Co-design and artificial intelligence: generated artefacts do not amount to understanding
Note revised on 25 May 2026. Article originally published in January 2026 — full rewrite.
The appearance of tools driven by generative models capable of producing mock-ups, wireframes and interface variants in a few moments raises an apparently simple question: what is the point of a co-design approach costly in time and workshops, when the production of artefacts has massively accelerated?
The answer deserves to be posed clearly. Artificial intelligence produces artefacts. Co-design builds the understanding of the needs from which a product or service can be relevant. These two activities are not substitutable. Confusing the speed of production with the quality of design is the most frequent operational error that design teams have observed for two years.
This note sets out the distinction, qualifies what generation tools actually absorb, identifies the zones where human practice remains structuring, and proposes a framework that makes the articulation operational.
What generation tools effectively absorb
The capacities of tools driven by generative models on design activities are real and deserve to be clearly recognised, rather than minimised through professional defensiveness.
The rapid production of wireframes first. From a textual description, the tools now produce structured and usable mock-ups in very little time. This capacity changes the exploratory phase of a project: what classically demanded several days of manual production can now be handled over a clearly shorter horizon.
The generation of variants next. Where a team classically produced three variants for an A/B test, the tools now make it possible to produce a substantially higher number easily. This plurality changes the quality of exploration: teams are no longer constrained by their production capacity to consider only a restricted number of options.
The analysis of reference sets completes the list of strengths. Examining hundreds of existing interfaces to identify recurring patterns, sector conventions, approaches that have emerged on a given product type, now takes a few minutes rather than several days of manual analysis. This acceleration feeds the inspiration phase without replacing it.
Interactive prototyping closes the list. The tools now produce interactive interfaces — not just static images — that allow testing a user journey under conditions close to the final product. This capacity substantially shortens the lead time between the idea and the first concrete test.
For a Swiss SME with a constrained design budget, these capacities constitute a real accelerator. They do not, however, dispense with the activities that follow.
Three zones where human practice remains structuring
Beyond the artefacts produced, three zones remain largely foreign to what generative tools absorb, and it is on these zones that the durable value of co-design is built.
Empathy and active listening first. Co-design begins with listening that exceeds the verbal. An experienced practitioner captures hesitations, contradictions, non-verbalised emotions of a user in situation. They perceive that a client says "it's fine" while slightly frowning, and they explore that dissonance. This capacity for embodied observation is not reproduced by model-driven systems, which work from textual or structured data, not from body language or the emotional context of an exchange.
Local cultural context next. The Swiss market — and even more so French-speaking Switzerland in its specificities — has characteristics that models trained predominantly on Anglo-Saxon data do not grasp by default. The relationship to discretion, the value given to quality rather than volume, the sensitivity to French-German-English multilingualism, the high compliance expectations on data-protection subjects, shape the needs of local users. A co-design workshop with clients in Geneva and an equivalent workshop in Zurich reveal subtle but decisive differences in experience expectations that generic visual patterns do not capture.
Unexpressed needs complete the list. The most structuring needs for a product approach are often those users do not express — either because they consider them obvious, or because they do not know a solution exists. Co-design brings out these latent needs through observation of real practices, through creative exercises that exceed verbal formulation, through confrontation with prototypes that reveal preferences a direct interview does not surface. Generative models work on the patterns present in their training data; they do not reveal what is not there.
A four-phase articulation
The observable practice of teams that succeed in articulating human co-design with acceleration by generative models typically deploys in four distinct phases, each with a different dosage between the two.
User research first — entirely human phase. This phase consists of conducting interviews, field observations, co-design workshops with target users, and documenting insights, frustrations, expressed and unexpressed needs. No model-driven tool replaces this direct confrontation with reality. Teams attempting the economy of this phase produce visually convincing artefacts that do not solve the right problems.
Concept generation next — a phase combining humans and models. From the insights collected in phase one, the tools accelerate the production of variants and visual directions that the team selects, combines, repurposes. This ideation phase derives a tangible benefit from the acceleration, provided the insights of the first phase actually feed it.
Testing and iteration completes the sequence — a predominantly human phase. The concepts are presented to users in test sessions where direct observation of reactions, hesitations, unexpected preferences, constitutes the material of improvement. This confrontation with reality substitutes for no algorithmic metric, because metrics quantify what they can quantify, not necessarily what counts for product relevance.
Refinement closes the sequence — a phase combining humans and models. On the concepts validated by testing, the tools allow rapid iteration on micro-interactions, journey variations, detail optimisations. The human designer supervises and arbitrates; the model executes the variations.
The human supervision framework
Integrating generative tools into a co-design approach requires an explicit governance framework, without which the biases of the models propagate silently into the final products. Three principles structure this framework.
Systematic validation before exposure first. Every deliverable produced by a model — mock-up, journey, visual content — goes through a human review before being presented to users. The tool proposes, the designer disposes. This discipline is not an administrative formality; it constitutes the mechanism that prevents the models' default choices from becoming the company's choices through negligence.
Traceability of decisions next. Documenting which parts of the design come from automated generation and which parts result from user research facilitates subsequent debugging of the experience, and prepares the organisation for the transparency requirements that are taking shape in the European regulatory framework. This traceability does not require heavy infrastructure — a few methodical annotations suffice.
Primacy of user feedback over algorithmic recommendation completes the list. During test sessions, user feedback always takes precedence over recommendations produced by the tools. If a system suggests a journey optimised by engagement metrics, but testers find that journey confusing, the human feeling prevails. This hierarchy avoids drift towards optimisation of metrics at the expense of actual satisfaction.
This framework is particularly relevant in the Swiss context, where the Federal Act on Data Protection[1] imposes greater transparency on the use of automated systems in processes affecting users.
What practice allows us to observe on the Swiss market
Three observations emerge from observable practice on the Swiss market.
Interfaces co-designed with real users present a quality of use that interfaces produced exclusively by models do not reach. This quality is not measured solely by engagement metrics, which can be satisfactory in both cases; it is measured over time by retention, by spontaneous recommendations, by the loyalty of the client who recognises in the interface an attention to their context.
Co-design becomes a relative competitive advantage in a market where the production of visual artefacts has been democratised. When every competitor can produce visually satisfactory interfaces in a few days, the difference is made on the quality of the understanding of needs — precisely what co-design builds and what the tools do not reproduce.
Accessibility of the practice to SMEs widens thanks to the acceleration of the production phase. A short co-design sprint — typically a few days, articulating a field research phase, an assisted generation phase, and a test and iteration phase — now becomes accessible to organisations that could not, until recently, mobilise a design team for several weeks. This democratisation benefits the companies that accept to invest in the user-research phase, and it leaves aside those satisfied with the automated production phase alone.
The discipline that distinguishes serious practice
Generative artificial intelligence does not replace co-design. It modifies the dosage between the phases of the approach, by substantially absorbing the production of artefacts and freeing up time for the activities where human value remains irreplaceable.
For a Swiss company adopting this articulation, three orientations distinguish durable use from a fashion effect.
Investment in the user-research phase first. The temptation to shorten this phase because the tools now produce artefacts so quickly is the main source of observable failure. This phase remains the foundation of everything else, and its shortening costs more than it saves.
Training in supervision next. Using generation tools with discernment requires a competence built through sustained practice, not by reading manuals. This skills progression deserves to be integrated into the development plan for design teams.
Setting up the governance framework completes the list. Without systematic validation, traceability of decisions and primacy of user feedback, the use of tools drifts towards a homogenisation that clients perceive, sometimes unconsciously, and that progressively erodes the differentiation the company was striving to build.
Human co-design and acceleration by generative models do not oppose. They articulate. This articulation distinguishes, once again, organisations consolidating their product capacity over time from those carried away by the apparent ease of the tools of the moment.
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/en [↩]
Jérôme Deshaie is CEO of MCVA Consulting SA, a Swiss firm specialising in strategic consulting on artificial intelligence, based in Valais.