How digital Accessibility is pushing Artificial Intelligence
This post is about how accessibility can help AI, specifically AI agents.
I have spoken several times about how AI can have positive effects on accessibility. However, there is also the reverse perspective: accessibility could also be of great importance for the development of AI—especially for autonomously operating agents.
This refers to AI agents that independently perform tasks on websites or within applications. Whether this form of operating graphical user interfaces will actually catch on remains to be seen. However, the trend is pointing in this direction.
What makes this interesting is that accessibility could potentially be a crucial prerequisite for such agents to function reliably. This would add another argument in favor of accessibility in the future: it could benefit not only humans, but also improve interactions with AI.
Whether this scenario comes to pass remains to be seen. At present, it is more of a forecast than a guaranteed development.
Accessibility and SEO - A Strange Pair
A related topic is the relationship between accessibility and Search Engine Optimization (SEO). Websites that are easily accessible for assistive technologies often feature characteristics that also benefit search engines.
One example is a clear heading structure. Headings make orientation easier for both humans and search engines and make the thematic relevance of a text more explicit.
JavaScript can also play a role. If content is generated dynamically via JavaScript, it can be harder to access for certain user groups and, under certain circumstances, for search engines as well. Although Googlebot can execute JavaScript, it cannot always do so like a full browser.
The same applies to alternative text and image descriptions. They provide additional information about image content and can thus support image recognition and image search.
Subtitles are another example. They make audiovisual content more accessible while simultaneously providing text data that can be used to train and improve automatic speech recognition.
A lean and cleanly structured website can offer benefits as well: it consumes fewer resources and can be processed more efficiently. However, I would no longer unconditionally support the former assumption that fast, lean websites are fundamentally indexed better.
Overall, there are clear overlaps between accessibility and search engine optimization. Accessibility may never have been the decisive factor, but it could certainly offer an additional advantage. That is why some SEO specialists have engaged with the topic of accessibility—and some still do today.
The Agentic Web Becomes Accessible
What we are currently experiencing is the beginning of a development in which AI agents could increasingly take over tasks on websites. If this vision is realized, they will independently fill out forms, purchase products, book trips, process government forms, or even file tax returns. In short: they could take over many of the tasks we dislike doing ourselves.
A major challenge for AI agents is the enormous demand for tokens and, consequently, computing resources. Fundamentally, there are two ways an agent can analyze a website.
- The first and rather inelegant method is analyzing screenshots. While this can be helpful in individual cases, it is inefficient and can lead to erroneous results.
- The more obvious approach is accessing the DOM—the Document Object Model. Here, the agent attempts to recognize which elements serve which function based on the website's structure: Is this a button? Is it meant to submit the form? Is this field intended for the first name, last name, street address, or IBAN?
The problem: such information cannot always be derived reliably from the DOM. Unique labels are frequently missing, or the available information does not allow for a clear conclusion about an element's function. In those cases, the AI must additionally analyze the visual interface and infer, for example, which input field belongs to which label based on the spatial arrangement of elements.
This becomes particularly difficult with purely graphical control elements. With an icon, it may not even be clearly recognizable whether it is interactive, what function it serves, or what happens when clicked. An AI can deduce such relationships using training data and comparable websites, but that process is effort-intensive and prone to errors.
This is where the Accessibility Tree comes into play. It contains the information that assistive technologies also access. Put simply, it is a simplified representation of the DOM reduced to the information relevant to accessibility.
The DOM describes the entire structure of an application. The Accessibility Tree, on the other hand, contains the information relevant to assistive technologies: Which elements exist, what function do they have, how are they labeled, and what is their current state?
For AI agents, this offers significant advantages—provided the website has been correctly implemented for accessibility. Elements become uniquely identifiable and labeled. For example, the agent can recognize that an element is a button, an input field for a first name, a checkbox, or a language selection dropdown.
This also makes the function of an element explicit: Does text need to be entered, an option selected, or a button activated? At the same time, the agent can recognize the status of an element, such as whether it is enabled, selected, or expanded.
Another advantage is clear and understandable labeling. This is precisely what the WCAG requires for forms. In practice, however, labels are often ambiguous or not correctly linked to their corresponding input fields. A cleanly constructed Accessibility Tree provides the AI agent with far more reliable information.
The same holds true for error messages. If they are correctly assigned to a form element—for instance via aria-describedby—and phrased clearly, the agent can identify what problem occurred and how it can be fixed.
As a result, accessibility could become a crucial factor for AI agents: the cleaner a website is structured for assistive technologies, the easier it is for an AI agent to understand its functions and reliably execute tasks—ideally even without human intervention.
Another major advantage is the clear structure of an accessible website. Navigation elements can be clearly identified, and a breadcrumb shows where you are within the page hierarchy.
The sequence of content and control elements is also structured logically. The WCAG refers to this as Meaningful Sequence and Focus Order. A logically built sequence makes it easier for an AI agent to navigate reliably through a website.
Why We Can Be Hopeful
But why should this work any differently for AI agents than it did for search engine optimization? There, too, the benefits of a clean structure were repeatedly emphasized without resulting in widespread website accessibility.
First of all, it is by no means certain that AI will actually become this driver for accessibility. It is a plausible assumption currently being discussed among experts, but development could take a different path.
However, there are several good reasons why AI agents stand to benefit from cleanly structured and accessible applications. One essential factor is efficiency. While AI agents can handle complex applications, they likely prefer those that contain less clutter and fewer potential sources of error.
Imagine, for instance, an online shop that contains numerous advertisements, recommendations, and other content alongside the actual product. A leaner application could be significantly more efficient for an AI agent. The less information that needs to be processed and the clearer the UI controls are, the fewer tokens and computing power are required.
This applies to navigation as well. If, for example, two buttons have similar labels, an agent is more likely to favor an application where functions are clearly named and distinguished from one another.
A cleanly structured application can therefore save tokens, reduce errors, and speed up task execution. If tokens remain a relevant cost factor in the future, this efficiency is likely to become increasingly important for AI agents.
This concept also translates to selecting between different websites. If an AI agent can choose between two travel portals, for example, and one of them is leaner, more clearly structured, and accessibly implemented, it might prefer that portal. From its perspective, it would be the faster, more reliable, and resource-efficient way to book the desired trip.
An intriguing question is whether we will see separate websites for humans and AI agents in the future. For example, a stripped-down interface for agents alongside a richer user interface for humans is conceivable.
What seems more likely to me, however, is a hybrid form: websites will continue to be geared toward human users, but will overall become leaner and more structured. If AI agents account for a significant portion of traffic in the future, providers can hardly afford to ignore that traffic or run separate portals for humans and machines.
Whether this development actually occurs cannot yet be said today. However, the chances are good that AI will become another driver for accessibility. If accessible websites work more efficiently and reliably for AI agents, an economic argument for accessibility will emerge alongside the social and legal ones.