Automated PDF remediation handles the mechanical bulk of accessibility work. Software can tag structure (auto-tagging is where a reviewer looks first), set the document language, fix tab order, name links, build bookmarks and repair fonts, then re-check each change. Software cannot settle questions of meaning on its own. Whether alt text says the right thing, whether content reads in the order the author intended, and whether a color change is acceptable are judgment calls, so a person makes them. AI helps by drafting answers for that person to confirm. The table below maps each WCAG 2.1 AA check for PDFs to who resolves it and why.
How much software fixes
Software clears most mechanical defects. In Ockem's measurements across 12,689 findings in three document libraries in August 2026, automatic passes cleared 63–74% of defects. Counted by document across production documents in the same month, the median document still needs one manual check, and the 90th percentile needs two. The manual check we most often see left over is logical reading order, because deciding the correct order takes judgment.
Three ways a check gets resolved
- Software. A deterministic fix is applied, then re-checked by the rule engine. Nobody needs to decide anything.
- Software, then a person confirms. Software detects the problem and drafts or proposes a fix. A person accepts, edits or rejects it before it counts.
- A person. Software finds candidates and shows the evidence. A person decides, and the decision is recorded.
Check by check: who resolves what
This table comes from the WCAG coverage map for Ockem Proof, our product, which runs a rule engine built on veraPDF plus our own detectors. Other tools may draw some of these lines differently. Linked checks have a deep dive.
| Check | WCAG criterion | Who resolves it | Why |
|---|---|---|---|
| Document title | 2.4.2 Page Titled (A) | Software, then a person confirms | Spotting a missing title, or one that repeats the filename, is mechanical. Choosing a good title is a naming decision, so AI suggests one for a person to review. |
| Document language | 3.1.1 Language of Page (A) | Software | The language can be detected from the text and written into the file. Non-English detections are worth a glance. |
| Language of parts | 3.1.2 Language of Parts (AA) | A person | Software can find passages in a different script. Which passages are in another language, and which language, is a content decision. |
| Headings and lists | 1.3.1 Info and Relationships (A) | Software, then a person confirms | Rebuilding lists and closing up skipped heading levels is mechanical. PDF/UA requires unbroken heading levels, while WCAG treats them as best practice. Whether a large bold line is a heading or decoration is a judgment about meaning. |
| Tables | 1.3.1 Info and Relationships (A) | Software, then a person confirms | Common export damage can be repaired automatically. Ambiguous tables, multi-row headers and header columns need a person to say which cells are headers. |
| Reading order | 1.3.2 Meaningful Sequence (A) | Software, then a person confirms | Software can build structure and flag pages where the order looks wrong. Whether the order matches what the author meant takes a person. |
| Alt text | 1.1.1 Non-text Content (A) | Software, then a person confirms | Software can find every image and draft a description. Whether an image is meaningful or decorative, and what it must convey, takes a person. |
| Scanned pages | 1.1.1 Non-text Content (A) | Software | OCR can rebuild a scanned page with real text and structure. Heading levels recovered from a scan are estimates, so flagged pages deserve a look. |
| Color contrast | 1.4.3 Contrast (Minimum) (AA) | Software, then a person confirms | Measuring contrast is mechanical. Changing colors alters the design, and text over photos needs a design choice, so a person approves. |
| Images of text | 1.4.5 Images of Text (AA) | A person | Software can detect text that exists only inside an image. Whether that use is acceptable, as in a logo, is a judgment. |
| Non-text contrast | 1.4.11 Non-text Contrast (AA) | A person | Software can measure form-field borders. Deciding which graphics a reader needs to understand the page is a judgment. |
| Sensory characteristics | 1.3.3 Sensory Characteristics (A) | A person | Software can flag phrases such as "the box on the right." Whether the instruction still works without seeing the page is a judgment about meaning. |
| Resize text | 1.4.4 Resize Text (AA) | A person | How text behaves when a reader zooms or reflows it is reviewed and attested by a person. |
| Form fields | 4.1.2 Name, Role, Value (A) | Software, then a person confirms | Tagging fields is mechanical. What a field is called is content, so software drafts a name from the printed label for a person to accept. |
| Links | 2.4.4 Link Purpose (A) | Software | Links can be tagged and named from their visible text, and bare web addresses turned into real links. |
| Bookmarks | 2.4.1 Bypass Blocks (A), 2.4.5 Multiple Ways (AA) | Software | Bookmarks can be generated from the heading structure of a long document. |
| Tab order | 2.4.3 Focus Order (A) | Software | Tab order can be set to follow the document's structure. |
| Fonts and character encoding | Supports 1.1.1 and 1.4.5 | Software | Embedding fonts and repairing character maps is mechanical, and the page looks the same afterward. |
| Accessibility permission flag | 4.1.2 Name, Role, Value (A) | Software | A security setting that blocks assistive technology can be detected and switched off. |
| Embedded audio and video | 1.2.1 Audio-only and Video-only (A) | Software, then a person confirms | Media can be identified and given a placeholder description for a person to word. Captions and transcripts are authoring work. |
| Keyboard, timing, flicker and focus visible | 2.1.1, 2.1.2, 2.2.1, 2.3.1, 2.4.7 | A person | These mostly matter in PDFs with scripts or interactive forms. Software flags candidates, and a person tests and attests. |
Why some checks need judgment
Alt text quality
WCAG 1.1.1 asks for a text alternative that "serves the equivalent purpose." Purpose depends on context. The same bar chart might need a one-line takeaway in a press release and the full figures in a budget. A photo of a building might be decorative on a cover and essential in a zoning notice. Software can describe what an image shows. A person knows why it is there. Drafts help a reviewer move quickly, and a quality check can reject empty answers such as "image."
Reading order and meaningful sequence
WCAG 1.3.2 says that "when the sequence in which content is presented affects its meaning, a correct reading sequence can be programmatically determined." Two parts of that sentence need judgment, whether the sequence affects meaning and what the correct sequence is. In a two-column newsletter with a sidebar and a pull quote, software can compare the visual layout with the tag order and flag pages where they disagree. Whether the sidebar belongs before or after the main story is the author's call.
Sensory characteristics
WCAG 1.3.3 says instructions must not "rely solely on sensory characteristics of components such as shape, color, size, visual location, orientation, or sound." Software can find phrases such as "the round button" or "see the chart at right." It cannot tell whether the surrounding text already gives a way to follow the instruction without seeing it. That takes a reader.
Contrast over images
WCAG 1.4.3 requires a contrast ratio of at least 4.5:1 for body text and 3:1 for large text. Measuring is mechanical for text on a solid background. Over a photograph, the background changes from pixel to pixel, and text that is part of a picture with significant other visual content has no contrast requirement. Deciding which case applies, and whether a fix such as darker text or a solid backing band is acceptable, is a design decision. Teams that rely on Acrobat should know that Acrobat's checker does not test color contrast. It always lists contrast as a manual check. This comparison was written by Ockem.
What AI does in PDF remediation
In its April 2026 interim final rule, DOJ wrote that "advanced technology, such as generative AI, does not yet reliably automate the remediation of inaccessible content at scale." The table above reflects the same reality. AI is useful for drafting the parts that need language.
- Alt text and longer descriptions for images.
- Document titles, when the title is missing or repeats the filename.
- Form field names, where the page prints no label.
- Table summaries.
In each case a person approves the draft before it is written into the file. In Ockem Proof, these drafts run on hardware Ockem operates, and no customer document content goes to a third-party model.
When you evaluate any tool, ask which checks a person confirms, who that person is, and how the decision is recorded.
How to plan the human part
- Expect a short list per document. Plan for about one decision on a typical document and two on harder ones.
- Batch similar decisions. Contrast failures across a document often trace back to a few template colors, so one decision can clear many of them.
- Fix the source. Templates with real heading styles, alt text and accessible colors mean fewer decisions on every new document.
- Record who decided. A decision with a date and an accountable party behind it is evidence. Ockem's certificate for each document states what software verified, what your organization attested, and what is not claimed.
- Listen to a sample. A certificate is evidence of conformance checks and attestations, not a usability test. For high-stakes documents, such as forms people must complete and notices people must act on, have someone test a sample of finished files with a screen reader such as NVDA, JAWS or VoiceOver.
- Decide who does it. Your staff can make the decisions, or an outside service can. Ockem's Assured option has our experts do the human-review steps for a quoted price, as shown on the pricing page.
See it on your own files
You can try one PDF (opens an external site) of up to 10 pages free, with no account, or read the product overview. To count your PDFs first, use the free PDF inventory tool. For the rules behind these checks, read Title II and PDFs and archive, keep or fix. For what remediation costs, see the cost guide.
Questions
Can AI make PDFs accessible?
AI can draft parts of the work, such as alt text and titles, and software can fix most mechanical problems. Judgment calls such as reading order and the meaning of images still need a person, and DOJ wrote in April 2026 that generative AI "does not yet reliably automate the remediation of inaccessible content at scale."
Which PDF accessibility checks are manual?
Checks that depend on meaning need a person, including alt text quality, reading order, sensory characteristics, images of text, language of parts, resize text and focus visible. Software measures color contrast, but color changes need a person's approval.
Can software alone make a PDF meet WCAG 2.1 AA?
Some documents need no human decisions, but many need at least one. In Ockem's measurements of production documents in August 2026, the median document needs one manual check, and the 90th percentile needs two.
Does Acrobat check color contrast?
Acrobat's checker does not test color contrast. It lists contrast as a manual check, so you need another way to measure text against the 4.5:1 ratio WCAG 2.1 sets for body text. This comparison was written by Ockem.
This guide explains the rules in plain language. It isn’t legal advice. Check how they apply to your organization with counsel.
Sources
- DOJ interim final rule extending Title II compliance dates, full text (April 2026) (opens an external site)
- DOJ interim final rule (Federal Register 2026-07663, April 20, 2026) (opens an external site)
- W3C, Web Content Accessibility Guidelines (WCAG) 2.1 (opens an external site)
- W3C, Understanding Success Criterion 1.3.2: Meaningful Sequence (opens an external site)
- W3C, Understanding Success Criterion 1.3.3: Sensory Characteristics (opens an external site)
- W3C, Understanding Success Criterion 1.4.3: Contrast (Minimum) (opens an external site)
- W3C, PDF techniques for WCAG 2.1 (opens an external site)
- veraPDF, open-source PDF validator (opens an external site)
- Clemson University, manual checks for PDF accessibility (Acrobat contrast check) (opens an external site)