Context

My engagement with Novonesis began in June 2025, working across smaller projects tied to the same employee base and to the organization's early experiments with AI in design and development workflows. In September 2025, I joined the EnterpriseAI project as the sole UI/UX designer, embedded in a team of four developers and a UI architect, three weeks from Beta soft-launch. The product was a greenfield enterprise AI assistant, purpose-built on Novonesis infrastructure for employees across scientific, regulatory, and business functions. I worked in Agile from Beta launch through the V1 revision phase, ending in February 2026.

Problem

Novonesis employees are highly educated: PhDs, researchers, legal and regulatory specialists. But domain expertise doesn't map to digital fluency, and it doesn't map to familiarity with AI interaction patterns. These were new. The same person who held sophisticated scientific judgment was encountering a conversational AI interface for the first time. The design challenge: introduce a fundamentally new class of tool to users with no established mental model for it, in an environment where misuse or misplaced trust had real operational consequences.

Role

I arrived with three months of context about Novonesis employees and their relationship to AI tools. When I joined the EnterpriseAI project, I became the sole designer on a team three weeks from Beta soft-launch. I owned the full design layer: the revamped UI shipped at go-live, the compliance cards on the new chat page, the file upload experience, the selective copy interactions across output types, and the Thinking component that makes LLM reasoning visible to users. In parallel, I built a local component library in Figma and aligned their documentation for developer handoff. I reported on design progress to product leadership weekly.

After the Beta launch, the dynamic shifted. The product was in active use, the core design decisions were established, and the frontend team had started generating UI with AI-assisted development tools. Developers brought those implementations to me with UX questions: does this pattern hold? Does this interaction create confusion? My role became less about proposing and more about evaluating. I was the reference point for UX quality in a team that had become capable of generating interface without waiting for design to lead.

Key Decisions

Chat as the default interaction paradigm. I chose the conversational interface over more structured alternatives because it offered the lowest barrier for users encountering AI for the first time. Scientists and regulatory specialists don't need a new mental model to start a conversation. The chat format didn't make the tool familiar because users already knew AI interfaces; it made it familiar because everyone already knows how to send a message. The tradeoff was flexibility: a conversational interface constrains structured interaction and limits the product's ability to guide users toward better prompts. That was a known cost, accepted in exchange for adoption.

From a compliance line to a governance surface. The initial request was minimal: add a line below the prompt input confirming appropriate use. The compliance requirements were real but still evolving, and the team needed something implementable before the regulatory framework had fully matured. I expanded that request into a design problem: how do you communicate governance to users in a way that stays current, doesn't create update debt, and doesn't block access to the tool?

I explored three approaches: a guided onboarding with contextual animation for first-time users, a persistent information layer on the empty chat page, and a dismissible version of the same layer. The first was studied and deferred. The second and third were simplified into static compliance cards on the new chat page. The cards shipped fast, were easy to update as the regulatory language evolved, and kept governance in view without interrupting the workflow.

The tradeoffs were known. On desktop, the card density was higher than ideal. On mobile, the cards occupied the full vertical space of a small screen: a problem flagged at the time and committed to address in the next iteration cycle.

Making LLM reasoning visible. In a regulated environment, users evaluating AI output need to assess not just the answer but how it was reached. I designed the Thinking component to surface the model's reasoning process as a collapsible layer attached to each response. Hiding the reasoning entirely would have produced cleaner output but removed the verification affordance that builds appropriate trust in a context where acting on an AI answer has professional consequences.

Granular copy interactions per output type. I defined separate copy behaviors for tables, images, code, and diagrams rather than a single copy-all interaction. Scientists and regulatory staff extract specific outputs for downstream use in lab notebooks, reports, and submissions. Granular copy reduces the friction between the AI tool and the workflows it feeds into.

Result

The product launched to 187 users on November 19, 2025. Sixty-five days later, it had reached 2,540 sign-ups, 14,000 conversations, and 56,707 user messages. The error rate dropped from 0.2% to 0.002%, a 100x improvement after identifying and resolving expired authentication token failures. On January 28, 2026, the product recorded 505 conversations in a single day, the first time it crossed that threshold.

What Remains Open

The contract ended in February 2026 with the product in active growth and the team operating with design autonomy. Two items remained in design but not in the shipped interface: the file upload conversion communication, fully specified and ready for implementation, and the mobile-optimised revision to the compliance cards. Both had documented specifications. What the team inherited was not just a component library. It was a set of interaction principles they already knew how to apply: the governance patterns, the transparency layer, the system of selective copy. The product would keep building on that foundation without needing the designer present to explain it.