Leading Design for an Agentic AI Product Group
I led 11 designers across four product suites and owned design for iX Hero and iX Hello, Concentrix Catalyst’s public agentic AI products. I organized design reviews across a distributed team and assembled the development team for a React design system with automated QA.
Both products are public: iX Hello 2.0 launched in February 2025 and iX Hero in May 2025. The enterprises whose contact centers ran them are under NDA, so this case study covers the approach and the outcomes without product screens. Happy to talk through specifics in a conversation.
The Challenge
I joined Concentrix Catalyst in 2022 and spent my first 18 months leading a service design engagement embedded with Microsoft. I then moved internally to lead design for a new agentic AI product group: four product suites, and I owned design across all of them. Two of the products are public. iX Hello is an assistant a business stands up on its own data so its employees can work faster. iX Hero is a set of agentic apps for the people handling customer service calls and chats, and it is the one this case study is about.
Each suite had its own management, its own roadmap, and no line of communication to the others. The designers were the ones who talked across them, comparing notes on what each suite was building, which is how we knew when two teams were solving the same problem twice. Engineering rebuilt the same components for every suite. On top of that, agentic interfaces had no settled conventions to fall back on. Nobody had a pattern library for what an agent should see when a model is working on a task.
iX Hero was the harder problem. The answer to a customer’s question might sit in the client’s own data systems, in Salesforce, or in whichever third-party integrations that client happened to buy, and every client’s stack was different. All of it had to fit on a laptop screen the agent was already working in. On a phone call the agent is talking while they read, so anything that takes hunting is too slow to be any use. What we had to work out was which information belongs on screen at which point in a conversation, and how someone gets to the rest without a pause the customer can hear.
My Approach
I structured the team so designers stayed close to their own product suite and its customers while working from one shared set of patterns. That meant weekly critique across suites, not just within them, and a single quality bar I held rather than four different ones.
The team spanned India, Poland, and the United States, so reviews ran across about twelve hours. Anything that needed everyone in a room at once was not going to work.
On the foundation, I pushed for a design system in React rather than a Figma library alone. I wanted the design patterns to be implemented in the component library, so designers and engineers could work from the same foundation. Integration across the suites was still in progress when the design team was eliminated. I came up through design and engineering before management, so I still know what it costs the team to ship.
For the agentic AI work, the design problem was the handoff between the model and the person. We had to decide when to show the model’s reasoning and when to stay quiet, when to ask an agent to confirm before acting, and how to let someone step in mid-task without losing the context the model had already gathered.
On iX Hero, most of the work was defining the conversation workflow before we designed anything. If you know where you are in a conversation you know what the agent is about to need, so the screen can get there first rather than wait to be asked.
Designing the human and model handoff
The team worked through what an agent needs while the model is acting, when confirmation is needed, and how someone can intervene without losing the conversation context.
- 1. Make context visibleWhich booking or customer details does the agent need at this point?
- 2. Show the proposed actionWhat is the model working on, and what needs human review?
- 3. Support confirmation or correctionWhen should the agent confirm, override, or take over?
- 4. Carry context forwardWhat must remain available when the work moves between the model and the person?
Reconstructed for this case study to explain the design questions. This is not a product screen or a claim that every pattern shipped.
The Result
Design systems had failed at Concentrix more than once before I got there, so the hard part was not building one. It was getting anyone to back another attempt. I put together the team to do it properly: JavaScript developers, testers, and a two-person design system team, one of them doubling as product owner because that was what it took to get it funded.
We finished the token library, the component library, and the admin side, with automated QA running in the pipeline. Integration across the four suites and the component audit were still underway.
iX Hello 2.0 shipped in February 2025 and iX Hero in May 2025. Both are still in market.
What I would change
I would make direct access to customer-service agents and a concrete evaluation plan explicit requirements earlier. Our user input was secondhand. I asked for agent access and raised monitoring more than once, but accepted assurances and the work continued.
In a future engagement, I would document the research access, the evaluation owner, and the evidence needed for the next decision. If those were missing, I would put the tradeoff in front of the product leads while there was still time to change the plan.
