Accessibility deadlines in Europe and the United States have created a flood of untagged PDFs that suddenly need fixing, and a flood of AI tools promising to fix them in one click. Adobe, specialist remediation vendors and newer startups all use machine learning to detect headings, tables, lists and reading order. So, does it work?
The honest answer: it is a big time saver and not a finish line.
What AI auto-tagging does well
- Detecting text blocks and paragraphs on clean, single-column pages.
- Identifying headings based on size, weight and position, especially in consistently formatted documents.
- Recognising simple tables with clear gridlines.
- Proposing alt text for images with multimodal models, which is far better than the blank alt text most PDFs ship with.
- Speed. Tagging a 100-page report from scratch by hand takes hours. AI gets you a draft structure in minutes.
Where it still fails
- Complex tables: merged cells, multi-level headers and tables split across pages.
- Reading order in magazine-style layouts with sidebars, pull quotes and captions.
- Heading levels: a tool may tag every bold line as a heading, or flatten the hierarchy.
- Meaningful alt text: a model can describe a chart's appearance ("a bar chart with blue bars") without capturing its point ("sales doubled in Q3").
- Decorative versus informative: logos, borders and background shapes are often tagged as content instead of artifacts.
- Forms: field labels, tooltips and tab order still need human attention.
A workflow that holds up
- Start from the source if you can. A properly styled Word or InDesign file exported with tags beats any remediation. See tagged vs untagged PDFs.
- Run AI auto-tagging on legacy PDFs with no source.
- Validate with an automated checker such as PAC to catch structural errors.
- Review by hand: headings, table headers, reading order and every alt text.
- Test with a screen reader for important documents.
- Fix the template that produced the document, so you do not repeat the work.
Beware of "accessibility overlay" promises
Some vendors claim automatic, guaranteed compliance. Regulators and disability advocates have criticised overlay-style products on the web, and the same scepticism applies to documents. An auditor checks the result, not the tool that produced it. Automated checkers can only test about a third to a half of the relevant criteria.
When is AI good enough?
For internal archives where some access is better than none, a reviewed AI draft is a reasonable outcome. For customer-facing documents under the European Accessibility Act or the ADA, use AI to get 80% of the way, then finish by hand.
Takeaway
AI auto-tagging turns accessibility remediation from a blank page into an editing job. Treat it like a first draft from a capable junior: useful, fast, and in need of review. For the full set of requirements, see the European Accessibility Act checklist and the PDF accessibility guide.
