AI found 47 accessibility issues.
Amazing? Maybe.
It might be useful or it might be a bunch of noise.
I don't think the right question is how many issues AI managed to find. The number alone is meaningless. How do you know if, as a result of what it found, you managed to meaningfully reduce real barriers your users face? Ah, now that's a better question.
A good accessibility audit, whether you use AI for it or not, should do five things well.
First, it should look in the right places. I'm talking keyboard navigation, form labels, focus order, colour contrast, accessible names, error messages and screen readers.
Second, the findings need to be accurate. The last thing your busy team needs is false positives to waste their time on. False negatives are even worse because they create false confidence.
Third, it should rank issues by user impact. A missing button name that blocks checkout matters more than a minor heading-level nitpick.
Fourth, it should give guidance your team can use. Something like "fix ARIA" is ridiculous. This is where AI could maybe help a lot. Could it show the broken experience, explain why it matters and suggest a fix?
And fifth, it should be checked against real use. Automated results, even AI-generated ones, are not the same as a person with disability completing that task with assistive technology.
All this takes time and resources. AI could speed that up, and, if done well, be even useful.
But the metric is not the volume of issues it can find.
Can your team can make better release decisions thanks to it? That's where the value is.