One of the lesser adopted capabilities in the AI wave is accessibility. With the current capabilities, there is no reason to still treat accessibility as negotiable.

That reason usually has less to do with principle than with familiar excuses: the customer did not ask for it, the budget is tight, the regulator is not forcing it, the team lacks expertise, there are more pressing issues.

I think the opposite is true.

As interfaces are becoming more dynamic, more stateful, and more autonomous, accessibility becomes more important, not less.

Many of us are already starting to interact with software through mediation. I increasingly ask agents to move across multiple websites, gather evidence, group it, reason over it, and bring back a finding. Once that happens, the interface is not just a screen I click through directly. It is also terrain that software traverses on my behalf.

People who rely on assistive technology have been living with mediated interaction for much longer. In that sense, the rest of the industry is belatedly moving toward a condition accessibility work has understood for years: interfaces need to stay legible under mediation.

Agentic systems intensify that requirement. These interfaces are full of live status changes, shifting focus targets, generated summaries, expandable evidence, inline approvals, and keyboard-driven intervention points. If those things are not built on top of sound semantics, the interface gets harder to use exactly where clarity and consistency matter most.

That is why semantic markup, focus management, keyboard operation, status announcements, and contrast are not cleanup work. They are product readiness.

Current LLM tools also weaken an old excuse. A straightforward prompt can often help implement semantic structure, keyboard handling, focus behavior, and status messaging. Another can help review flows against a checklist and catch some obvious regressions before they ship. That does not replace accessibility expertise or user testing. It does make it harder to argue that accessibility is too specialized or too slow to begin.

There is a second implication. Users of assistive software, and the tools they rely on, will adapt to this complexity too. They will bring their own agents and machine mediation to interfaces that are getting harder to parse.

The field should design for that. We may need an agent-understandable spec for interface intent and behavior, something closer to a source map layered on top of semantic HTML, ARIA, and WCAG, so a user’s own agent can understand what changed, what state the interface is in, what actions are available, what interruptions are pending, and what reliable paths through the workflow exist.

That is not only an accessibility-side story. It is a general interface story. More and more users who do not usually think of themselves as accessibility users are now adopting workflows with the same shape: delegated traversal, machine summarization, machine interpretation, supervised action. One side of the world just got there earlier.

And the baseline here is not mysterious. WCAG 2.2 and the WAI-ARIA Authoring Practices Guide should be closer to the beginning of the spec, not the end.

Too much AI product work still treats accessibility as a compliance pass after the demo works. That is backwards.

Accessibility belongs earlier in the spec for agentic products not as charity or polish, but as core interface engineering.

Previous: Not every app should become an agent

Related: The prompt is not the product

Sources and grounding: WCAG 2.2 and the WAI-ARIA Authoring Practices Guide.