Nielsen Norman Group: AI agents as users
What the article is about
Nielsen Norman Group’s Sarah Gibbons and Kate Moran present a structural argument: the category of “user” now includes AI agents alongside humans, and design practice needs to reflect that. The article does not treat agent compatibility as a future concern — it treats it as a present decision that organizations are either making intentionally or by default.
Context
The article maps out the three ways agents currently interact with digital interfaces: vision-based navigation, where the agent reads screens the same way a human would; accessibility tree parsing, where the agent reads the semantic structure underlying the visual layer; and direct API access, where the agent communicates with backend services without interacting with a visual interface at all.
The practical upshot of this taxonomy is that interfaces optimized for human use often work reasonably well for vision-based agents and poorly for accessibility-tree-based ones — unless the underlying semantic structure is clean. This connects agent compatibility directly to accessibility work. Semantic HTML, clear labeling, logical hierarchy, and predictable patterns benefit human screen reader users and agent navigation alike. Organizations with rigorous accessibility practices have, in effect, been building agent-compatible systems without naming that goal.
Key takeaway
The article’s central recommendation is strategic rather than tactical: decide whether to enable agent access, block it, or design for both, and make that decision deliberately rather than by neglect.
The case for enabling agents is straightforward — products that agents can use become more useful to people who rely on agents to act on their behalf. The case for restricting agent access is also real in specific contexts: ad-supported models where engagement matters, financial or healthcare services with regulatory constraints on automated access, and platforms protecting competitive advantages in pricing or recommendation data.
What Gibbons and Moran flag as a genuine risk is opting out without a reason. If competitors allow agents to act within their products while others block them, user preference may drift over time toward the more agent-accessible option. The article also describes how the gap between human-optimized and agent-optimized interfaces may grow into separate design concerns over the longer term — two distinct surfaces for two distinct populations of users.
Who should read this
UX designers and product managers making architecture decisions about whether and how their products interact with AI agents. Accessibility practitioners will find the section on semantic structure useful as an additional argument for existing investment in accessibility standards.