Designing interfaces that AI agents can actually use

By on

An industrial robot arm on a workbench

Gartner expects around 40% of enterprise apps to embed task-specific agents by the end of this year. Those agents will use your product without looking at it, and several common design habits stop working entirely.

Gartner expects roughly 40% of enterprise applications to have task-specific agents embedded by the end of this year, up from under 5% in 2025. Whatever the exact number turns out to be, the direction is not in dispute.

Those agents are going to use your product. They are never going to look at it. That single fact breaks a set of assumptions most interfaces are built on.

What stops working

  • Meaning carried only in colour or position. A red badge that means overdue is invisible unless the state is also in the markup.
  • Flows that depend on remembering the previous screen. Anything held in component state and never reflected in the URL or the DOM cannot be resumed.
  • Error messages written as reassurance. "Something went wrong, please try again" gives an agent nothing to act on, so it retries the same failing call.
  • Confirmation patterns that rely on hesitation. A human pauses at a destructive dialog. An agent does not have a flinch.

The error message point is bigger than it looks

Once an agent is a consumer, your error copy is an API contract. It is the only channel through which the system tells the caller what to do differently.

"That date is not available" is a dead end. "That date is outside the booking window, which opens 90 days ahead" tells the caller what to try instead. The second version is better for people too, which is the pattern with most of this.

The upside nobody expected

Almost everything that makes an interface legible to a machine also makes it legible to a screen reader. Semantic structure. Honest labels. State you can address directly rather than infer from pixels. Errors that say what to do.

The accessibility argument for all of this has been on the table for fifteen years and mostly lost budget fights. It is about to get funded for the least romantic reason imaginable, which is that enterprise customers want their agents to work. I will take it.

Where to start

  • Take your most common task and try to complete it from the accessibility tree alone. Whatever you cannot do is what an agent cannot do.
  • Audit your ten most frequent error messages. Rewrite each one to name the constraint and the valid alternative.
  • Make sure every meaningful state has a text or attribute representation, not just a colour.
  • Put the important state in the URL. If a flow cannot be resumed from a link, it cannot be resumed by anything.

Tags: ai, ux, agents