AI Accessibility Checklist: Interfaces Everyone Can Use

AI interfaces introduce accessibility problems that conventional pages do not have. This checklist covers them, drawn from FISTA Solutions' web and mobile product work.

What is specific to AI interfaces?

Six issues beyond standard accessibility practice.

Issue	Why it is specific
Streaming output	Text arrives over time
Uncertainty signalling	Frequently visual only
Chat-only entry	Prompting is itself a skill
Variable response length	Unpredictable structure
Review interfaces	Dense comparison tasks
Processing delays	Long waits with no structure

Streaming and announcements

The problem unique to generated interfaces.

• Streaming output announced in meaningful units, not per token
• Live region politeness set so it does not interrupt constantly
• An option to announce on completion rather than progressively
• Generation start and end announced
• Cancellation reachable and announced
• Errors mid-stream announced clearly
• Tested with more than one screen reader

Input alternatives

Chat should not be the only route in.

• Suggested actions offered alongside free text
• Templates or structured inputs available for common tasks
• Voice input supported where appropriate
• Input field does not require precise phrasing to work
• Examples shown of what the system can do
• Previous inputs reusable without retyping
• No time limit on composing input

Keyboard and focus

Standard practice, frequently broken by dynamic content.

• Every action reachable by keyboard
• Focus order logical as content is added
• Focus not stolen when output arrives
• Focus visible at all times
• Generated content reachable and navigable
• Modal dialogs trap and restore focus correctly
• Skip links where output is long

Non-visual information

Anything conveyed by colour or icon must also be text.

• Confidence and uncertainty conveyed textually
• Generated versus human content distinguished in text
• Citations reachable and identifiable by a screen reader
• Status indicators have text equivalents
• Error states described, not just coloured
• Contrast ratios met for all states
• Content does not rely on shape or position alone

Cognitive load

Affects everyone and affects some people decisively.

• Response structure predictable across interactions
• Plain language used, jargon explained
• Long output broken into sections with headings
• Summary available before full detail
• Instructions kept short and concrete
• No requirement to remember earlier content
• Undo available for actions taken from suggestions

Waiting and feedback

Long processing needs structure, not just a spinner.

• Processing state announced, not only shown
• Progress conveyed where the wait is long
• Agent steps described in text as they happen
• No time limits the user cannot extend
• Timeout warnings given with a way to continue
• Failure after a long wait explained clearly
• Results persist so they can be reviewed at the user's pace

What are the most common failures?

Live regions announcing every token. Confidence shown only by colour. Chat as the only input. Focus stolen by arriving output. And testing with an automated checker rather than a screen reader.

Who should own this?

The product team owns accessibility for its interfaces, with specialist review for significant releases. Accessibility owned by a central team that reviews at the end produces findings too late to act on.

How often should it run?

Reviewed at design, tested before release, and audited annually. Any change to streaming behaviour or input methods should trigger a re-test.

What evidence should it produce?

Test records naming the assistive technology used, findings and their resolution, and any user testing conducted. Automated scan results alone are not evidence for these issues.

What about agent interfaces specifically?

They add longer waits and multi-step progress, both of which need non-visual equivalents. A user who cannot see the step list has no idea what is happening during a two-minute task.

Describe each step in text as it begins, keep the sequence reviewable afterwards, and make cancellation reachable throughout. See streaming UI patterns for AI apps.

What should you do first?

Turn on a screen reader and use your AI feature. The streaming behaviour alone usually reveals the work to be done.

How FISTA Solutions helps

FISTA Solutions builds and operates production AI systems through AI agents, AI enablement, and forward deployed engineering: streaming output announced in meaningful units rather than per token, and input alternatives offered so composing a prompt is not the only route in, decisions documented with their reasoning, and handover that leaves your team able to maintain what was delivered. The record is 150+ projects for 50+ companies across 12+ countries.

To adapt this checklist to your environment, message FISTA on WhatsApp, or read streaming UI patterns for AI apps.