Automated Accessibility Tagging
Teaching a million old PDFs to speak to screen readers — semantics applied at a scale no manual team could reach.
Automated accessibility tagging applies document AI to the remediation problem: taking documents that lack semantic structure — scanned pages, untagged PDFs exported from design tools, decades of accumulated archives — and adding the tags that assistive technologies require: heading levels, paragraph and list structure, table semantics with header associations, reading order, figure alternative text, and form field labels. Manual tagging by a specialist takes minutes to hours per document; organizations sitting on repositories of millions of documents have no manual path to compliance at all.
The AI stack maps naturally onto the tagging task because it is the same problem as document understanding. Layout analysis identifies the structural regions; reading-order models sequence them; table-structure recognition reconstructs row, column, and header relationships; classification distinguishes decorative images from content-bearing figures; and vision-language models generate draft alternative text describing charts and photographs. The output is written back as the document's tag tree, transforming a flat visual artifact into a navigable structure a screen reader can traverse by headings, skim by list, and read tables from intelligently.
Quality assurance is where automated tagging programs succeed or fail, because accessibility standards include criteria machines can't fully judge — whether alt text is meaningful, whether the reading order makes sense for this content, whether a table's structure reflects its logic. Mature pipelines pair automation with targeted human review: automated checkers validate structural conformance, sampling or confidence-based routing sends uncertain documents to remediation specialists, and the highest-traffic or legally-critical documents get full manual verification. The result is a tiered program that reaches the whole repository while concentrating human expertise where judgment is genuinely required.
PDF/UA and WCAG walk into an audit — making the world's default document format usable by everyone, provably.
A document isn't really published until everyone can read it — including people using screen readers.
Before reading the words, a model has to see the page the way a human does — headings here, table there, footnote at the bottom.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo