Your Website Works For Humans, But Does It Work For AI Agents?

Margarita Simonova is the founder of ILoveMyQA.com.

gettyA customer visits your website, finds a product, adds it to the cart and completes checkout. The flow works on desktop and mobile. The payment goes through and your QA team signs off.

Now, imagine the same customer asks an AI assistant to do it for them.

The assistant finds the product, but cannot tell which button adds it to the cart. Or, it selects the wrong size because the options are poorly labeled. Maybe it reaches checkout and gets stuck on a form that makes perfect sense to a person looking at the screen.

This is a customer experience problem QA teams need to consider.

We have spent years designing websites around human behavior. We test whether people can navigate menus, find information, complete forms and recover from errors.

AI agents introduce another way of using those same websites. Instead of asking for directions, a person may ask an assistant to compare products, book an appointment or handle a purchase.

Google has introduced standards for agent-driven commerce, while Chrome is exploring ways for websites to describe available actions directly to browser agents.

Not every company needs to support every agent, but websites may no longer be used only by people clicking through them.

A human visitor can often work around small usability problems. We understand that a shopping-bag icon probably means cart. We can see that a calendar opens below the date field. We notice a confirmation message in the corner.

An agent may depend on the labels, page structure and feedback it can actually interpret. A button without a useful name, a custom dropdown or a message that appears visually but is not clearly exposed to software can turn an otherwise simple task into guesswork.

These are not entirely new problems. Accessibility teams have been asking developers to use meaningful labels and predictable controls for years. Now, others benefit from getting those basics right.

Still, accessibility and agent compatibility are not identical. A website can meet accessibility expectations and an agent can still misunderstand business rules or fail to complete a complex workflow. Both need to be tested.

Imagine a customer asking an AI assistant to book a hotel room for two nights, with free cancellation and a total price under a certain budget.

The agent might find the right hotel and select the dates, but does it understand the difference between a refundable rate and one that only allows cancellation within 24 hours? Does it notice resort fees? What happens when the final price changes at checkout?

A human tester might have verified that every field works. That does not tell us whether an agent can complete the customer’s actual request correctly.

I would start with a few common tasks and test them from beginning to end. Not just whether the agent reaches the confirmation screen, but whether it chose the right options, understood the conditions and stopped when it needed the customer’s approval.

A successful click is not always a successful outcome. If the reservation is wrong, the customer is unlikely to care whether the mistake came from a human or an AI agent.

There is also a security and privacy question.

If agents can interact with websites on behalf of customers, businesses need to decide what they should be allowed to do.

Reading product information is different from changing an address, canceling a reservation or submitting a payment. Some actions should require explicit confirmation. Others may need a clear way to reverse a mistake.

And there are legitimate reasons for sites to restrict automated traffic. Supporting agents should not mean removing fraud protection or letting unknown software act without limits.

The goal is not to make every action available to every agent. It is to make permitted actions understandable, predictable and safe.

I would not begin by redesigning an entire website for AI. I would begin with the customer journeys that matter most.

• Can an agent identify the right product, service or action without guessing what a button means?

• Can it complete forms and understand errors when required information is missing?

• Does it notice important conditions such as price changes, cancellation rules or subscription terms?

• Does the process stop for confirmation before an agent makes a consequential decision?

• Can the team see where an agent got stuck, what it attempted and whether it completed the customer’s original task?

Run these checks across the real situations your customers face. A straightforward purchase may work, while a discount, an address change or an expired session causes the same agent to fail.

It is also worth checking that improving a flow for agents does not make it worse for people. Those are not competing goals.

I would treat agent usability as an emerging part of digital quality, not another trend to chase.

Start by finding out whether your important workflows are understandable to an automated assistant. Identify where the experience breaks. Decide what agents may do and where customers must stay in control.

Then let QA test the full outcome, not just the interaction.

We have always tested websites to make sure customers can use them. Now, we need to consider what happens when a customer sends an AI agent instead.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?