2025 was a year of architectural shifts, moving mandates, and timelines that didn't leave much room to breathe. Finding myself backed into a corner, I used levers to navigate through scope negotiations, stakeholder influence, pruning complexity, and building cross-functional coalitions to get us through in one piece.

In regulated, high-stakes environments, the gap between what a GenAI tool can do and what people will actually trust it with tends to be wider than anyone budgets for. I watched it play out across two years of beta rollouts inside a live banking environment. The friction wasn't purely technical. At the heart of it was a simpler, harder question. If something goes wrong, who answers for it?

How a design team is structured tends to reflect a deliberate set of calculated trade-offs to be made. The harder question isn't which model to choose. It's what needs to be in place to keep it working as intended. Without the right support, any setup will drift, and when it does, design stops shaping decisions and starts fulfilling them.

Across 2024 - 2025, I worked with business units to deploy beta GenAI features. Stakeholders were ready but we exercised caution until we understood what we were actually building and for whom. In a high-trust environment, a poorly received beta is hard to recover from. Prudence turned out to be the more interesting design challenge.

Enterprise interfaces demand precision under pressure. Users managing hundreds of concurrent data points, workflows and deadlines in a single session aren't browsing. They're operating with little margin for error. I drew on cognitive science to develop a signal-centric, composable semi-opinionated DLS, architected to hold brand coherence with processes to scale sustainably across 32 platforms and multiple countries.