Designing human experiences for what’s next.

Are we building the right agentic experience for enterprise?

The supply chain officer didn't want a chat interface. Her job wasn't a conversation, it was a series of decisions she made every morning, informed by options she'd learned to read at a glance. Going chat-first added work, not removed it. She'd have had to keep asking for every option instead of seeing them. The pushback was specific: show me more upfront, why this recommendation, what else was scanned, and the reasoning behind all of it, so trust in the agent's capability could actually start building.

So we went back. We're now testing a different approach: what if the interface showed her the agent had already done the job, and explained why its top recommendation made the cut? Not a chat, but a view — her options surfaced and ranked, the reasoning visible right alongside them. The same cognitive work she did every morning, done in advance, waiting for her confirmation.

It's early, but the direction came from listening to her, not from assuming that what works for a developer tool works for a procurement workflow.

A second deployment is testing this further. The compliance analysts I'm working with process upwards of 1,000 reports a day. Their job is operational: see the state of a queue, understand what's changed, identify high-priority exceptions, and act on them in sequence, at volume, with accountability attached to every decision. Based on their mental model and the nature of the work, a conversation-led interface feels like it would land the way it would if you suggested reading your emails by asking a chatbot to describe each one. These are hypotheses, not conclusions. We're building an iterative approach, bringing in the actual users so we hear their day-to-day problems directly, and solve for those using the technology.

This tracks with where Anthropic has pointed publicly. Their stated long-term direction is that chat stays the universal fallback, while models generate dynamic UI on the fly when a UI serves the task better than conversation does. The key word is fallback — not default, not primary.

There's a widespread assumption in enterprise AI that agentic means chat, that the interface question has already been answered by the products that most visibly succeeded with that paradigm. Chat works for tools like Claude.ai or Cursor because the job they serve is fundamentally linguistic: you produce something by describing it, and the interface and the output share the same medium. That's the structurally correct choice for that task. But most enterprise work isn't linguistic in nature. A supply chain officer makes sequential decisions across structured data. A compliance analyst monitors a large, changing queue. These jobs aren't conversations, and treating them as conversations isn't an oversight. It's a category error.

The long-term vision for agentic AI is compelling: less information, fewer screens, the operator simply takes the action. But reaching that vision means building through it, not skipping to it. Users need to see the agent doing the job they recognize before they'll trust it with the job they can't see. Gartner predicts nearly 40% of agentic AI projects will be cancelled by 2027, and a significant part of that will be adoption failure rather than technical failure — the cost of products that skipped the trust-building and jumped straight to the end state.

Every step toward less interface has to be earned. Interface choices in enterprise agentic AI need to come from the user's mental model of their job, not get inherited from consumer products that solved a different problem. The question isn't whether your agent is capable. It's whether the surface you've built matches the way your user actually thinks about their work. We're still figuring that out. The supply chain officer and the analysts are telling us what's wrong. We're listening.

What human in the loop actually means

Human in the loop is often treated as a compliance checkbox — a confirm button before the agent acts. That's not what it is.

Designed well, human in the loop is the trust architecture. It tells users: the agent knows your job, it's already done the thinking, and you're in full control of every decision. Not because we don't trust the agent. Because adoption depends on users feeling confident, not just informed.

The supply chain officer doesn't need less information. She needs the right information, presented in a way that reflects how she already thinks, with clear sight of the agent's reasoning and full control over the outcome. When that's right, the confirmation step isn't friction. It's what makes the product feel like it was built for her.

That surface doesn't design itself. It comes from someone who spent time understanding how she works before anyone started building.

What this role actually is

Design in this context isn't producing artefacts. It's not wireframes handed across a wall. It's not a Figma file attached to a Jira ticket.

By the time a Forward Deployed Designer gets involved, a prototype already exists. Engineers have built something. The client has expectations. The product is taking shape. The value I bring isn't starting from scratch — it's seeing what's wrong before users do: looking at a live experience and knowing whether the trust arc is right, whether the interface choice fits the job, whether accountability actually lives in the surface or is just implied somewhere in the backend.

Why this role matters now

The FDE makes the agent technically possible. The FDD makes it something people will actually use.

Right now, most enterprise AI deployments have one and are missing the other. The model works. The integration holds. The experience fails — not dramatically, but quietly: adoption numbers that look worse every quarter, users who were politely resistant from day one, contracts that don't renew.

In regulated enterprise, where accountability is real and the tolerance for ambiguity is essentially zero, trust isn't a feature you add after the product works. It is the product. And designing it takes someone whose job is specifically to understand users deeply enough to know when the agent is earning that trust, and when it's quietly losing it.