PDF Accessibility

Manual vs Automatic PDF Remediation: Which Work Belongs in Each Process?

Learn which PDF accessibility tasks software can accelerate and which require human judgment, from tags and text layers to reading order, image meaning, equations, and complex tables.

People in the office

Updated November 18, 2026.

The practical answer: automate structure, reserve judgment for people

For manual vs automatic PDF remediation, use this rule: automate repetitive fixes when the intended structure is clear; use a person for meaning, context, complex layouts, and final review. Software can accelerate tagging, scan text layers, simple structure repairs, and initial checks, but it cannot decide what an image means, whether a complicated table is understandable, or whether a screen reader gets the document in the right sequence.

Manual remediation means a person inspects and corrects a PDF. Automatic remediation means software applies or suggests structural repairs. The strongest process uses both: automation to get through the mechanical work, human review to make sure the result matches the document.

A tag is metadata inside the PDF that identifies what an item is: a heading, paragraph, list, table, link, figure, form field, or decorative artifact. Tags are not decoration. They are the structure screen readers and other assistive technologies use to navigate and announce the file.

The mistake is treating a repaired file or a cleaner checker result as proof. It is evidence for triage and review, not a certification that the PDF is accessible, compliant, or conformant with PDF/UA or WCAG.

What automatic PDF remediation can usually accelerate

Automatic remediation is best at work that is repetitive, detectable, and structurally clear. That usually means the software can infer the intended structure from the page layout, existing metadata, font patterns, link annotations, form fields, and the relationship between nearby objects.

This is where automation earns its place in PDF remediation: it reduces the time spent doing the same low-judgment corrections page after page.

Good candidates include:

  • Basic tag identification and repair. Headings, paragraphs, lists, links, figures, form fields, and tables often follow visual patterns that software can detect. A title at the top of a page in a larger font is a likely heading. Bulleted lines are likely a list. A visible URL or annotated link may need a link tag.

  • Marking decorative content as artifacts. An artifact is content that should be ignored by assistive technology, such as a border, background shape, repeated divider, or purely decorative icon. Software can often flag obvious decoration so a screen reader does not announce it as meaningful content.

  • Link and form-field structure. Where link annotations and form fields already exist, automation can help connect them to the tag tree and surface missing labels or unclear field names for review.

  • Scanned-page text layers. A scanned PDF is often just an image of text. Optical character recognition, or OCR, can create a text layer so the words can be selected, searched, and read by assistive technology.

  • Straightforward tables. A simple table with clear rows, columns, and header cells may be a reasonable automation candidate when the relationships are unambiguous.

OCR deserves special caution. A text layer is not the same as a usable document. OCR can misread characters, split words, flatten columns into nonsense, or miss handwritten and low-resolution content. The output still needs review for recognition errors, reading order, and meaning.

Simple tables are similar. If a table has one header row and regular columns, software may be able to apply a useful first structure. If it has merged cells, nested headers, blank spacer rows, footnotes, rotated text, or multiple header levels, it belongs in human review.

The useful distinction is candidate repair versus confirmed repair. Automation can quickly apply a likely structure. A person has to check whether that structure reflects the actual content.

Penn State Libraries puts the tradeoff plainly in its PDF remediation tool guidance: automated remediation “typically achieves approximately 85–90% accessibility,” depending on document complexity and structure, and still requires manual review for the rest (Penn State Libraries). The exact percentage is less important than the principle: the last gap is not just more clicking. It is judgment.

What needs human remediation and review

Human remediation is for the parts of a PDF where the correct answer depends on understanding the document.

The biggest one is logical reading order: the sequence in which content should be read by assistive technology. A screen reader does not experience a page as a sighted reader does. It follows the PDF’s structure. If the structure says to read a sidebar between two halves of a sentence, or to announce a footer before the main paragraph, the document may have plausible tags and still be confusing.

This is why multi-column layouts are hard. A visual reader can jump from the top of the left column to the bottom, then across to the right column. A screen reader needs that sequence encoded. Sidebars, footnotes, pull quotes, captions, callouts, marginal notes, and overlapping visual elements often require a person to decide what belongs in the main flow and what should be read later.

For a detailed treatment of this specific problem, see Ordecto’s guide to logical reading order.

The second major category is alt text. Alt text is a written description of an image’s purpose or meaning for someone who cannot see it. It is not always a literal caption. The right description depends on why the image is in the document.

A chart in a financial report may need the trend and key values. The same chart in a design mockup may need its visual treatment described. A photograph in a campus brochure might be decorative; a photograph in a medical, legal, or research document may carry substantive information. Software can detect that a figure exists and may suggest a description. A person must decide what matters in context.

Other human-review tasks include:

  • Complex tables. A person needs to verify header relationships, reading sequence, footnotes, blank cells, row groups, and whether the table should be simplified or explained in surrounding text.

  • Equations and formulas. Equations may need both a correct visual presentation and an accessible representation, such as MathML or spoken text. MathML is a markup language for representing mathematical notation in a structured way.

  • Ambiguous headings and lists. A bold line may be a heading, a label, a table cell, or just emphasis. A sequence of indented lines may be a list, an outline, or a set of examples.

  • Context-dependent structure. Forms, appendices, glossaries, legal clauses, references, and academic papers often contain repeated patterns that look similar but function differently.

  • Assistive-technology sense-checking. The real question is not whether every object has a tag. It is whether the file can be navigated, understood, and used.

The non-obvious failure mode is a PDF that looks remediated because every visible element has a technically plausible tag. The screen reader still reads the abstract after the references, a table caption after the table body, or a footnote in the middle of a sentence. Nothing about that is meaningfully accessible to the reader who has to rely on the structure.

That is why what automated checkers cannot test reliably matters as much as what they can detect.

Why an automated checker result is not a final accessibility decision

PDF/UA is the accessibility standard for PDF documents. WCAG, the Web Content Accessibility Guidelines, is a set of accessibility requirements used for web and digital content. Both are useful frameworks for reviewing PDFs, but neither turns an automated result into certification.

A checker can identify many failed conditions. It can tell you about missing tags, empty headings, untagged annotations, missing document language, form-field problems, suspicious table structure, or figures without descriptions. It can also surface findings that need a person and areas it did not test.

That is not the same as deciding that the PDF is accessible.

Automated tools struggle with the parts that require interpretation:

  • whether alt text communicates the image’s actual purpose;

  • whether a complex table’s header relationships are understandable;

  • whether the reading order matches the intended reading experience;

  • whether an equation’s accessible representation is correct;

  • whether a form’s instructions make sense before the user reaches the field;

  • whether the document works for its likely audience.

Ohio State University’s accessibility guidance describes automated checkers as giving a “quick overview” of accessibility issues found in PDFs, not as a substitute for ensuring accessibility through review (Ohio State University). That is the right mental model.

Treat checker output as a prioritised worklist:

  • Machine-detectable failures. These are the items software can identify with enough confidence to send straight into repair or review.

  • Findings needing a person. These are the ambiguous items where the software can point to the issue but cannot decide the answer.

  • Not-tested areas. These are outside the tool’s reliable scope and need another review method.

A PDF/UA checker is useful because it gives structure to the review. It does not certify the file. For the deeper standards distinction, Ordecto’s explanation of PDF/UA vs WCAG for PDFs is the better next stop.

A repeatable manual-and-automatic workflow for many PDFs

The hard problem is not one PDF. It is a department, agency, publisher, school, or research group with hundreds or thousands of PDFs and a deadline. The answer is not to pretend every file needs the same level of work. It is to build a triage process that reserves expert time for the files and issues where it changes the outcome.

Step 1: inventory the collection and prioritise files

Start by finding the PDFs and sorting them. Do not treat every archived file as equally urgent.

Prioritise by:

  • Audience impact. Public forms, course materials, patient information, legal notices, benefits documents, and high-traffic reports usually outrank low-use archives.

  • Publication date. Current and recently updated documents often matter more than old materials kept only for reference.

  • Use frequency. Files linked from main pages, portals, syllabi, application flows, and public guidance pages should move up the queue.

  • Risk. Documents tied to legal rights, government services, healthcare, education, finance, employment, or public participation deserve earlier review.

  • Complexity. A scanned 200-page manual with tables and forms needs different planning from a 3-page text report.

  • Deadline. Procurement, course launches, public-comment windows, regulatory dates, and grant deadlines change the order of work.

The University of Georgia’s guidance on document accessibility prioritisation starts in the same place: inventory the PDFs and make the work manageable rather than treating the backlog as one undifferentiated mass (University of Georgia).

Step 2: run an automated check to create the first worklist

Run an automated check across the files you plan to address first. The goal is not to declare winners and losers. It is to estimate the type of work each file needs.

Look for:

  • missing or broken tags;

  • missing document language or title metadata;

  • untagged links and annotations;

  • scanned pages without a usable text layer;

  • form fields without labels or instructions;

  • tables with missing headers or unclear structure;

  • figures without descriptions;

  • repeated layout patterns that may be repairable in batches;

  • findings marked as needing a person.

This step separates obvious mechanical repairs from documents that need deeper human attention.

Step 3: apply candidate repairs to repetitive structural issues

Next, apply candidate repairs where the structure is clear enough. Preserve the original file and record the version being worked on. This matters when a later reviewer needs to understand what changed or roll back a bad repair.

Good candidates at this stage include headings, paragraphs, lists, links, form fields, obvious decorative artifacts, simple tables, and scan text layers. The point is to reduce the amount of manual tagging a person has to do from scratch.

Do not bury uncertainty. If the software is guessing, the file should carry that forward into review.

Step 4: manually review uncertain areas

This is where expert time belongs.

Review:

  • reading order across the whole document, not just one page;

  • image purpose and alt text;

  • complex tables and header relationships;

  • equations and formulas;

  • form labels, instructions, errors, and keyboard flow;

  • headings and lists where visual styling is ambiguous;

  • footnotes, captions, sidebars, pull quotes, and appendices;

  • scanned text for OCR errors.

For long documents, review the highest-risk sections first: instructions, forms, tables, figures, summaries, eligibility criteria, deadlines, and anything a reader must act on.

Step 5: test with assistive technology where risk warrants it

Validator output is not the reader’s experience. For high-impact PDFs, test with appropriate assistive technology and keyboard navigation.

The question is simple: can the reader find, understand, and use the content? If a report’s tags are technically present but a screen reader announces a page in the wrong order, the review is not done.

This does not mean every archived one-page flyer needs the same testing depth as a public benefits form. It means the test method should match the document’s audience, complexity, and risk.

Step 6: retain the repaired PDF and the evidence trail

Keep the repaired PDF and a record of what was checked, what changed, what remains unresolved, and what still needs a person. An audit trail for PDF remediation matters when work is split across teams, vendors, reviewers, and future updates.

At minimum, retain:

  • the original file;

  • the repaired file;

  • the checker report;

  • the date and version reviewed;

  • human-review notes for uncertain areas;

  • unresolved issues and their owner;

  • any assistive-technology testing notes.

This lets teams repeat the same triage, repair, review, and evidence process across a large collection without pretending every file received identical treatment.

Where Ordecto fits in the division of work

Ordecto, a PDF accessibility tool from the team behind Otio, fits where the workflow needs the automated check, candidate repair, and report, while keeping human review in the loop.

Ordecto checks every page against PDF/UA-1 with WCAG 2.2 as a referenced test model. It labels findings as failed, needs a person, or not tested, which is the right framing: a machine result is not shown as “accessible.”

For repair, Ordecto can tag common structure in reading order: headings, paragraphs, lists, tables, links, figures with descriptions, form fields, and decorative artifacts. Scanned pages can receive a text layer so their words can be read.

The review step is still part of the process. A person can inspect tags on the page, edit a tag, description, or reading order, and review the returned report instead of treating the repair as a certificate.

The output is the repaired PDF plus an accessibility report showing what was checked, what changed, and what remains unresolved or needs a person. That makes it useful for teams managing many PDFs under deadline: automation handles the repeatable work, while the report preserves the handoff to human judgment.

If you want the first triage step, use Ordecto’s free PDF accessibility checker and review every finding marked as needing a person before publishing.

A decision rule for assigning each remediation task

The cleanest way to divide the work is to ask two questions.

First: is the task repetitive, detectable, and structurally clear? If yes, automate it or use automation to produce a first repair.

Second: does the task depend on interpretation? If yes, assign it to a person.

Automate when the work looks like this:

  • applying common tags to headings, paragraphs, lists, links, figures, form fields, and simple tables;

  • identifying likely heading levels from consistent visual patterns;

  • marking obvious decorative content as artifacts;

  • creating a text layer for scanned pages;

  • surfacing missing form labels, missing descriptions, or untagged annotations;

  • repairing simple table structures where rows, columns, and headers are clear;

  • producing an initial report of failed, uncertain, and untested items.

Use human remediation when the work looks like this:

  • deciding the logical reading order of multi-column pages, sidebars, footnotes, callouts, captions, and appendices;

  • writing alt text that explains what an image means in this document;

  • resolving complex table relationships;

  • checking equations and accessible math representations;

  • interpreting unusual layouts;

  • deciding whether a bold line is a heading, label, emphasis, or table content;

  • checking whether a form’s instructions make sense before the user reaches each field;

  • testing whether the PDF is understandable with assistive technology.

Escalate rather than rely on a machine-only result when the file has high audience impact, legal or institutional risk, complex design, fillable forms, multilingual content, substantial scanned material, equations, or dense tables.

The practical next action is simple: check one representative PDF with Ordecto’s free checker, separate the automated failures from the findings that need a person, and use that split to plan the rest of the collection.

FAQ

Q: Is manual PDF remediation always more accurate than automatic remediation?
A: Not necessarily. Automation can handle repetitive structural work consistently, while human review is better for meaning, context, complex layouts, and final verification; the most dependable process combines both.

Q: How should a team decide which PDFs to remediate first?
A: Prioritise documents by audience impact, deadline, frequency of use, institutional risk, and complexity. Run an initial automated check to estimate the work, then give the highest-impact or most uncertain files earlier human review.