Key takeaways
- The single-human-at-a-screen model of UX is breaking as AI agents act autonomously on users' behalf.
- Design for two layers: the human principal (intent, trust, control) and the agent operator (clean, machine-readable structure).
- Let AI handle the average user; spend human research budget on edge cases and bad days where trust actually breaks.
- Experience design now spans discovery, shaping AI-native builds, and continuous post-launch evaluation — not just the final screen.
- Keep humans in the loop at every irreversible, high-stakes or trust-critical handoff.
For twenty years, the user sat at the centre of everything we built. Personas, journey maps, usability tests — all of it orbited a single human tapping, scrolling and deciding. That model is now cracking. When an AI agent books the flight, fills the form and completes the checkout, the human never touches your interface. So the uncomfortable question for every product team becomes: who is the user now, and what are you actually designing for?
At Evocative we see this not as an existential crisis for UX but as a long-overdue expansion of its scope. The field is quietly dropping the narrow definition of the user and widening its lens to include the agents acting on people's behalf — while still, crucially, serving human goals underneath.
Why the single-user assumption is breaking
Recent thinking from the Nielsen Norman Group and the wider design community points in the same direction: AI-native software development is collapsing the old lifecycle. Code gets generated in hours, not sprints. In that environment, UX risks being reduced to mere "context" fed into a model — a prompt, a guardrail, a bit of tone-of-voice. That would be a serious mistake.
The other shift is that software increasingly acts autonomously. An agent wakes up before you do, scans your inbox, drafts replies and reorders supplies. The last pillar of user-centred design — that a person is present at the moment of interaction — simply isn't guaranteed anymore. Several major platforms pushed this change almost simultaneously, which tells you it's a structural move, not a fad.
Two users, not one
The practical reframe we recommend to clients is to stop designing for a single actor and start designing for two distinct layers:
- The human principal — the person whose goals, trust and accountability are ultimately at stake. They set intent, grant permission and judge outcomes.
- The agent operator — the AI acting on their behalf, which needs clean, machine-readable structure, clear affordances and reliable feedback to do its job well.
These two have different needs. The agent wants unambiguous data, stable APIs and predictable states. The human wants transparency, control and the ability to intervene when something feels off. Good experience design now serves both at once — and never lets the agent's efficiency quietly override the person's interest.
Where human research still wins
There's genuine temptation to let synthetic research do the heavy lifting. AI can simulate the "average" user response surprisingly well, and for high-volume, common-path validation it's genuinely faster and cheaper. We're not precious about that; use it where it earns its keep.
But the average day is rarely where products fail. The real risk lives on the bad days — the edge cases, the frustration, the moment a user's trust snaps because an agent did something irreversible without asking. Synthetic models smooth those over precisely because they regress to the mean. Capturing lived frustration, ambiguity and emotional stakes is still stubbornly, valuably human work. That's where your research budget belongs.
What experience design means now
The discipline map hasn't vanished — interaction design, service design, UX research, visual design and DesignOps are all still here. What's changed is the surface area. "Experience design" is becoming the better umbrella term precisely because "user experience" implies a single human at a single screen. The work now spans:
- Discovering real needs before anything is built, so agentic systems optimise for the right outcomes.
- Shaping what gets built, feeding structured intent into AI-native pipelines rather than decorating the output afterwards.
- Evaluating before and after launch, because autonomous systems drift and need continuous oversight, not a one-off usability test.
A practical starting point
If you own a product roadmap heading into agentic territory, start small and concrete. Map one journey where an agent could act autonomously. Identify every point where the human must stay in the loop — permissions, irreversible actions, money, reputation. Design those handoffs deliberately. Then make your interfaces machine-legible: semantic structure, clear states, documented actions. You'll find this also makes them more accessible and more robust for humans, which is a healthy sign you're on the right track.
The user isn't disappearing. The person — their goals, their trust, their bad days — matters more than ever. What's changing is that you now design the path between their intent and the systems that execute it. That's a bigger job, and a more interesting one.
Sources
- Jev: How to use it for better, faster, and cheaper user decisions in UX — UX Collective
- Let AI have the average user: Why User Research should focus on bad days — UX Collective
- UX Disciplines: Glossary — Nielsen Norman Group
- The agent woke up before you did — UX Collective
- The Role of UX in Agentic Software Factories — Nielsen Norman Group
- Experience Design: Dropping the “User” from “User Experience” — Nielsen Norman Group
