AI Design is Service Design
Practical framework for bridging the frontstage and backstage of agentic systems.
A few months ago, I was using presentation software and decided to try its AI functionality. I wanted to create a first version of a deck and prompted it to do so. The AI replied that it couldn’t do that; it could only generate images for the slide I was on. I was disappointed. On the surface, it looked like any other AI, but it wasn’t aware of the workflow—the presenter’s overarching goal and the steps they would take in that moment. For AI to provide a useful and satisfying experience, someone should have thought not only about the UI and typical interaction patterns, but also about the workflow underneath.
In AI experience design, there are two layers to design.
The first layer is the one you already design for: the human-AI interaction—how a person interacts with AI, delegates to it, reads what comes back, and decides whether to trust it.
The second layer is the agentic experience — the conditions the agents themselves work in: how they’re briefed, what they can reach, the steps that they follow, and how they pass work along.
Think of AIX as the sum of these two parts: the human-AI experience plus the agentic experience.
I'm not the first to talk about this distinction. The field has started calling it AX, or agent experience — a term Mathias Biilmann floated in early 2025 — and it already has two readings: designing the experience humans have with agents, and designing products so agents can use them. That maps closely onto my two layers.
Where I’d push this further is twofold. First, agentic experience should be treated as workflow design: AX is not just about the behavior of one agent, but about how work is sequenced, handed off, and completed. Second, much of the agent-facing work still assumes a single agent as the customer, with clean tool definitions and machine-readable docs, when in reality several agents coordinate to deliver an output. That makes AIX, in a very real sense, a service-design challenge: it is about designing interactions among agents that still create value for people. My former Carnegie Mellon colleagues have been circling this idea for a while. Shelley Evenson was asking years ago what it means to design not just people-to-machine but machine-to-machine interactions that still bring value to people.
If you know service design, you already know the concept of line of visibility. There’s frontstage above it, and backstage below. The frontstage is what people see and feel; the backstage is the work they never see, but that makes the system work. Think of a restaurant: the dining room is frontstage, and the kitchen is backstage.
That said, traditional service design doesn’t cleanly map to AI. The split between front and back stages is not as straightforward as it sounds, so it’s tempting to say human-AI experience belongs to the frontstage and agentic experience belongs to the backstage — but both cross the line. An agent can be frontstage, for example as a customer-facing AI employee in a live demo. And a whole crew can stay backstage, hidden from view.
Humans are on both sides too: the user out front, and the accountable reviewer in back, the human-in-the-loop who signs off before the work ships. Think of middle-office personnel who need to approve the agents’ credit-check analysis before it goes back to the client on the front stage.
So frontstage and backstage are not simply human versus agent; they can involve both. Nielsen Norman Group has argued that AI pushes service design to evolve across both stages, and Jodi Forlizzi takes it a step further: AI is a third kind of actor the classic blueprint never accounted for, so traditional service design has to be rethought with this new actor in mind.
What kind of actor is AI, then?
AI is indeterministic, opaque, and autonomous by nature. It learns and improves itself. In a way, it acts like a sentient being, like a teammate.
I’ve been studying this metaphor for the past two years. I ran a summit and conducted numerous workshops on the topic. What I landed on is that teammate’s a useful north star and makes it easier for people to grasp this new emergent material. However, we also need to be cautious about it, since it comes with caveats. First, it cannot be trusted 100% of the time; it makes silly mistakes. AI reminds one of our guests, Jess Holbrook, the Alf character, who was great most of the time but then ate a cat. Second, when AI makes mistakes, it doesn’t have accountability. The honest shape today is then neither tool nor teammate. It exists in a fluid middle ground.
The role of an 'Orchestrator' can help us mitigate this in-betweenness. An orchestrator role shifts dynamically between human and AI, depending on the task and context. Rather than forcing a single label on the AI, we could focus on designing the orchestrator, ensuring the human retains ultimate accountability while empowering the AI to orchestrate when needed.
The orchestrator
The orchestrator seat becomes the crucial hinge where the two actors meet. While the active direction of tasks can shift to an AI—acting as the agent orchestrator—the ultimate responsibility cannot. Someone has to answer for the result, and accountability is the one thing that doesn’t move. It always stays with the human. Underneath, the AI doesn’t run its agents as a flat peer team either: the AI becomes an orchestrator that directs, and the specialist agents report up; direct peer-to-peer chatter is the fragile exception. That accountable human reviewer from a moment ago lives right here, tethered to this shared seat, stepping in and out of the active flow to make the final calls.
This is also where a framing I’ve been working with fits—what I’ve called the AI’s distributed personalities: the way a single assistant shifts between advisor, assistant, and executor as the moment demands. That shifting belongs to the human-layer experience. The executor mode—the AI going off to actually do the work—is powered by the backstage and by the orchestration underneath. Designing this fluid handoff, and ensuring accountability remains crystal clear even as the AI takes the executor reins, is the core of intentional AIX design.
So what should you carry out of this if you design or research AI experiences for a living? Three things.
Think in Stages like a Service Designer, Not Just Surfaces. Don’t only design the surface the user touches; design how the agents behind it deliver that surface. Understand the workflow, then design the agents and agentic workflow as well.
Author the Behavior, Not Just the Screens. Notice that our medium has changed. Agent behavior isn’t screens and flows. You author it in system prompts and markdown — closer to a character builder or a world builder, using your design chops on a new material. What you’re writing there is tone: when the agent should act versus ask, what it refuses.
Bridge the Front and Backstage. Design the AI across both stages as one holistic thing. That’s usually where things break. If we stay on the surface of a single stage and leave the agents’ voice — how the agents actually work — to someone else, the AI experience comes out incomplete, like the AI assistant in my presentation workflow. The holistic system view is something you intentionally craft with your product and engineering peers. Which means designers have to be in the driver’s seat not only in the front-stage decisions but also in the backstage, such as shaping the model behavior
And that’s a wrap for this issue. Until next time, take good care of yourself and your loved ones.
References
Shelley Evenson, How (might we) design the ideal integration between humans and machines?
Jodi Forlizzi (2025), Service Design Tools Can Inform the Design of Agentic AI Services.
Nielsen Norman Group (2025 How Service Design Will Evolve with AI Agents
Mortati & Viana Mundstock Freitas (2026) Journal of Service Research
Mathias Biilmann (2025) “Agent Experience (AX).”
MAST failure taxonomy (Cemri et al.) arXiv:2503.13657
Nice write-up, thank you!
- Hakan