It is tempting to paste your homepage into ChatGPT, ask it to check for accessibility problems, and call the resulting list your WCAG audit. It takes five minutes. It feels thorough. It even produces a tidy bulleted list that looks like real work.
It skips the exact step that actually finds the issues that get companies sued.
The demand letters and lawsuits that keep landing under the ADA, Section 508, and international accessibility laws are rarely about the things a model can catch from a single screenshot, like a missing alt attribute on a logo. They turn on things no AI can verify without actually operating the site: whether a custom modal traps and returns focus correctly, whether a screen reader user can complete checkout without getting stuck, whether the reading order of a data table still makes sense once you turn on a screen reader and try to use it for real. Used well, AI is a legitimate and valuable part of a modern audit process. Used as the entire process, it is a liability wearing a nice user interface.
The Stat: Deque's own published research finds automated scanning alone catches roughly 57.38% of WCAG issues on a page. (Source: Deque Systems)
What AI Actually Catches (And What "57%" Really Means)
That Deque figure is worth sitting with. It means that if your entire audit process is "run a scanner and fix what it flags," you are, on average, addressing a little over half of the actual conformance failures on the page and calling it done. Automated tools are excellent at what they were built for: parsing the DOM and flagging objective, rule-based violations. Missing form labels, insufficient color contrast ratios, empty headings, missing document language attributes, images without alt attributes at all. Tools like axe-core are the right starting point for every audit, and there is no good reason not to run one across every unique template on your site before doing anything else.
What automated tools structurally cannot do is judge meaning or behavior. A scanner can confirm an image has an alt attribute. It cannot confirm the alt text actually describes the image, or that it is redundant with adjacent link text, or that it communicates the same information a sighted user would get from the image in context. A scanner can confirm a custom widget has role="dialog". It cannot confirm that focus is trapped inside it correctly, that Escape closes it, or that focus returns to the triggering element afterward. Those require someone to actually operate the interface with a keyboard and a screen reader. That gap between what a scanner reports and what conformance actually requires is precisely where AI-assisted triage, and then human verification, earn their place in the process.
The 2026 Audit Order of Operations
A reliable audit in 2026 runs in three stages, in this order. Skipping a stage, or running them out of order, is how teams end up with a report that looks thorough but misses the findings that matter most.
Stage 1: Automated scan, always first
Run an automated tool such as axe-core or Lighthouse across every unique page template, not just the homepage and a couple of landing pages. Most sites are built from a small set of reusable templates (product page, checkout flow, search results, blog post), and a single unresolved issue in a shared component can multiply across thousands of URLs. The W3C's guidance on testing and evaluating for accessibility is a good baseline for deciding how broad this first pass should be and which tools to combine for better DOM coverage.
Stage 2: AI-assisted triage and drafting
Once you have raw scan output, an LLM becomes genuinely useful. This is the stage where AI earns its keep: deduplicating hundreds of near-identical findings down to a handful of root causes, drafting plain-language descriptions of each issue for non-technical stakeholders, prioritizing by frequency and severity, and mapping each finding to the specific WCAG success criterion it violates so nothing gets logged against the wrong requirement. Every mapped finding should still be checked against the actual WCAG 2.2 conformance requirements, since an AI-drafted mapping is a first draft, not a citation you can hand to legal counsel unverified.
If you want the specific prompts we use for this stage, day to day, we published the exact working set in 11 Prompts That Turn ChatGPT Into an A11y Assistant. That article is the task-by-task toolkit; this checklist is the process those tasks fit inside.
Stage 3: Human verification and sign-off
Nothing leaves an audit as a confirmed finding until a person has verified it by hand. That means a full keyboard-only pass through every critical flow (navigation, forms, checkout, any custom widget), a screen reader pass with something like NVDA, JAWS, or VoiceOver, and a manual check that AI-drafted alt text and ARIA labels are actually accurate rather than just plausible-sounding. This is also the stage where you confirm focus order and focus trapping on modals, tooltips, and menus, none of which a static scan can verify.
It is worth saying plainly: this stage cannot be replaced by a bolt-on widget that promises automatic remediation. We ran that experiment ourselves and documented exactly what happened in AI Overlays: We Tested 5, All Failed WCAG. An overlay is not a substitute for a human verifying real conformance, any more than a scanner is.
What AI Is Good At, and What Still Needs a Human
- Good for AI: parsing DOM-level rule violations at scale, deduplicating and clustering findings, drafting plain-language issue summaries, suggesting a first-pass code fix for engineers to review, mapping findings to likely success criteria.
- Still needs a human: judging whether alt text and ARIA labels are accurate and meaningful, testing keyboard and screen reader behavior on custom widgets, verifying focus management and reading order, confirming legal conformance against the actual WCAG success criterion text, and signing off that a fix genuinely resolves the underlying issue rather than just satisfying the scanner.
Treat the AI-drafted output from Stage 2 the way you would treat a junior team member's first pass: fast, useful, and never final without review.
Bringing It Together
The teams that stay out of trouble in 2026 are not the ones with the fanciest AI tooling. They are the ones who treat automated scanning, AI-assisted triage, and human verification as three distinct, sequential stages, each catching what the previous one structurally cannot. Skip Stage 3 and you are shipping a report built on the roughly 57% a scanner can see. Skip Stage 2 and your engineers drown in duplicate, unprioritized findings. Skip Stage 1 and you are paying human reviewers to find issues a scanner would have caught in minutes.
If you would rather skip the setup and get a professional read on exactly where your site stands today, our team runs this same three-stage process as a free accessibility audit, and we will show you precisely which findings came from automation, from AI triage, and from a human set of eyes.
Have questions about where your site fits into this process, or want a second opinion on an audit you already ran. Reach a real person on our team at experts@wcag.world, or start with a free WCAG audit and we will walk you through the results stage by stage.