Automated Accessibility Coverage
The proportion of accessibility criteria a tool can decide mechanically — around a third — and the explicit acknowledgement that the rest needs people.
Automated accessibility checking catches a specific and useful class: missing alternative text, insufficient colour contrast, absent form labels, invalid ARIA, missing document language, unlabelled controls. Run in CI on every page and component, it prevents regression cheaply and it should be there.
The number worth internalising is that this is roughly a third of the criteria. The remainder are judgements a machine cannot make: whether the alternative text is meaningful, whether the focus order matches the visual and logical order, whether an error message explains how to fix the problem, whether a custom component is operable by keyboard alone, whether a screen reader announces a dynamic update at a useful moment.
The risk of the automated layer is therefore false assurance — a green build reported as "accessible" when two thirds of the standard has not been assessed. The honest arrangement pairs continuous automated checks with periodic manual testing using an actual screen reader and keyboard-only navigation, ideally including people who use those tools daily, and states the split rather than letting the dashboard imply completeness.