NN/G Proposes Experience Design for When the Real User Is an AI Agent

In an article published on October 30, 2026, on the Nielsen Norman Group website, researcher Raluca Budiu proposes a name change that is, in practice, a change of method. The term is "experience design," and the central idea is easy to state and uncomfortable to accept: in many digital products, the user who touches the interface is no longer the human being who has the goal.

Budiu says she was skeptical when NN/G internally changed the job title "user-experience specialist" to "experience specialist." It looked like marketing rebranding. She changed her mind for a reason the company hadn't anticipated: AI agents acting on people's behalf break the premise that sustained the entire discipline, the idea that the human is always the one interacting directly with the company, through whatever channel.

The Problem the Multichannel Journey Didn't Anticipate

UX already went through a similar expansion in the 2010s, when the smartphone became ubiquitous and people started moving between app, website, physical store, and call center to resolve a single task. The answer at the time was journey maps and service blueprints: tools to ensure that all touchpoints worked coherently together.

What didn't change, in any of those iterations, was who was on the other side of the screen. Now it does. A person can ask an AI assistant to do grocery shopping, cancel a subscription, organize photos, or book a flight, and the agent becomes the product's immediate user, operating outside the set of channels the company designed for people.

For Budiu, that's the real reason to abandon "user experience" as the name of the field: the journey doesn't just cross channels, it also crosses multiple actors, some of them not human.

Three Scopes, One Single Goal

The researcher divides the field of experience design into three fronts, and this is where the article becomes useful for those who design products on a daily basis:

• Humans using digital products directly: the traditional UX scope remains intact, usability, information architecture, interaction, content, accessibility. Nobody stops needing to fill out an income tax form alone.
• Agents using systems on people's behalf: a new discipline, which some already call agent experience, focused on how an AI agent consumes an interface or an API to complete a delegated task.
• People and agents within the same journey: the newest point and, according to Budiu, the most neglected today, the place where the experience breaks during control transitions between human and machine.

In short: a product can work well for humans and work well for agents, and still fail when the two need to take turns on the same task.

The Example That Explains Everything: the Plane Ticket

Budiu describes a case that sums up why this third scope matters. A person asks an AI agent to book a flight. The agent recommends an option, the user accepts, and to complete the purchase the agent asks the person to log in on the airline's website. The problem: the user's password manager extension doesn't work inside the agent's embedded browser.

Nothing, in isolation, is broken. The agent works well. Asking for authentication is reasonable. The internal browser was designed for the agent to execute steps and for the person to consume content, not for the person to operate their own tools.

The experience breaks when the task moves from agent to person and the person is asked to use a tool that they normally don't use and that is presumably designed for agents.

Raluca Budiu, Nielsen Norman Group

The central point isn't the bug itself, which is easy to fix. It's that it only appears when someone studies the human-agent system as a whole. Studying the agent alone doesn't reveal the problem; studying the human alone doesn't either.

What Changes for Those Who Design the Product

Designing for agents isn't the same as designing for humans, even when the end goal is the same. An agent can access code, structured data, and APIs directly; a person depends on visual hierarchy, spatial organization, and perceptual cues that make no sense at all to a machine.

This raises questions that traditional UX never had to answer with the same urgency:

1. When should an agent act without asking the person anything?
2. When does it need to stop and ask for explicit confirmation?
3. What should an agent store as memory between one task and the next?
4. What kind of context and which tools should it have access to?
5. At what point does control return to the person, and how can they inspect or reverse an action the agent has already taken?

None of these questions has a generic answer. They depend on the domain, the risk of the task, and how much the person trusts the agent, which means qualitative research remains the foundation, not an extra.

Research Methods Are Changing Too, but Still Lack Validation

Budiu identifies three ways AI is changing research and design work, not just the object being studied:

This table already points to where most product teams will feel the friction first: you can't simply run a classic usability test on an agent and call it valid research.

What Doesn't Change, and Why That Matters

Expanding the field doesn't mean abandoning human-centered design. Budiu is direct on this point: the ultimate success criterion remains whether the person achieved their goal, not whether the agent performed flawlessly on a technical level.

A system aimed at agents that process banking transactions, for example, still needs to protect the financial security and interests of whoever is behind that agent. Blindly optimizing for the agent's success, without considering the well-being of the person who delegated the task, is the kind of mistake experience design exists to prevent.

What Remains Open for Those Designing Products Today

NN/G's article doesn't deliver a closed playbook, and it's honest about that: it names an emerging field, it doesn't describe a validated process. For those who work with product, the practical takeaway is to think of the journey in three simultaneous layers, the interface for the human, the interface (or API) for the agent, and the transition between the two, where most current products have never been tested.

It's worth asking, for any product that today has some AI agent in front or behind the scenes: what happens when the task needs to go back from the machine's hands to the person's hands? If the answer is "we've never tested that," there's the next research item, not a hypothesis, a real blind spot.