Survey data from more than 2,500 professionals shows AI has accelerated parts of the software development lifecycle unevenly. Respondents reported gains in planning (58%), design (61%), and development (51%), but only 30% said AI had sped up QA and testing. That split creates a two-speed workflow: teams race through design and build, then stall when someone asks whether the result is actually accessible. Accessibility has long been treated as a QA bottleneck, and AI is now pushing more work into that stage, faster.

The stakes are visible in the data. WebAIM's latest scan of the top million home pages found automatically detected errors per page rose 10.1% in a year, from 51 to 56.1. WebAIM cited AI-assisted coding as one possible contributing factor.

Mature Practices, Not AI Adoption, Set Teams Apart

Among organizations improving across every accessibility metric they track (the "improvers"), AI adoption made the smallest difference of anything measured. Improvers use AI at roughly the same rate as everyone else; their edge lies in the maturity of surrounding practices such as training, communication, and a culture of accessibility. This is survey data, so it shows correlation only, and mature programs may simply be better at tracking and reporting wins.

The gap widens further on specific uses. Professionals in mature accessibility programs were nearly six times as likely to say AI helps them prioritize issues, and nearly three times as likely to say it helps them address accessibility more proactively during development. When context, testing, environments, and guardrails are in place, agents can take on a large share of the work from drafting through operations, reducing time lost to waiting and handoffs. People remain responsible for deciding what to build and approving what ships.

One Diagnostic Exercise

Trace a recently shipped feature from the first Figma frame to production, marking every moment a person decided whether it would work for someone using a keyboard, a screen reader, or a browser at 400% zoom. If most marks cluster at the end of creation, the team will likely either ship barriers or spend upstream savings fixing them later, when fixes are harder and costlier.

Embedding Judgment: Find, Fix, Prove

As AI absorbs more work, people need to stay involved at three points: finding issues, fixing them, and documenting evidence. Each attaches to an existing step such as a handoff, a pull request, or a review, so nothing new is added to the timeline.

Find: Check While Context Is Fresh

  • Start at the handoff. Spell out focus order, accessible names, states, heading structure, and what gets announced when something changes or fails; anything vague gets filled in with model defaults. Put it where the model reads it: design system components, rules files (Cursor rules, AGENTS.md, Copilot instructions), and prompt templates, not just a spec document. GitHub's Annotation Toolkit provides Figma components for marking headings, landmarks, focus order, and keyboard behaviour on the canvas so intent travels with the file.
  • Gate the pull request. Run checks in CI/CD that block or warn on merge, and compare against a baseline so only changes are reviewed. AI-assisted pull requests arrive faster than reviewers can read them, so the automated gate matters more, not less. In React projects, eslint-plugin-jsx-a11y flags issues before commit, while Pa11y and Lighthouse CI scan builds for new errors. Level CI also brings accessibility testing into CI/CD, including pre-merge detection, delta findings, quality gates, and fix recommendations in pull requests.
  • Automate the sweep; route judgment to people. Run automated checks across pages, screens, and documents for broad coverage—QualWeb, for example, checks pages against the W3C's ACT rules. Testing across several engines such as axe-core and Equal Access can help catch what one misses. But usability still needs human assessment: a dropdown can pass automated checks while its submenu remains inaccessible to keyboard users.
  • Give every check an owner. Designers specify focus, errors, and states; developers run a keyboard pass and a screen reader smoke test on custom components (screen reader testing can require introductory training); UX or QA navigates the full journey. Put accessibility metrics in acceptance criteria and the definition of done, backed by component-level tests, because "everyone owns it" tends to become "nobody did it."

Fix: Act Before the Context Cools

When issues become tickets, the context should travel with them: the element, the reproduction, the relevant WCAG criterion, a suggested fix, and the PR link.

  • Put guidance next to the work. Wherever the fix happens—editor, design file, or component library—the answers should be at hand. An MCP server can pull accessibility findings and fix guidance into an AI coding assistant so developers work without leaving the editor; Level Access offers one that connects to tools such as Cursor and Claude.
  • Let AI draft, with guardrails. Assistants optimize for a passing check, not a correct experience. Supply grounded fix guidance for the specific issue and treat output as a draft. The accessibility instructions file in the Awesome Copilot collection steers Copilot toward accessible patterns. After a draft, re-run the check and review semantics and behaviour, since a green result can come from markup that looks right and isn't—for example, role="button" on a div, redundant ARIA, or invented alt text.
  • Review before merge. Test the change in context: did it solve the original problem without breaking anything else, and does the interaction now behave and announce as intended? This gate does different work than the per-check testing earlier—that caught the issue; this confirms the fix holds.
  • Work the backlog by impact. Pull older findings starting with anything blocking a key journey, such as paying a bill or resetting a password, then take the quickest fixes. Include the component, what went wrong, and the remediation guidance.

Prove: Leave a Receipt

Every check, fix, and judgment call should leave a trail. A fix isn't done until something fails if the issue returns: have the AI assistant write the regression test alongside the fix and record the finding link and any judgment call in the PR description. A passing test shows a known issue hasn't returned; it doesn't show the experience works. For what automation can't cover, add a short PR note covering what was tested, with which assistive technology, and any judgment calls made.

Tooling can later assemble what was tested and who approved it from those PRs and tests. The Level Access Reporting Agent cut VPAT/ACR authoring time by more than 70% this way, but only because the evidence already existed; the tool compounds a practice rather than replacing one.

Move the Judgment With the Speed

Closing the validation gap isn't about adopting more AI everywhere or abandoning the toolset that made teams faster. The opportunity is to build the practice around that toolset so human judgment is flagged, embedded, and acted on earlier and more often.

Smashing Editorial