I have spent a lot of time recently working with AI agents, particularly ChatGPT, and one of the most useful things I have done is simply watch them work.
Not just looking at the answer they produce, but watching an AI agent navigate a website, interpret instructions, make decisions and attempt to complete a task.
It has changed the way I think about AI, websites and digital experience.
First lesson: AI can be a very capable teaching assistant
My initial experiments were largely around education.
Give an AI agent access to a clearly defined task, along with a good Standard Operating Procedure (SOP), and it can work through surprisingly complex processes.
That distinction is important. The AI is not necessarily effective because it is “clever”. It is effective because it has been given a clear process to follow. That has obvious implications for education.
We can use agents to help students practise processes, demonstrate how tasks are completed, review work against defined criteria or support repetitive teaching and administrative activities.
But watching the agent work also reinforced something I already believed: the quality of the process matters just as much as the quality of the AI.
A vague instruction produces unpredictable behaviour. A well-designed SOP gives the agent boundaries, context and a route through the task.
In many ways, designing an AI agent is beginning to look remarkably similar to designing a good employee onboarding process.
Then I realised I was watching a UX test
Something else became apparent while watching the agent navigate websites. I was effectively watching a usability test. The agent would sometimes hesitate. It would misinterpret navigation. It would struggle to identify which button represented the next logical step. It would encounter pages where the visual hierarchy made perfect sense to a human but was much harder for an automated agent to interpret.
That makes AI agents potentially interesting tools for UX analysis.
Watching an agent attempt a customer journey can highlight confusing navigation, unnecessary steps, inconsistent terminology and poorly structured information.
If an AI agent cannot understand what the website expects it to do next, there is a reasonable chance some human users are struggling too.
But this led me to a much bigger realisation…
UX may no longer be enough
For years, digital teams have designed websites around User Experience (UX). How quickly can somebody understand the page? Can they find the information they need? Can they complete the task with as little friction as possible?
Those questions remain important.
But increasingly, the entity visiting your website may not actually be a human.
It might be an AI agent acting on behalf of a human.
An agent might be asked to:
- compare courses;
- research suppliers;
- find product specifications;
- book an appointment;
- identify the best option;
- complete a form;
- check availability;
- gather information from multiple websites;
- or eventually complete a purchase.
That introduces another design challenge.
Agent Experience — AX.
What does good Agent Experience look like?
A website designed for humans often relies heavily on visual clues. We understand that the large button is important. We recognise that something that looks like a card can probably be clicked. We understand icons, layouts, colours and conventions because we have encountered them thousands of times.
An agent may interpret that environment differently.
For an agent, clarity is incredibly valuable.
- Clear headings.
- Descriptive links.
- Consistent terminology.
- Logical page structures.
- Structured data.
- Transparent pricing.
- Explicit product attributes.
- Predictable navigation.
- Well-labelled forms.
- Accessible content.
- Clear confirmation that an action has been completed.
Interestingly, many of these things are already considered good practice for UX, accessibility and SEO.
AX does not necessarily replace those disciplines. It adds another audience.
The website of the future may have two users
This is perhaps the biggest change in my thinking. When we design a digital experience, we may increasingly need to ask two questions:
Can a human understand this?
And:
Can an AI agent understand and act upon it?
Sometimes those requirements will overlap. Sometimes they may not.
A beautifully designed interactive product selector might create an excellent human experience while being difficult for an agent to interrogate.
A technically perfect database of products might be very easy for an agent to understand but provide a terrible experience for a human visitor.
The challenge will be designing for both.
From mobile-first to AI-first?
The web has gone through similar transitions before. We moved from designing primarily for desktop screens to thinking mobile-first. Search engines encouraged businesses to think carefully about how machines crawled and interpreted their websites. Accessibility encouraged us to recognise that users interact with digital content in many different ways.
AI agents may represent the next shift.
Perhaps websites will increasingly need to become AX First.
Not because humans stop visiting websites, but because AI agents increasingly become the intermediary between the organisation and the customer.
Your next customer may never personally browse your website.
Their agent might.
And if that agent cannot understand your products, navigate your content or complete a task successfully, your organisation may effectively become invisible to that customer.
A new digital question
Watching ChatGPT navigate websites started as an experiment in using AI to support teaching. It became an unexpected lesson in UX. But the most important insight was the emergence of another design consideration altogether.
For years we have asked:
“Is this website easy for people to use?”
Increasingly, we may also need to ask:
“Is this website easy for AI to use?”
That is the question I think digital businesses need to start exploring now. Because the next generation of web traffic won’t just be visitors. It will be agents…
