“Two hundred errors. One colour value. And nobody to fix them.”
Almost every company we talk to already has an accessibility report. Eighty pages, neatly sorted, colour-coded. It sits in a folder.
Nobody works through it. Not out of negligence — but because every single line in it is a small research project. Where does this value come from? Which pages does it affect? What breaks if I change it? At two hundred lines that isn't an afternoon. It's a project nobody has a budget for.
sp8 reverses the order. The assistant doesn't hand you the report. It hands you the change.
Accessibility checkers have existed for twenty years, and they are good. They reliably find what's wrong. What they can't do is the second step — and the second step is all of the work.
A contrast failure doesn't live in a page. It lives in a colour value that applies to every page. A checker crawling the published site therefore reports it two hundred times — without knowing it is the same failure two hundred times. Work through that list without knowing the cause and you do the work two hundred times.
The WCAG expert in sp8 doesn't inspect the finished page. It inspects the system behind it. So it doesn't see two hundred findings. It sees one design token — and names it.
A contrast ratio of 3:1 instead of the required 4.5:1 doesn't mean a page looks bad. It means someone with failing eyesight can't read the price.
A skipped heading level doesn't mean the markup is untidy. It means someone having the page read aloud can't make out its structure and has to start over.
That's why the check returns more than an error message: the elements checked, the WCAG criterion violated, the severity as A, AA or AAA — and for contrast failures the design token responsible, for structural failures the segment and field. Not 'contrast too low'. Which value, in which place, against which criterion.
The second capability fixes things. It has write access to the instance's design tokens and to segment content. So it can change the contrast value, not annotate it.
This is the point where you wouldn't normally trust an AI. Rightly so.
Anyone who works with AI knows the moment. It says 'done' — and nothing happened. It meant well, it described the operation, it never carried it out. And it doesn't even know.
In sp8 that isn't a matter of trust, it's a field. Every capability has postconditions, and the WCAG fix capability states it literally: only report success once the WCAG issues are resolved after the change — a fresh check is required.
So the assistant can't declare itself finished. It has to go and look. That's a single sentence in a configuration field — and it's the reason you can believe the result.
An assistant judging accessibility must not quote the criteria from memory. Language models are remarkably good at that and occasionally wrong — and with a standard, occasionally wrong is the same as useless.
The WCAG guidelines and techniques therefore sit in the system as their own searchable knowledge base. The capability responsible has exactly two tools: list knowledge bases, search a knowledge base. Nothing else.
And its answers have to name the base used and cite the criteria with links to the official documents. Every assessment can be checked — including by someone who doesn't take our word for it.
There are tasks where a machine is no match for a person. Accessibility isn't one of them.
It consists of hundreds of near-identical, rule-based cases that have to be worked through doggedly and with undiminished attention. People are bad at that — not because they couldn't do it, but because nobody brings the same care to the two-hundredth contrast value as to the first. That is exactly where a machine is better.
Which is why accessibility in sp8 isn't a campaign you run once and tick off. It's a state somebody maintains.
The effort shifts. Instead of 'someone spends weeks working through a list' it becomes 'someone reviews the proposed fixes and approves them'.
And because the check runs against the actual state of the system, the findings don't go stale between report and implementation. That is the classic reason the first audit was money down the drain.
The WCAG expert works on request today. Having it check on its own initiative and on a schedule is configurable and not yet used in production.
And it judges what sits in the system — not how a screen reader actually reads the finished page out loud. It does not replace testing with assistive technology by real people. Anyone promising that hasn't understood the job.
We'll show you live how the WCAG expert in sp8 finds a design token, fixes it and re-checks the result.