Designing with AI without losing control is possible — provided you don't confuse two things that speed tends to blur together. Generative tools produce mockups, variants, and interactive prototypes in minutes. They do not produce the understanding of needs that makes an interface right. Production is accelerating; understanding, meanwhile, is still built with real users. That is exactly where control is won or lost.
Published January 2026, revised May 2026 — this version is a full rewrite.
What generative tools genuinely do well
Let's start by acknowledging the capabilities, without professional defensiveness. From a written description, a tool today produces structured, usable wireframes. Where a team once prepared three variants for a test, it can now explore dozens. Reviewing reference material — scanning hundreds of existing interfaces to extract a sector's conventions — now takes minutes instead of days. And prototyping has become interactive: you test a clickable journey, not a static image.
I experienced this firsthand on my own site. During the redesign of mcva.ch, I generated variants of entire pages in a single evening — structures, hierarchies, visual directions. That volume of exploration was out of reach for a solo consultant three years ago. For a Swiss SME with a tight design budget, this acceleration is a genuine boon.
But here's what that evening also showed me: none of those variants knew what my visitors were actually looking for. The choices that shaped the final site came from elsewhere — from positioning work, from the real questions business leaders ask, from the pages that language models cite or ignore, which is what I measure through the Score GEO™. The tool explored. The decision was made on different ground.
A mockup is not a design decision
The point worth remembering fits in two sentences. Co-design is the process that builds, together with real users, the understanding of needs — both expressed and latent — from which a product can become relevant. In 2026, generative AI massively accelerates the production of design artifacts; it does not produce that understanding.
Three areas remain out of reach for models, and it's on these that lasting value is built.
Active listening, first. An experienced practitioner picks up on what an interview doesn't put into words: the user who says "it's fine" while frowning, the hesitation before a click, the contradiction between what someone says and what they do. Models work on text and structured data; they see neither body language nor the emotional context of an exchange.
Local cultural context, next. The Swiss market has its own codes — an approach to discretion, a demand for quality over volume, multilingualism, high expectations around data protection. Models trained predominantly on Anglo-Saxon data don't carry these codes by default, and the generic visual patterns they produce don't capture them either.
Unspoken needs, finally. The most defining needs are often the ones users never articulate — either because they consider them obvious, or because they don't know a solution exists. A model generates from the patterns present in its training data; it doesn't reveal what isn't there. Observing real-world practice does.
How do you combine the two, phase by phase?
The useful question for an executive commissioning a project: at what point is acceleration legitimate, and at what point does it become a shortcut that costs you? The table below summarizes the balance that works, phase by phase.
| Phase | Who leads | Role of generative tools |
|---|---|---|
| User research | The human, entirely | No decisive role |
| Concept generation | The team, which selects | Mass production of variants |
| Testing and iteration | The human, in direct observation | Kept in the background |
| Refinement | The designer, who decides | Rapid execution of variations |
In plain terms: user research — interviews, field observation, workshops — remains entirely human, and teams that skip it produce convincing artifacts that solve the wrong problems. Concept generation makes full use of the tools, provided it is fed by research. Testing with real users becomes human again, because a metric quantifies what it can quantify, not necessarily what matters. Final refinement combines both: the designer decides, the model executes the variations.
This sequencing explains why a short co-design sprint — a few days spanning research, AI-assisted generation, and testing — has become accessible to organizations that could never mobilize a design team for weeks at a stretch. This is the tight format I apply in the solutions I build for SMEs.
What guardrails keep you in control?
Without an explicit framework, a model's default choices become the company's choices, through simple negligence. Three rules are enough to prevent that.
Validation before exposure: every generated deliverable — mockup, journey, visual — goes through human review before it is shown to a user or a client. The tool proposes, the designer decides.
Traceability: note what comes from automatic generation and what comes from user research. A few methodical annotations are enough, and they prepare you for transparency requirements that are becoming more precise — Switzerland's Federal Act on Data Protection already requires increased transparency around automated processing that affects users[1].
Primacy of user feedback: if a tool recommends a journey optimized for engagement but testers find it confusing, human judgment wins. Always. This hierarchy prevents optimizing metrics at the expense of genuine satisfaction.
These guardrails are set at the proposal stage, not along the way — it's one of the clauses I systematically frame in a custom project before generating anything at all.
What this changes for a Swiss SME in 2026
When every competitor can produce visually correct interfaces within days, the artifact stops being a differentiator. What differentiates you is rightness — an interface that shows the client you've understood their context. That rightness comes from understanding needs, precisely what co-design builds and what tools don't reproduce. In other words: the acceleration of production has made user research more profitable, not less.
The same shift runs through every discipline of digital production, incidentally — I documented it on the code side in AI and Developers. And to place design within a complete AI adoption roadmap, I laid out the overall approach in What should a Swiss SME do about AI in 2026?
Key takeaways
— Generative AI accelerates the production of mockups and prototypes; it does not produce the understanding of needs, which remains the core of design. — The balance that works: human user research, AI-assisted generation, human testing, supervised refinement. — Three guardrails to set at the proposal stage: validation before exposure, traceability of decisions, primacy of user feedback over the metric.
FAQ
Can AI replace a UX designer? No. It replaces part of the production — wireframes, variants, prototypes — but not the ability to understand users, capture unspoken needs, and make judgment calls. A project run without that understanding produces plausible interfaces that solve the wrong problems.
Is co-design still worthwhile if AI produces everything so fast? It's worth more than before. Accelerated production has commoditized correct-looking interfaces; differentiation now hinges on rightness, which comes from user research. And the format has tightened: a sprint of a few days is often enough where weeks used to be required.
What are the risks of letting a tool generate my interface without a framework? The model's default choices become your choices: generic aesthetics, journeys optimized for metrics rather than for your customers, and a homogenization the market eventually notices. There's also a compliance dimension, since the FADP requires transparency around automated processing that affects users.
How do I know if my provider is using AI correctly in design? Ask three questions: what user research precedes the generation, who validates each generated deliverable before it's shown to anyone, and how test feedback takes precedence over the tools' recommendations. A serious provider answers all three precisely.
And for your own design projects? The AI Usage Diagnostic: sixty minutes to lay out your real workflows, identify what deserves custom work, what stays in SaaS, and what needs no AI at all. Book a diagnostic
Sources
[1] Federal Act on Data Protection (FADP), revision of 25 September 2020, in force since 1 September 2023. www.fedlex.admin.ch/eli/cc/2022/491/fr [↩]
Jérôme Deshaie is CEO and founder of MCVA Consulting SA, an augmented agency based in Valais, Switzerland. Fifteen years serving major international brands, now working directly with Swiss SMEs. Background.