The data is clear: AI is accelerating design and development, but accessibility testing is being left behind. The result is a web that is measurably less accessible than it was a year ago, and the gap is likely to widen unless teams rethink where human judgment fits into their workflows. This is not a problem that better AI will solve on its own. As the Level Access report shows, the teams that are improving across every accessibility metric are not the ones using the most AI, they are the ones with mature, embedded accessibility practices that AI can amplify.
The core insight from Andrew Chung's analysis is worth sitting with: AI adoption made the smallest difference among the factors that separate high-performing teams from the rest. What matters is training, communication, and a culture where accessibility is treated as part of the creation process, not a final checklist. This aligns with the governance gaps we explored in When AI Agents Act Autonomously, Five Governance Gaps Emerge, in both cases, the challenge is not the technology itself but the human systems that surround it. If you have not already embedded accessibility checks at the handoff, the pull request, and the component level, AI will simply help you ship inaccessible work faster.
The practical takeaway is specific and actionable. Trace a recently shipped feature from design to production and mark every point where a person decided whether it would work for a keyboard user or a screen reader. If most of those marks cluster at the end, you are building a backlog of fixes that will cost more to address later. The teams that are keeping pace have moved those judgment calls earlier: designers specify focus order and accessible names at the handoff, developers run keyboard and screen reader tests on custom components, and automated gates in CI/CD catch regressions before merge. This does not require new tools for most teams, it requires deciding who owns each check and making that ownership explicit.
One detail from the report stands out as a warning for anyone using AI coding assistants. The data shows that professionals in mature accessibility programs were nearly six times more likely to say AI helps them prioritize issues, and nearly three times as likely to say it helps them tackle accessibility proactively. That is not because they have better AI, it is because they have the context, guardrails, and testing environments in place for AI to be useful. Without those, an AI assistant optimizes for passing a check, not for delivering a correct experience. The same principle applies whether you are training a local model on your editing style, as described in Why settle for presets when your own local AI can learn your editing style, or deploying agents across a product workflow. The question to answer this week is straightforward: where does human judgment sit in your accessibility workflow today? If you cannot trace it, you have already lost ground.
