WCAG 2.2 for AI Products: Accessible Agents and Chat UIs
WCAG 2.2 applies to AI products like any other web content, but chat interfaces and agents stress specific criteria: status messages for streaming responses, focus management when content appears dynamically, target sizes for suggestion chips, accessible authentication when agents need sign-in, and understandable generated content. Getting these right makes AI features usable for screen reader, keyboard and voice control users.
This article walks through the WCAG 2.2 criteria that matter most for chat and agent interfaces, with practical patterns. It complements our European Accessibility Act checklist, which covers the legal side.
What is new in WCAG 2.2?
WCAG 2.2 became a W3C Recommendation in October 2023. It keeps the criteria of WCAG 2.1, removes the obsolete 4.1.1 Parsing criterion, and adds nine new success criteria. Six of them are at levels A and AA, the levels most organisations target:
- 2.4.11 Focus Not Obscured (Minimum): focused elements must not be completely hidden, for example by sticky chat composers or cookie banners.
- 2.5.7 Dragging Movements: anything done by dragging must also work with a single pointer without dragging.
- 2.5.8 Target Size (Minimum): pointer targets of at least 24 by 24 CSS pixels, or sufficient spacing, with some exceptions.
- 3.2.6 Consistent Help: help mechanisms such as a chat or contact link appear in the same relative place across pages.
- 3.3.7 Redundant Entry: do not ask users again for information they already provided in the same process.
- 3.3.8 Accessible Authentication (Minimum): no cognitive function test such as memorising or transcribing for login, unless alternatives or assistance exist.
How should streaming responses work with screen readers?
Streaming text token by token is the most common accessibility failure in AI chat. If the response container is a live region that announces every change, screen readers either stutter through fragments or are flooded. If it is not announced at all, blind users do not know an answer arrived. Criterion 4.1.3 Status Messages requires that such updates can be presented to assistive technology without moving focus.
A reliable pattern: announce a short status such as 'Assistant is responding' when generation starts, then announce the complete message, or complete sentences in batches, rather than every token. Provide a visible and programmatic 'stop generating' control. Let users move to the latest message with a single shortcut, and make each message a navigable unit with a heading or list structure.
How do you handle focus and keyboard use in chat?
After sending a message, keep focus in the input so users can continue typing; do not move focus to the response unless the user asks. When the agent opens a confirmation dialog, move focus into it, trap it there while it is open, and return it afterwards. Ensure the sticky input area never covers the focused element, which is exactly what 2.4.11 targets.
Every control needs a keyboard path: copy, retry, rate, expand sources, attach files. Suggestion chips and reaction buttons are often tiny; check them against 2.5.8 and give them accessible names that describe the action, not just an icon.
Want an expert opinion on your product?
Talk to a UX expert or AI Code Engineer — 30 minutes, free, no obligation.
Book a call for freeWhat about generated content itself?
Accessibility does not stop at the container. Generated answers should use real structure: lists as lists, tables with headers, code in code blocks. Images generated or shown by the agent need alternative text, which the system can draft but should not leave empty. Long answers benefit from a short summary first.
Language level matters for users with cognitive disabilities. Offer a way to request simpler language, avoid unexplained jargon by default, and keep instructions for agent actions concrete. AI disclosure labels required under the EU AI Act must themselves be accessible, which means text that assistive technology can read, not only an icon.
Which agent-specific issues should you test?
These checks fit naturally into an agentic UX audit, since accessible agent flows and well-designed agent flows share most of their requirements.
- Can a keyboard-only user start a task, review the agent's plan, confirm or cancel, and undo?
- Are progress updates for long-running tasks announced without stealing focus?
- Do time limits on confirmations allow users to extend them, in line with 2.2.1 Timing Adjustable?
- When the agent needs the user to sign in, does the flow avoid puzzles and allow password managers, per 3.3.8?
- Does the agent reuse information the user already gave instead of asking again, per 3.3.7?
- Is the route to a human or to help in a consistent place, per 3.2.6?
Can AI help make products more accessible?
Yes, with limits. AI can draft alternative text, summarise long content, convert between modalities and power voice interfaces. It can also generate inaccessible code at scale if prompts and reviews do not demand accessibility. Treat AI output as a draft that must pass the same checks as human work.
Automated testing tools catch a portion of WCAG issues; manual testing with assistive technology finds the rest. A focused UX review of your chat or agent interface combines both perspectives.
Frequently asked questions
Do we need WCAG 2.2 or is 2.1 enough?
Many legal requirements in the EU reference WCAG 2.1 AA through EN 301 549 as of late 2026. WCAG 2.2 is backwards compatible and improves usability, so targeting 2.2 AA is a sensible choice.
Should the chat log be an ARIA live region?
Usually not the whole log. Use a dedicated, polite status region for announcements and keep the log as a navigable structure. Test with several screen readers, as behaviour differs.
Are voice assistants automatically accessible?
No. Voice helps some users and excludes others, such as deaf users or people with speech impairments. Offer text and voice in parallel.
Can you audit our AI chat interface for accessibility?
Yes. Our reviews of chatbots and agents include accessibility findings with commented screenshots and concrete fixes. Book a call for free to discuss your product.
Book a call for free
From UX audits and Agentic UX reviews to AI development, agentic coding workshops and publishing your AI-built app on the App Store and Google Play — we help you ship. 24/7 call support in English and German.