AI and product design: building trust, with your users and with your teams
Your AI works technically. What remains is whether your users understand what it shows them, and whether they will know when it gets things wrong.
Everyone has already run into hallucinations. A designer who uses AI asks themselves, like the user of your feature, when to believe what it suggests, and we ask ourselves the same thing when we use Claude.
Making AI a tool for the people who use your service or your product is our job. We don't build the model: we design, with your data or AI team, what the user sees of it and what they do with it. You can bring us in for an audit, project support or training.
When you call on us
- You are a product owner or in product management, and AI is already in your product. It works technically, but not for your users, and nobody knows when the AI gets it wrong or how the user will find out.
- You built a prototype with Lovable, Claude or another tool, you are using it as the basis for what comes next, and you still have doubts about its quality.
- You run an SME or a startup, an innovation department or an IT department, at a time when everyone is wondering how to bring AI into their site, through a chatbot, a smart search engine or something else. The first question we ask with you is what it can bring you, as an invisible building block that surfaces content that classic algorithmic approaches could not offer.
- You lead a design team, and you wonder what it changes to equip yourselves with Claude day to day for research, UX and UI.
Why now
In France, the job postings Indeed files under its designers and artists category fell by 28 % in a year (28-day average at 25 September 2026, compared with the same period in 2025). This category does not measure UX alone, but it is the only public French series we found. For our part, we see a market restructuring deeply. In-house design teams are gradually doing the acceptance review of mockups that product owners produce with Claude, or of prototypes made with Lovable. Research analysis is handed to Copilot, and developers receive mockups in HTML, CSS and JavaScript coded by GPT. In international surveys, half of the 900-plus designers polled by Designer Fund and Foundation Capital for “AI in Design 2026” say they have already pushed code, with a sample that leans toward AI enthusiasts, as the authors acknowledge. According to Figma's 2026 survey, the share of designers who take part in development has doubled in a year, to 41 %.
How we work with you
An audit, to see your product from the side of the person using it
The audit relies on an expert inspection, interviews and tests with your users. It looks at what the user understands of what the AI shows them and what they believe of it, how errors reach them and are flagged to them, and the interface and journeys built around the AI.
We design with your team and test with your users
We design the loop of use around the AI with you, working with your data or AI team. We make the level of confidence one can place in an answer visible in the interface, and we test with your real users, in iterations.
A training course in two modules
The training has two modules. One is reserved for design teams and shows what it means to equip yourselves with Claude day to day, for research, UX and UI. The other teaches how to make a design system readable by AI.
Three things to know before you call us
You are starting from an AI-generated prototype
A prototype made with Lovable, Claude or another tool can serve as a basis for what comes next. We look in it for the project owner's intent, the usable references the prototype gave them, and what has to be undone because it is generic output from the AI tool. The quality of these prototypes remains uneven today, even if it is improving very fast. Among other things, it is hard to get a pixel-perfect result on the first try, and they skip steps or idealise the user's need.
No synthetic users in place of real ones
Clients, in very different sectors, already ask us whether an AI could answer in place of their users. We ran the test in March 2026 with Claude Sonnet and Opus 4.6, on a questionnaire for which we had the answers of 1,000 real French people. In 27 tests out of 30, the AI's answers differ statistically from the real ones, and they pull everyone toward an average profile. We tell the story of the study in our Lab.
A design system readable by AI, on our own site
On akiani.fr, most code changes are made with Claude Code. We took the colours, typefaces and spacings of the site from the Figma mockup, then described them in a file in the design.md format, which AI agents know how to read. An integration guide goes with it. A script then checks that the code produced respects the design system. We did it for our own site before teaching it in the second module of the training.
How we stay up to date
Ten of the twelve people at Akiani have taken at least one course from Anthropic's Claude Academy, from Claude 101 to the introduction to the Model Context Protocol.
We also put AI to the test of our own methods, as in the Lab study on synthetic panels, and we are currently testing interviews run and analysed by AI.
Möko, an English-learning platform built around AI
Möko, a start-up from Nice, wanted an English-learning platform made adaptive by AI. We first shaped the journey and the screens of the learning loop, where each notion follows the same pattern: learn, practise, then be assessed. A journey workshop mapped the backtracking, those moments when the learner starts over and frustration lurks, to match the tone of the messages to their emotions. The interfaces borrow the codes of video games without distracting the learner during the modules.
Read the Möko case studyFrequently asked questions
You can plug a language model into a base of synthetic personas and have an agent navigate your site. Methodologically, though, nothing today replaces a human whose behaviour you observe. On our March 2026 questionnaire, the AI's answers pulled everyone toward an average profile, with no extremes and far fewer “I don't know” answers.
Quite the opposite. A well-structured, well-documented system is what lets an agent generate code that matches the product rather than a generic interface. When code is produced on an assembly line, the results look alike, and differentiation comes from what you give the machine to read.
We don't know yet whether that's feasible. What we do know is that it is far from deterministic, and that in a qualitative approach, both for running the interview and for analysing it, this makes comparability impossible to guarantee. We are testing it right now, to form a real conviction on the subject.
Agencies that go by that name often do AI integration or automation. At Akiani, we approach AI from the side of use: what your users see of it and do with it, which we design with your data or AI team and test with them.
Adding AI to a product or a service, or starting from a prototype ?
Tell us where you stand, we reply to every message within 48 working hours on average.
Let's talk about your project