Designing for Two Users: Service Design in the Age of AI Agents

What is agent experience?

Agent experience, or AX, is how easy it is for an AI assistant to understand a business and complete a task for someone. Mathias Biilmann, the chief executive of Netlify, introduced the term in January 2025 and set it in a clear line: Don Norman's user experience in 1993, developer experience in 2011, and now agent experience. His central point was that companies need to treat agents "as a persona" they design for.

That word, persona, is where a service designer's ears prick up. We know how to design for a persona. We just haven't had one like this before.

Is this real, or still a demo?

It is early, but it is real. In 2025, OpenAI and Stripe launched checkout inside ChatGPT with an open standard for other shops to join. Google announced a standard for AI agents to make payments, built with more than 60 payment and technology companies, and began testing restaurant booking inside its AI search. In June 2026, BrightEdge (which sells tracking tools, so read the figure as a signal) reported 88 requests from AI agents acting for users for every 100 visits from ordinary search.

For the practical, business-owner version of this, including a readiness checklist, UX Caribbean has written From UX to AX. Here I want to look at it as a designer.

What does the agent need that a person doesn't?

A person can cope with a lot. They can read a photo of a price list, work out that "from TT$450" probably includes breakfast, phone to check, and forgive a clumsy form. An assistant is less forgiving in some ways and more demanding in others:

• It needs facts stated, not implied. Anything in an image, a PDF or someone's head effectively doesn't exist.
• It needs to know what each control does. Many AI browsing tools read a page through the same hidden labels screen readers use. OpenAI's guidance for its Atlas browser says better accessibility helps its agent use buttons, menus and forms correctly.
• It needs predictable steps. A booking flow with a surprise pop-up, a step that only works on a phone, or a "WhatsApp us to confirm" is a dead end.
• It needs to prove it has permission. Google's payment standard is built on the idea of recording exactly what the customer approved before the agent pays. Consent becomes part of the design.

How do the classic service design tools change?

This is the part I find most interesting. The tools don't need replacing. They need one more lane.

• Personas gain a machine persona. Alongside "Sasha, visiting from Toronto", you now write down what her assistant needs: which facts, in what form, and what it is allowed to do without asking her.
• The service blueprint gains an agent lane. The service blueprint, as Shostack introduced it and Bitner and colleagues later refined it, maps the customer's actions, the visible staff actions, the backstage work and the support systems. Add a row for the assistant: what it reads, what it does, and where it hands back to the person.
• The line of visibility becomes the line of access. It used to separate what the customer sees from what they don't. Now it also decides what an assistant may read, what it may do, and what always needs a human.
• Moments of truth move into the chat. The first impression, the price check and the decision now often happen inside an assistant's answer. The business is judged there before anyone sees its website.
• Failure paths matter more. Every good blueprint shows what happens when something goes wrong. Now you also design what happens when the assistant can't finish: a clear handover to a person, not a silent drop-off.

Why does it all start with knowing your own facts?

Because you can't design a clean agent journey on top of a messy business. In my last article I described how service design techniques such as shadowing, mapping and testing on real cases now produce the material AI systems run on. AX is the same idea turned outward. The prices, availability, terms and exceptions that staff carry in their heads have to be written down in one checked place first. That record becomes the source for the website, the listings, the structured data and, eventually, the answers an assistant uses to book.

In practice, the order is always the same:

1. The brain: capture what the business knows, in one checked place.
2. The layers: make it findable and understandable, in search and in AI answers.
3. AX: make it easy for an assistant to act, with a person approving what matters.

What should a small business do now?

Very little that is exotic. Most of the agent-ready work is also good design for people: plain facts, labelled buttons, short booking steps, accurate listings on the platforms customers already use, and a person confirming anything that commits money or time. The standards are new and competing, and many of the early features are limited to the United States, so I wouldn't build anything elaborate yet.

What I would do is add the second user to the map. The next time you walk through how a customer books with you, walk through it again as their assistant. Wherever it gets stuck, a hurried customer probably does too.