Digital accessibility experts and advocates have spent more than two decades making the case that digital accessibility is not optional. Our arguments have rested on a combination of compliance, ethics, and the sheer size of the audience organizations were excluding. Today, this argument has a new dimension, introduced by the rapid rise of AI.
AI agents and the accessibility tree
We know that today’s web is actively being rebuilt around AI agents: tools that browse, click, fill out forms, and complete tasks on a person’s behalf. As recently reported by CNET, agentic AI bots now account for more web requests globally than humans do.
What is less well-known is that the infrastructure making this possible is the same infrastructure accessibility teams have been building for years: a structure comprising semantic HTML, ARIA attributes, and WCAG (Web Content Accessibility Guidelines) conformance. Neil Patel addresses this convergence in his recent video “The Marketing Opportunity of a Decade (But Not for Long),” breaking down how AI shopping agents evaluate a website and pointedly observing that “Accessibility for humans became visibility for AI.”
AI agents perceive web pages in essentially the same way as screen readers. To truly understand the full significance of this, we need look no further than the definitive force in search and AI summaries. Google. A recent Chrome for Developers article on Lighthouse agentic browsing scoring lays things out clearly, referring to the accessibility tree as “a core metric for agentic navigation,” and establishing that “agents rely on the accessibility tree as their primary data model.” They go on to clarify that visibility—”a specific subset of accessibility audits that are critical for machine interaction”—is about “confirming that content is not hidden from the accessibility tree while being interactive.” Their recommendation? “Ensure a sound accessibility tree.”
What is unmistakable is that digital accessibility has never been more important, and we now have even more reasons why the web must be accessible to all.
AI and accessibility: Identifying the gaps
Given this new reality, you’d think digital accessibility would be priority number one for every developer and every engineering leader at every organization with digital properties. Unfortunately, that doesn’t seem to be the case.
According to WebAIM’s 2026 Million report, 95.9% of home pages now contain detectable accessibility failures. That number has gotten worse for the first time in six years.
Meanwhile, engineering leaders trust their AI-generated code. In our own survey of 200 engineering leaders using AI-generated code and coding agents, 88% said they trust their AI-generated code is accessible, and 96% said they actively prompt their AI agents to produce accessible output.
The code itself doesn’t back up that trust. In Microsoft’s evaluation framework, six of 10 AI models failed every automated accessibility test. In the GAAD Foundation’s AIMAC benchmark, 35 of 37 leading AI models produced multiple critical and serious accessibility issues by default.
The same pattern shows up after code ships. 64% of engineering leaders say accessibility issues are a top reason for reworking code after release. 72% say AI-assisted development has increased their compliance risk. But when the same leaders rank their compliance concerns, accessibility comes last.
AI and accessibility: Closing the gaps
None of this means the moment is lost. It means the opportunity is still wide open, and the organizations that close this gap now, before it becomes standard practice, get to lead rather than catch up. The data backs this up: catching an accessibility issue at the design stage costs around $21 and takes less than 15 minutes. Catching that same issue after it reaches production costs over $637, a 30x increase. For a mid-size organization, shifting that work earlier can save more than a million dollars a year.
The takeaway is clear: you need to build accessibility into how AI writes code from the first prompt, not fix it after the fact. And because the disability community has been perfecting human-computer interaction for different input and output modalities for decades, their expertise isn’t just morally important; it’s technically essential for building robust, reliable AI that actually works. Let’s not forget the “accessibility” in the accessibility tree!
AI that works
When I say “AI that works,” I’m talking about trust. About code you can rely on. This is mission-critical for agentic AI. How can we hand over responsibility for our tasks to an agent, if we can’t trust that the agent will perform its tasks accurately and correctly—or even get started, getting lost in the maze of bad markup? The answer is: we can’t. And the answer is the same for everyone—not only people with disabilities. Clear structure, plain labels, and robust and predictable behavior make a page easier for a screen reader to read. They also make it easier for an AI agent to read. And the easier it is for the AI agent to read, the more likely it is to perform its task correctly.
When a person with a disability can’t process a web page, we know what could happen. They might have a frustrating experience that results in the business losing a customer when they abandon the page. They may become frustrated and leave a complaint or a negative review, which can damage brand reputation. They may enter incorrect information without realizing it, or forfeit their privacy rights to get past accessibility barriers, and require someone on the support team to fix it by hand. There are very real time and money costs associated with all these outcomes. In some cases, they may file a legal complaint, and an organization could have a lawsuit to contend with. Inaccessibility means risk.
There are also real-world risks if an AI agent can’t process a page. The agent may abandon the task, or the task may fail outright. It may fall back to reading a screenshot instead of the page’s underlying structure, which is slower and less accurate. It may fill in what it couldn’t read with a guess instead of a fact, and that guess can look exactly like a correct answer: an order placed, a form submitted, a decision made, all based on an AI hallucination. And because an agent can repeat that mistake across many transactions in the time it takes a person to make one, the same error scales far faster than a single human issue ever could.
New solutions for the agentic AI era
Agentic AI may be new, but these problems aren’t. We’ve been developing solutions for them for thirty years. And we can solve these issues for the agentic AI era. We have the lived experience of people with disabilities and the expertise of advocates, allies, and practitioners. We have the tools, technologies, and data. We have the regulations, standards, and guidelines. We have everything we need.
Everything we need, except one thing: alignment. We need uniform buy-in at every level, from the leadership teams making company-wide investments in AI to the development teams who are producing new code every minute of every day. We need everyone to align on the mandatory embrace of digital accessibility.
This is the real choice facing every organization building with AI right now. Treat accessibility as foundational to your AI strategy, and you have a genuine opportunity for your business to thrive in the agentic AI era and lead the market while doing so.
We go deeper into what that opportunity looks like in our new report, From debt to dividend: How engineering leaders can advance accessibility in the AI era. It’s a practical guide for integrating AI to achieve and maintain compliance while minimizing financial, legal, and reputational risk, and building accessibly from the start.
Accessibility at the speed of AI
The web has always been a dynamic, ever-changing place. But what we’re experiencing now is a truly radical shift. The web is being read by more AI agents every day, and that number will only grow. Human-centric digital accessibility and agent-centric accessibility are converging, redefining our very sense of accessibility in the process. Fortunately, despite the seismic nature of this transformation, we have the expertise to handle it. The expertise has always been there. For decades, a global community of disability advocates, experts, and practitioners, many of whom are people with disabilities, has been laying the foundation for an accessible web. We can leverage this expertise. The need is urgent. The time is now. We can achieve accessibility at the speed of AI. We can, and we must.
