What our AI does, and who checks it.
Our AI does three jobs: it finds a page’s main task, explains findings in plain words, and drafts code fixes.
It doesn’t decide whether a page meets the standard or the law. Tools test only part of the standard, and people judge the rest.
The jobs our software does, and what checks each one.
We say “our AI” only when an AI model does the job. Our experts do the rest of the work.
Managed accessibility is new, and not everything is ready yet. The last column says what’s live today.
| Job | What does it | What checks it | Is it live? |
|---|---|---|---|
| Find a page’s main task and open it, in the free scan | Our AI agent, in its own browser | It makes no accessibility judgment. The automated tests do that. | Live in the free scan |
| Explain each finding in plain words, in the free scan’s report | Our AI. When it isn’t available, the report uses our standard wording. | Software checks the wording for claims we do not make. No person reads it. | Live in the free scan |
| Draft a code fix, or advice when a fix needs a decision, in managed accessibility | Our AI | Software checks that the draft fits one place in your code and avoids claims we do not make. In managed accessibility, an expert then approves, edits, or rejects every fix. | Built, not live yet |
| Detect barriers on a page, in the free scan and in managed accessibility | Automated tests from axe-core, an open-source rules engine. This is not AI. | Nothing reviews the result in the free scan. In managed accessibility, a person reviews each result. | Live in the free scan. Built for managed accessibility, not live yet. |
| Run a website’s main tasks each day, in managed accessibility | Scripted tasks, which find each control by its role and name. This is not AI. | A person reviews each result. | Built, not live yet |
Automated tests cover only part of the standard, so a page with no detected barriers can still stop people. A script doesn’t use a screen reader or a keyboard, and a clean run doesn’t mean a task works for everyone.
An example scan of our practice page.
We built a practice page with barriers on purpose and ran our scan on it. These views show the saved output of that run.
It’s one page on one date. It shows how a check reports, not what one would find on your website.
Scan your own page for freeWhat the automated tests found
The report lists each detected barrier in plain words and says how serious it is.
Tools miss what only people can judge, so a barrier that isn’t on this list can still stop someone.
- CriticalThe arrow button in the form shows only an icon, with no words, so nothing on the page says what pressing it does.
- CriticalThe dining room picture has no text description, so nothing explains what the image shows.
- CriticalThe name field in the form has no label, so a visitor can’t tell what to type in it.
- SeriousThe small print about the 15-minute table hold is light grey on white, which makes it hard to read.
- ModerateThe page turns off pinch zoom on phones, so visitors can’t enlarge the text.
Our AI writes this wording when it’s available. No person reads it.
What our AI drafted
Where a code change can remove a barrier, our AI writes the change and explains it. A free scan doesn’t draft fixes, so we ran this step separately on one barrier.
Some barriers need a design or content decision, and we tell you which ones. In managed accessibility, an expert approves, edits, or rejects every fix.
practice/index.htmlDraft fixRemoved: <button type="submit"><span class="icon-arrow"></span></button>Added: <button type="submit" aria-label="Book a table"><span class="icon-arrow" aria-hidden="true"></span></button>This submit button shows only an arrow from a CSS background image, so it has no accessible name. People who use a screen reader or voice control hear only "button" and can't tell what it does. The change adds aria-label="Book a table", which matches the form's heading, and hides the empty icon span with aria-hidden="true". The button now announces its purpose and can be targeted by name with voice commands.
In managed accessibility, an expert approves each fix. No expert reviewed this example.
What people do, and what happens when our AI is wrong.
Our AI can get things wrong. In managed accessibility, a person reviews each fix before it reaches your code.
See what our experts doApprove fixes
In managed accessibility, an expert approves, edits, or rejects every fix before it reaches you. A rejected draft doesn’t reach your code.
Test with a keyboard, a screen reader, and magnification
In managed accessibility, our experts test your main tasks this way on an agreed schedule.
They test only the tasks and pages in scope. The free scan doesn’t include this testing.
Make the calls software can’t
People decide whether a description fits its image, whether a task makes sense, and what to do when a fix needs a design or content decision. Our AI gives advice instead of a fix when the choice is yours to make.
If you find a mistake
Tell us what our AI got wrong and where you saw it.
Tell us about a mistake