For enterprise agents, the hard design problem is rarely the model. It's transparency, boundaries, and the human checkpoint.
When I designed Evai — a concept for an AI agent that works inside enterprise workflows — I expected the hard problems to be about the model: accuracy, latency, capability. They weren't. The model was the easy part. The hard part was designing the interface that decides whether anyone trusts it enough to use it.
I learned this the way you're supposed to: by talking to the people who'd use it. In discovery interviews, the same theme kept surfacing. Enterprise users don't want a more eloquent chatbot. They want an agent that acts inside the workflow they already have, and they will only let it act if they can see why.
A generic chat interface puts all the weight on the model's words. In a low-stakes consumer setting, that's fine. In an enterprise workflow — where an action might reroute a shipment or adjust a price — eloquence is not enough. People need to understand the reasoning, trust the limits, and stay in control of consequences. That's an interface problem, and it has three parts.
The single biggest driver of trust in my research was being able to see why the agent recommended something. Not a raw model dump — a legible summary of the inputs and the logic. When Evai proposes an action, it shows the context it pulled and the reasoning that led to the recommendation. The user evaluates a case, not a verdict.
People don't trust an answer they can't interrogate. They trust a recommendation they can follow.
Competitor agents tended to be unbounded — they'd attempt anything, which made them impressive in a demo and unreliable in practice. Users told me the opposite was what earned trust: an agent that is confident inside its domain and honest when something falls outside it. Scoping the agent isn't a limitation to apologize for; it's the thing that makes it credible.
For any consequential action, the research was unambiguous: there must be a human approval step. So the journey I designed always ends the same way — the agent gathers context, proposes an action with its reasoning, and then stops for a person to review, edit, and approve. The agent does the work; the human keeps the authority.
Transparency, boundaries, and the human checkpoint weren't features I added at the end to make an AI product feel safe. They were the concept. With AI, the design work that matters most is the interface of trust around the model — and that's a job for research and craft, not for a bigger model.