AI-generated interfaces look polished but often fail WCAG on keyboard access, focus visibility, form labels, colour contrast and custom components built from divs. An audit combines automated scanning, which finds a minority of issues, with manual keyboard and screen reader testing of key user journeys. Under the European Accessibility Act many consumer-facing digital services in the EU must meet these requirements since June 2025.
On this page
Ask an AI builder for a modern dashboard and you get rounded cards, subtle grey text, icon-only buttons and custom dropdowns. It looks professional. For a keyboard user, a screen reader user or someone with low vision, it may be unusable. Accessibility is no longer optional for many products in Europe, and AI-generated interfaces have characteristic weaknesses that a focused audit can find quickly. This article explains what to test, how to test it and how to fix issues efficiently with the same agents that created them.
Why accessibility matters now
The Web Content Accessibility Guidelines (WCAG) are the reference standard; WCAG 2.2 is the current version, and level AA is the usual target. In the EU, the Web Accessibility Directive already requires public sector websites and apps to comply, via the harmonised standard EN 301 549. The European Accessibility Act has applied since 28 June 2025 to many consumer-facing products and services, including e-commerce, banking services, e-books and passenger transport services, with an exemption for microenterprises providing services. In Germany it is implemented through the Barrierefreiheitsstärkungsgesetz (BFSG). Beyond legal duties, accessible interfaces are simply easier for everyone to use, and they tend to be better understood by search and AI engines.
Typical failures in AI-generated UIs
Generators reproduce patterns from their training data, including inaccessible ones. The issues we see most often are:
- Clickable
divandspanelements instead of buttons and links, which keyboards and screen readers cannot operate. - Icon-only buttons without accessible names, so a screen reader announces "button" and nothing else.
- Removed focus outlines for visual reasons, leaving keyboard users without orientation.
- Light grey text and placeholder-only form fields that fail contrast and labelling requirements.
- Custom modals, dropdowns and tabs that do not trap or return focus and lack correct ARIA roles.
- Error messages shown only in red, without text association to the field or announcement.
Component libraries such as Radix UI, which underlies shadcn/ui, used by many builders, handle much of this correctly when used as intended. Problems arise when generated code wraps or overrides them, or builds custom widgets from scratch.
How to run an audit
- Define the scope: the key user journeys such as sign-up, login, the core task, checkout or form submission, plus representative page templates.
- Run automated scans with tools such as axe DevTools, Lighthouse or WAVE on each page and state; automated tools detect only a portion of WCAG issues, so treat results as a starting point.
- Complete each journey using only the keyboard: tab order, visible focus, no traps, all actions reachable.
- Repeat with a screen reader, for example VoiceOver on macOS and iOS or NVDA on Windows, and check names, roles, states and announcements.
- Check zoom to 200 and 400 percent, reflow on narrow screens, contrast ratios and motion settings.
- Document each issue with the WCAG criterion, severity, location and a suggested fix, then prioritise by impact on the key journeys.
Fixing issues with agents
Agentic coding tools are effective at accessibility fixes once the issues are identified. Give Claude Code or Codex the audit findings and ask for changes per component rather than per page, because one fixed component often resolves dozens of instances. Ask the agent to replace clickable divs with semantic elements, add labels, restore focus styles and use the accessible primitives already in the component library. Then add automated checks, such as axe in Playwright or component tests, to CI so new screens do not regress; our article on CI/CD for vibe-coded apps shows where these checks fit. Add accessibility rules to your agent context file as well, so future generated code starts from a better baseline.
Prioritise by user impact rather than by count. A missing label on the sign-up form blocks people from becoming customers at all, while a contrast issue on a footer link is a minor annoyance. Fix the blockers in the key journeys first, then work through templates and shared components, then the long tail.
Re-test manually after fixes. An agent can add an aria-label that satisfies a scanner but says something meaningless.
Accessibility, UX and SEO overlap
Many accessibility fixes improve usability for all users: clearer labels, visible focus, readable contrast and predictable components. They also help discoverability, because semantic HTML with headings and link text is easier for crawlers to interpret, as explained in SEO for AI-built web apps. It makes sense to run accessibility and UX reviews together; our article on UX reviews for AI-built apps describes the broader review.
Getting an expert review
Automated tools and internal checks catch much, but an experienced reviewer finds the issues that block real users. uxreviewer provides structured, expert reviews of apps and websites with concrete action steps, which teams can hand directly to developers or agents. For public sector projects and B2B products with formal accessibility requirements, the OnlineAppCreator custom development team implements fixes and builds accessibility into delivery, including documentation for procurement.
Frequently asked questions
Which WCAG level should my app meet?
Level AA of WCAG 2.1 or 2.2 is the common target and the basis of EU requirements through EN 301 549. Check which version your contracts or regulators reference.
Does the European Accessibility Act apply to my app?
It applies to many consumer-facing products and services, such as e-commerce and banking, since June 2025, with an exemption for microenterprises providing services. Get legal advice for your specific case.
Are automated accessibility tools enough?
No. They detect only part of WCAG issues. Manual keyboard and screen reader testing of key journeys is required for a meaningful audit.
Can AI agents fix accessibility issues?
Yes, they are effective once issues are clearly described, especially at component level. Always re-test manually, because superficial fixes can pass scanners without helping users.