Discovery
We read your evals, model cards and support tickets, because the failure profile is the brief for most of the interface.
AI
AI products fail on trust, not capability
The model usually works. What is missing is the interface that shows its working.
Agent and workflow UX
Multi-step behaviour, task handoffs, interruption and resumption, and the boundary between acting alone and asking first every time.
Trust and transparency
Showing the working. Sources, confidence, audit trail, the reversible action. In a regulated category, this is the sale.
AI design systems
Components for streaming, partial and failed states, citations, corrections and confidence. Most systems assume one single correct state.
The process
Six stages. Twelve weeks. A date we hold.
Decision boundaries
We draw the line between what the system does alone, what it proposes for approval, and what it refuses to touch.
Prototypes
Clickable, in front of users, running against real output rather than a screenshot of a good day.
The trust layer
Sources, confidence, what changed, what can be undone and what was logged, readable in seconds by someone busy.
Interface design
Production UI with every state designed. Streaming, partial, empty, stale, wrong, rate-limited and unavailable.
System and handover
A component library your engineers build from rather than around, documented so it holds when the model changes underneath.
FAQs
Questions we get
Then we are the wrong studio. Features built to be seen rather than used get shipped, ignored, and quietly removed a year later, after eating a quarter of the roadmap. That is a much cheaper conversation before the build than after it.
That is a design brief, not a blocker. An interface that communicates uncertainty honestly outperforms one that performs certainty, especially with technical buyers who can smell the difference. Reliability sets what the system may do alone. It does not decide whether you can ship.
No. We design the product around it, working alongside your ML or engineering team. We will read your evals and your docs rather than ask for a summary, and we will not pretend to opinions about your architecture that we have not earned.
Good, in a sense. That is where transparency stops being a differentiator and becomes the product. Audit trails, reversibility, human sign-off and refusal states are all design work, and they are the parts of this we most like doing.
Designing the interface a person uses to work with a model: what the system does automatically, what it asks about, how the output is presented, how it is checked, and what happens when it is wrong. The model is one component. AI UX is everything around it.
Twelve weeks for a full engagement, discovery through handover. A trust layer on an existing product is shorter. You get the date at proposal stage, before you commit to anything, and we scope to the actual gap rather than sell you everything.
Pick the number before we start. Usually feature adoption, completion rate, override rate or how often a user accepts the first output. We agree it in discovery and report against it after launch.