing process: it reconfigures it phase by phase. Every phase gains capacity, but every phase also gains its own traps.

Augmented discovery (and the risk of plausible synthesis).

In discovery, AI lets us process volumes of qualitative and quantitative information that used to demand weeks of work:

Automatic synthesis of interviews and transcripts.
Pattern detection in feedback at scale.
Competitive analysis across hundreds of products.
Hypothesis generation based on aggregated data.

This drastically reduces analysis time. But it introduces a problem I’ve started calling plausible synthesis: the model produces a coherent summary, well written, with categories that seem reasonable… and that don’t correspond faithfully to what users actually said. The summary sounds like research, and that’s worse than a bad summary, because it lowers the reader’s defences.

Three concrete protections I apply today:

Mandatory traceability. No AI-generated quote or insight enters my synthesis without a pointer to the minute/second of the audio or the line in the transcript. If the AI can’t cite the source, the insight doesn’t exist.

Manual sampling. Of the total number of synthesised interviews, I manually review at least 20% in full to check whether the synthesis omitted important nuance or amplified the noisiest parts. In a recent case, the AI had summarised “users want more customisation options” for me; real listening showed “users want fewer options but better explained.” The difference was in tone, not in words, and the model missed it.

Separation between data and interpretation. AI can classify and cluster; causal interpretation I do myself. “38% mentioned confusion at checkout” is data. “This is because of a lack of visual feedback” is my hypothesis, and I mark it as such.

Speed without judgement isn’t research. It’s fiction in the format of a report.

Expanded ideation (and the homogenisation problem).

Generative AI widens the exploration space: dozens of concepts in minutes, UI variations, copy, structures, exploration of extreme scenarios.

But there’s a consequence becoming increasingly visible: when we all use the same models trained on the same data and prompted in similar ways, we produce the same designs. AI-generated onboarding screens tend to look alike. Financial dashboards look alike. SaaS landing pages look alike even more. Homogenisation isn’t hypothetical; it’s the default state.

This has two practical implications:

Differentiation goes back to human judgement. The initial 60% of an AI-generated screen is a commodity product today. Competitive value sits in the remaining 40%: specific art direction, brand voice, the non-obvious decisions, the details the model wouldn’t have proposed. That 40% is what users notice and what justifies the role.

The prompt is also a design brief. A generic prompt produces generic design. A prompt with concrete visual direction, clear references, system tokens, brand constraints and explicit decisions produces something different. Two-line prompts give two lines of value.

More options don’t mean better decisions. The designer’s value has shifted from “creating” to directing, selecting and discarding with judgement.

Prototyping with code (and the end of the handoff).

This is probably the deepest change in the process. With tools like Cursor, Claude Code, v0 or the generative capabilities embedded in Figma, the gap between design and code has shrunk to almost zero for certain product types.

What used to be:

designer produces mockup → engineer interprets → handoff → iteration → final approximation

Can now be:

designer describes intent → AI generates functional component → designer iterates on the real product → engineer reviews, stress-tests, takes to production

This changes several things at once:

We prototype with the product, not with its representation. Validating an onboarding flow in a clickable Figma prototype isn’t the same as validating it in a real app the user can use for a week. The quality of the feedback changes radically when what’s being tested is functional.

The traditional handoff blurs, but doesn’t disappear. AI-generated components need review of state, data handling, robust accessibility, performance, architecture. There’s still an engineer there, and they’re still making decisions a model can’t make well. What changes is that the starting point of that conversation is no longer an image, it’s imperfect but real code.

A designer who can’t read code is losing intervention capacity. And a real tension appears here: the “AI-native” designer of 2026 has at least basic code literacy. Not a developer, but someone who can read, modify with help from the model, understand what a component does and decide what should be changed. Without that, they depend blindly on the output.

Continuous validation (and the temptation not to observe).

AI also impacts validation: automated analysis of usability tests, simulation of user journeys, assisted heuristic evaluation, automated detection of accessibility and conversion issues.

The trap here is subtle: AI can produce a usability report without any real user ever touching the product. That’s useful as a pre-filter but disastrous as a replacement. Behavioural simulations assume an average user who doesn’t exist. Real users break hypotheses, ask off-script questions, fail in ways no model predicts.

My current rule is that AI can prepare, accelerate and cross-check validation, but can’t replace direct observation. A test with five real users is still more informative than an automated analysis over fifty simulated scenarios.

The deepest shift: the designer as accountable curator.

If I had to summarise how the role has changed in the last two years, I’d put it like this: I used to design what the user was going to see; now I answer for what a probabilistic system decides to show them.

We’re moving from deterministic interfaces to probabilistic systems. That means designing for:

Responses that aren’t completely predictable.
Interfaces that adapt in real time to context, intent and behaviour.
Personalised content at scale, generated on the fly.
System states that didn’t exist: “the AI is thinking”, “the AI isn’t sure”, “the AI got it wrong”, “the AI needs human oversight”.

In this kind of product, the designer no longer draws every possible screen. They design the space of what the system is allowed to produce, define invariants that can’t be broken, and accept that the final output is a collaboration between their direction and the model’s generation. It’s closer to directing a film than to painting a picture.

New design artifacts.

As a result, the designer’s output has expanded. Today, alongside screens and components, I also produce:

Structured, versioned prompts. A prompt that defines how the system responds to a user isn’t secondary documentation: it’s part of the product. I treat it like code: with versions, tests, pull-request review and a change history. A change in a prompt can modify the behaviour of the product for thousands of users without touching a single visual line.

Operationalised voice guidelines. Not as a PDF nobody reads, but as concrete rules the model consults and follows: what tone to use, what words to avoid, how to respond to emotional escalation, when to hand off to a human.

Conversational state machines. For products with agents or assistants, design includes defining valid states of the conversation: thinking, responding, awaiting confirmation, hallucination detected, unverifiable source, escalated to human. Each state with its visual representation, its microcopy and its accessibility behaviour.

Evaluations (evals). Probably the newest and most important artifact. These are test suites that verify the system’s behaviour against typical and adversarial inputs. Just as a visual component has regression tests, an AI-driven flow needs evals that detect when it starts to behave worse. Without evals, any improvement to the model can silently degrade the experience.

Provenance and confidence indicators. Components that show the user where a response comes from, with what level of certainty, whether it was verified and by whom. In 2026 this is no longer optional: it’s a regulatory obligation for AI-generated content.

Designing is no longer just defining how something looks. It’s defining how the system thinks, how it explains what it does and how it cedes control when it should.

New competencies for the designer (and which skills are at risk).

Integrating AI redefines the professional profile. There are five competencies that differentiate today, and one worrying erosion that deserves a name.

Critical thinking about AI outputs.

Evaluating the quality, coherence, biases and risks of what a model generates. It’s not a theoretical skill: it’s the daily muscle. Facing every screen, prompt or synthesis the AI produces, the operational question is: what assumptions are sneaking in here that I haven’t verified?

Prompting as a design discipline.

The prompt is the interface between the designer’s intent and the model’s capability. A good prompt is brief but specific, establishes context, constrains the output space and gives clear success criteria. It’s closer to writing a brief than to writing code. What differentiates designers is increasingly what they know how to prompt, not what they know how to draw.

Data and code literacy.

Understanding how the data feeding models is generated, processed and labelled. Reading enough code to modify, debug and evaluate. It doesn’t mean becoming an engineer; it means not depending blindly on the output.

Designing conversational and agentic systems.

Structuring non-linear dialogues, handling ambiguity, designing error states where the system doesn’t know something, designing escalation to humans. The conversational interface has its own heuristics and doesn’t resemble the visual interface.

Evaluation and measurement.

Defining what “good” means for a generative system, building evals to measure it, interpreting results. It’s an emerging skill that combines product judgement, technical understanding and test design.

The erosion: silent atrophy of baseline judgement.

Here comes the uncomfortable part. When a junior designer starts their career delegating production to AI, they risk not building the baseline judgement that would let them assess what the AI produces. It’s a generational paradox: the tool that accelerates those who already know how to design can prevent new entrants from learning to do it.

I’ve started seeing CVs where the portfolio is spectacular and the technical conversation is fragile. The portfolio was made with AI; the capacity to defend decisions was not. If you’re senior, this affects you as an employer and as a mentor. If you’re junior, the recommendation is hard but honest: for the first 500 hours of your career, design without AI. Not out of romanticism, but because you need to build the judgement you’ll later use to direct AI. Without it, you’re at the mercy of the output.

The designer stops being a translator of needs into interfaces. They become an orchestrator of complex systems and guarantor of quality in a pipeline they no longer fully control.

Risks and limits of AI in design.

Enthusiasm for AI hides risks that, unmanaged, end up in worse product and legal exposure.

Vibe designing: accepting what you can’t evaluate.

It’s the design equivalent of the “vibe coding” that has gone viral among developers: accepting the output because it “looks fine” without really understanding what it does. In design it takes the shape of validating a screen because the hierarchy looks right, without verifying contrasts, labels, keyboard focus order or responsive behaviour.

The defence is boring but effective: an executable review checklist that runs over every output before accepting it. Contrasts. Target sizes. Semantic accessibility. Copy reviewed. Error states. Internationalisation. If you can’t tick the box, you haven’t reviewed it.

Aesthetic homogenisation.

Discussed above, but worth repeating because of its importance: the mass use of models trained on similar data produces generic interfaces that lose differentiation. The brands that will stand out over the next few years are those that deliberately invest in their own visual identity, specific art direction and non-obvious aesthetic decisions. The value of design returns, paradoxically, to the human work on what a model can’t do by default.

Technological dependence and fragility.

Excessive dependence on specific tools creates operational risk. If your flow depends on four AI-powered SaaS products and one changes its model, its pricing or its usage policy, your process breaks. I’ve seen teams paralysed for a week by the silent change of a model that started producing worse outputs without prior notice. Diversifying tools and keeping baseline skills is risk management, not conservatism.

The environmental cost nobody measures.

It’s an uncomfortable topic, but present. Every image generation, every synthesis, every iteration has a computational and energy cost that until recently didn’t appear in any design equation. In 2026, with regulatory and corporate pressure to report emissions, this is starting to get metrics. Generating a hundred variants of a screen to pick one is an energy luxury whose cost-benefit deserves, at least, awareness.

Intellectual property and commercial risk of outputs.

This is the risk that has changed most between 2024 and 2026, and the one many teams keep ignoring. Recent litigation makes it clear that:

The training data of the main models contains copyrighted works. Some models trained on illegally obtained datasets, and several multi-million-dollar settlements have already cemented it as a real financial risk.
Although training may be covered by exceptions (fair use in the US, text and data mining exceptions in the EU with creators’ opt-out rights under the CDSM Directive), outputs can infringe copyright if they reproduce specific works sufficiently. New lawsuits are shifting the focus from training data to specific outputs.
The US Copyright Office has confirmed that content generated solely by machine is not registrable as a copyrighted work. Only the parts where there is substantial human authorship can be protected, and the rest must be disclaimed.

The practical implication for a designer in 2026: if you’re generating visual assets with AI for commercial use, you need to know which model, under which terms, and with what indemnity guarantees the provider offers. Some offer legal cover when their enterprise models are used; others don’t. It’s not a technical decision; it’s a legal one.

Ethics and responsibility: from abstract principles to concrete obligations.

Designing with AI means taking on new responsibilities. And, unlike two years ago, many of them are no longer ethics in the abstract: they’re law in the concrete.

The real regulatory framework in 2026.

The EU AI Act entered into force on 1 August 2024 and will be fully applicable from 2 August 2026. That activates obligations with direct implications for design:

Mandatory transparency for limited-risk systems. Chatbots, content generators, deepfakes: users must know when they’re interacting with an AI. This is literally a design decision. The component that communicates “you’re talking to an automated assistant” has to exist, be visible, and not depend on the team’s goodwill.

Labelling of generated content. Text, audio, image, video generated or manipulated by AI must be clearly marked as artificial when published for public use. This changes the interface: watermarks, visual indicators, metadata. It’s part of the product, not a legal afterthought.

Documentation and traceability in high-risk systems. If the product you work on falls into high-risk categories (hiring, credit, education, essential services, etc.), there are obligations for human oversight, event logging, and possibility of user intervention. The component that says “you haven’t been accepted” can’t exist without mandatory slots for reason, review path and human contact.

Training data transparency. Model providers must publish sufficiently detailed summaries of training data and respect rights reservations (opt-outs) creators express under the CDSM Directive. This doesn’t affect you directly as a designer consumer, but it affects which models are defensible for commercial use.

Penalties for non-compliance go up to 35 million euros or 7% of global annual turnover, whichever is higher. These aren’t theoretical threats.

Operationalisable ethical principles.

Given this context, the classic ethical principles (transparency, control, explainability, privacy, inclusion) stop being aspirations and become verifiable requirements. In practice, in my current work this means:

Transparency as a component. All AI-generated content leaves the system with a visible marker by default. Disabling it requires an explicit flag and gets logged for audit.

Control as a mandatory flow. Every automated decision has a review path and a human contact accessible in under two clicks. The design makes it impossible to “bury” this path.

Explainability as microcopy. When the system recommends, denies or personalises, it offers, at minimum, one sentence explaining why. It’s not perfect explainability; it’s the minimum threshold below which the product doesn’t ship.

Privacy by default. The system uses the minimum personal data possible. If it uses more, it’s explicit and granular opt-in, not a pre-ticked checkbox.

Verified inclusion, not assumed. Flows are tested with users from groups the model might be serving worse (minority languages, assistive technologies, non-Anglo names that break forms). It’s not enough to say “AI is inclusive”; it has to be checked.

Ethics isn’t a layer added at the end. It’s a set of decisions made in every component, every flow, every prompt.

The human factor remains essential (and now carries more weight).

For all AI’s capabilities, human judgement is not only still irreplaceable: it has become more critical, because it now applies over outputs generated at a speed no manual process could match.

AI:

Doesn’t understand context the way a person does.
Doesn’t take responsibility before a regulator, a user or a court.
Doesn’t define strategy or discriminate between contradictory priorities.
Has no taste, though sometimes it looks like it does.

The designer:

Interprets real needs, with their ambiguities and contradictions.
Sets priorities based on product, business and user context.
Makes decisions with measurable impact and answers for them.
Exercises aesthetic and ethical judgement which is, literally, what they bring to the process.

The right relationship isn’t replacement. Nor is it the friendly “collaboration” repeated in so many articles. It’s something more tense and more productive: critical collaboration. AI proposes; the designer accepts, modifies or discards with judgement, and answers for the result.

Conclusion: the question you have to be able to answer.

Artificial intelligence isn’t one more tool in the design stack. It’s a paradigm shift that transforms how we research, ideate, build and validate. But it also redefines what it means to design, what responsibility we take on, and what kind of products we create.

In 2026, adopting AI in the process isn’t about working faster. It’s about designing with greater reach, but also greater exposure: to errors that scale, to biases that amplify, to legal obligations that crystallise, to intellectual property risks that turn into litigation.

The operational question I ask any designer integrating AI into their work, and ask myself every week, is this: if tomorrow a regulator, a client or a user asks me why my product behaves the way it does, can I answer with anything more than “the model generated it”?

If the answer is yes — if I can trace the decision, explain the criterion, show the evaluation, document the limit — then I’m designing. If the answer is no, I’m delegating. And delegating to a probabilistic system without owning the result isn’t product design; it’s abdication in make-up.

The real value of the designer in 2026 doesn’t diminish with the arrival of AI. It becomes more strategic than ever, but also more demanding. Because now, beyond designing the experience, you have to answer for it.