Here's what enterprise UX taught me that applies directly to AI agent design:

• Cognitive load reduction is the whole game. The reason enterprises adopt AI agents is the same reason they adopted dashboards, workflows, and automation before them: their users are overwhelmed. The job of the AI agent — like the job of every enterprise UX — is to reduce the number of things a human has to hold in their head at once. Not to add a new thing.
   

• Multi-persona design is non-negotiable. In every enterprise system I've designed, the same workflow was used by radically different people — a legal expert and a junior procurement manager, a compliance consultant and a first-time importer. AI agents face the same challenge: the same agent serves users with vastly different expertise levels. Designing for both without patronising the expert or overwhelming the novice is hard, specific work.
   

• Audit trails are a design feature. In regulated industries, every action needs to be traceable. I spent years designing audit logs, approval chains, and version histories — not because users wanted them, but because accountability requires them. AI agents in enterprise contexts need the same: a clear, readable record of what the agent did, when, and why. Designing that log to be useful rather than just complete is a UX challenge most teams underestimate.
   

• The handoff is the hardest part. In enterprise workflows, the most friction always lived at the handoff between systems or people. The same is true for AI agents. The moment where control passes from agent to human — or human to agent — is where things break, where trust is lost, and where the experience either holds together or falls apart. Design the handoff first.
   

• Adoption is a design outcome, not a training problem. Every enterprise product I've worked on had an adoption problem at some point. The instinct is always to "train users better." The reality is almost always that the product wasn't designed to be adopted — it was designed to have features. AI agents face this exact trap. Adoption is designed in, from onboarding to first value to habit formation. It's not added on afterwards.

I didn't come to AI design from a machine learning background. I came from 17 years of learning what it takes for complex software to earn human trust — in the most unforgiving domains imaginable. That background is, I've come to realise, exactly the right training for the AI era.

​

The model is getting better every month. The experience of working alongside it is a design problem that's been waiting seventeen years to become mainstream.