AI vs Manual Document Remediation: What the Data Says

Ask ten vendors whether AI can remediate a document and you will get ten answers shaped by what they sell. Tool companies say mostly yes. Services companies say mostly no. Neither is being dishonest. They are describing different documents.

There is a better way to settle it than opinion. The federal government publishes data on which accessibility defects actually turn up most often in public documents, and those defects do not all behave the same way under AI. Some are nearly solved. Some are nowhere close. Sorting them is more useful than arguing about percentages.

This piece covers the five defects most common in federal documents, what AI handles reliably today, where it struggles and why a wrong guess is worse than an obvious failure, what still needs a person, and how to split the work between them.

What can AI actually do in document remediation?

Automation reliably handles structure it can detect from the file itself: tagging headings, setting reading order on clean layouts, identifying lists and simple tables, adding document titles and language settings, and running OCR so a scanned page has real text behind it. What it cannot do is decide meaning. Anything requiring a judgment about what content communicates still needs a person.

The five defects that show up most

GSA’s governmentwide assessment tracks the most common accessibility failures in top-viewed public documents. The same five keep appearing:

1. Missing text alternatives for images
2. Missing information and relationships, meaning structure that is visual but not tagged
3. Meaningful sequence, meaning reading order that does not match the visual order
4. Missing page titles
5. Missing or wrong headings and labels

Line them up against what automation does well and a clear split appears.

Where automation is strong

• Page titles are close to solved: A title is a metadata field. A tool can read the file, generate or copy an appropriate title, and write it. There is almost no judgment involved, and no reason for a human to spend time on it.
• Headings and labels are largely solved for detection: AI can spot that a line of text is functioning as a heading based on size, weight, position, and spacing. Modern tools do this well.
• Simple tables and lists are mostly solved: If a table has a clear header row and consistent columns, automation can tag it correctly.
• Reading order is solved for single-column documents: A straightforward report, a letter, a basic form: automation gets these right at a rate that makes manual work hard to justify.

Where manual is strong

• Reading order breaks on complex layouts: Multi-column pages, sidebars, pull quotes, floating captions, and footnotes all create ambiguity about what comes next. The AI guesses. Sometimes it guesses well. The problem is that a wrong guess produces a document that sounds fluent and is actually scrambled, which is harder to catch than an obvious failure.
• Nested and irregular tables resist AI: Merged cells, tables with two header levels, tables where a row label is really a category heading: these need someone to decide what relationship the layout is expressing.
• Heading levels need judgment even when detection works: A tool can see that something is a heading. Deciding whether it is an H2 or an H3 means understanding how the document is organized, which is a question about argument, not formatting.

Where automation needs a second pair of eyes

Alt text is where the pairing shows most clearly: Modern engines describe images well, and for straightforward photographs the output is often close to final. What a tool does not know is why the image is in your document. The same photograph needs a different description in a medical journal than in an annual report, because the information and context a reader needs from it changes. The draft comes from the tool. The fit comes from a person, and that check is usually seconds rather than minutes.

Charts and graphs ask more of both: A useful chart description is not a description of the picture. It is a summary of what the data shows. Automation can extract the underlying values and produce a solid first pass. Deciding which trend is the point of the chart is a judgment about the document’s argument, and that stays with a person.

This matters more than it sounds. Missing text alternatives is the single most common defect in the GSA data, which means it is also the biggest available win. An AI engine that drafts alt text across a whole library turns the most frequent failure in federal documents into a review task rather than an authoring task.

So which should you choose?

Neither, as a whole-document strategy. The useful question is not AI or manual. It is how to decide which parts of this document need a person.

A reasonable split looks like this:

• Let automation do the structural pass on everything: Titles, language, headings, basic tables, reading order on simple pages, OCR on scans. This is fast, consistent, and does not benefit from human attention.
• Route by document complexity, not document importance: A simple fifty-page report may need almost no human time. A six-page brochure with infographics may need a lot.
• Spend human time on fit, not first drafts: Confirming alt text and chart descriptions suit the document, resolving complex tables, and checking reading order on anything multi-column.
• Verify regardless: An automated pass that looks finished is not the same as a document that works with a screen reader.

What this means in practice

Here is the split in plain terms.

AI does the first pass on every document. Tagging, reading order on simple pages, page titles, language settings, standard tables, OCR on scans, and a draft of the alt text. This is the bulk of the work by volume and it needs no human attention.

A person does the second pass on what carries meaning. Confirming the alt text fits the document, writing chart descriptions, untangling nested tables, and checking reading order on anything multi-column.

How much of each you need depends on the documents. A large backlog of similar, simply formatted files is mostly an AI job, and the economics are obvious. The cost and speed side of that trade-off is an important but different question from the defect-by-defect one this piece is asking. Design-heavy, data-heavy, or legally sensitive documents need far more of the second pass, and budgeting as though AI alone will finish them produces files that pass a checker and fail a user.

FAQ