From 1.2% to 22% of customers in eight months — an AI Companion enterprise teams actually pay for

Context

AI where you work, not in another tab

The brief was a vision document. Everything between it and a shipping product was still open.

cplace’s users are enterprise employees in every role, plus the admins who keep the platform running. The Companion had to serve both, inside the interfaces they already use. The strategy ran on two dimensions — AI building blocks that provide capabilities, and interface elements that make them reachable throughout the platform — and on three principles I helped define after joining:

Success meant less time on routine tasks, more productive users, broader use of the platform’s advanced features and less need for training. Adoption of the Companion itself became the number we watched most.

Research

Three groups, one finding I didn’t expect

We benchmarked Planisware and monday.com for PMO, then interviewed three groups of people:

Design decisions

Sixteen tools, one card, four states

The hardest call wasn’t the chat. It was what each tool’s answer should look like in a side panel.

The Companion is a panel, not a full screen. Charts that shine in a demo don’t fit, and sixteen tools — search, people, data model, page management, knowledge bases — could easily have become sixteen interfaces. Instead, every tool call renders as the same collapsible card: what you asked, one line of intent, the tool’s name, the request, the result. The result has four states.

Big results didn’t get squeezed: complex charts open on a subpage through a URL in the answer. Power users get the raw request and response behind a code toggle — the same transparency the confirmation step gives everyone else.

Showing the thinking

Proposals
Four coded variations of the thinking UI — built to be argued about, not shipped.

Shipped
Status cards with Show details, then the answer. Less theatre, same transparency.

The PM hat

Prioritising with two questions

Alongside design I ran the day-to-day product work: which decisions first, what goes into the sprint, who is free across development, testing and design. Every request — internal or from a customer — got the same two questions: does it add a capability, or does it make one reachable where people work? Anything that answered neither waited.

Outcome

From 1.2% to 22%

Eight months after launch, 22% of existing customers had subscribed to AI Companion, up from the 1.2% who tried the paid feature at the start — and retention is strong. It sells as a quarterly subscription, and customers bring the model they prefer: the admin view above shows Azure GPT-4o and Claude models running side by side in one tenant.

How we work now

The part that changed me: designer to design engineer

With my UX colleague Adrian I set up an AI-first working mode for the whole product team. It is why my title now says engineer.

1. Prototype in Claude Code
   Working prototypes, not static mockups — several variations per question.
2. Discuss the proposals
   The team argues over things that run, then picks one.
3. Polish in Figma
   The chosen direction gets its final design decisions.
4. Figma MCP → VS Code
   The polished design connects straight back into the codebase.
5. Design gate, then ship
   A QA step before release; small tweaks I push to GitHub myself.

Three things became standard: Claude Code as the team’s prototyping tool, a design gate before anything ships, and one handover format from product owner to design to engineering. Rounds like the four thinking-UI variations above replaced static mockups as the thing we discuss, and the whole design → development → QA loop got faster and smoother. Underneath sits Helles, the design system I co-develop on PrimeNG v22 — the shared component layer that lets a Figma decision land in code without translation.

Reflection

What I took with me