5.1 The Accessibility Problem

This chapter illustrates

✓ Commodity Intelligence · ✓ The Modeling Principle · ✓ The Alignment Principle

Before the case can carry much weight, the bar has to be clear. DocAble was built against a real institutional obligation, heterogeneous documents, expensive manual work, and a quality problem in which an output could look correct while quietly destroying the accessibility it was meant to add.

DocAble, now deployed in beta at scholaccess.com, accepts PowerPoint, Word, Excel, and PDF documents and remediates them toward the accessibility standards a public university must meet.

5.1.1 A Real Law, a Real Deadline

The Americans with Disabilities Act requires a public university to communicate with people who have disabilities as effectively as it communicates with everyone else. That obligation is old; what changed is where communication lives. Teaching has moved into digital artifacts — PDFs, slide decks, Word documents, spreadsheets, learning-management systems — and materials a screen reader cannot read, or that carry their meaning only through color or layout, hand some students a worse version of the university than others. The obligation is legal as well as practical: inaccessible digital content exposes institutions to complaints, litigation, and remediation costs — settlements have reached sums in the thousands to millions of dollars ¹1. Laura Carlson, “Higher Ed Accessibility Lawsuits, Complaints, And Settlements,” 2026, https://www.d.umn.edu/~lcarlson/atteam/lawsuits.html..

For most of that history universities complied reactively: a student hit an inaccessible document, requested an accommodation, and the institution scrambled to respond — the burden of discovering the barrier falling on the person it had already excluded. In 2024 the Department of Justice changed the shape. A final rule set a blunt premise: if a public entity provides a service through a website, an app, or the documents distributed through them, the accessibility of that content is part of the service. It set a technical standard — WCAG 2.1 Level AA — and, for large institutions, a deadline. The burden moved from the excluded student to the institution, and from the single flagged artifact to everything already published. Universities sit on decades of accumulated digital sediment: course sites, admissions forms, scanned readings, lecture videos.

5.1.2 What It Costs to Make One Document Accessible

The cost is easiest to see in one deck. Open a real PowerPoint presentation and run the built-in Accessibility Checker: a typical instructional deck can produce dozens of findings, from missing descriptions to reading orders the checker cannot verify. One deck used during this project produced 42.

Even routine findings take time, and semantically dense figures take more. Multiplied across a course and then across an institution, the manual burden becomes large enough to explain why remediation is so often deferred.^** The per-course estimate — on the order of $20,000 in faculty labor for one teaching load's materials, against vendor quotes near $3 to $40 a page — rests on a small model: findings per deck, minutes per finding, decks per course, and a loaded hourly rate. The assumptions and arithmetic are collected in the evidence ledger: Appendix H. Outsourcing does not make the semantic problem disappear: a vendor can repair structure, but a graduate-level figure may require subject knowledge only the instructor possesses. The result was a large heterogeneous inventory, specialized judgment, a compliance deadline, and no manual path that scaled to the estate.

5.1.3 Documents as Semantic Artifacts

A document is a visual and semantic artifact: text, figures, tables, reading order, hierarchy, emphasis — the small tricks by which humans convey meaning to other humans. To make it accessible, a machine has to recover enough of that meaning to present it through another channel: to a screen reader, to a keyboard, to a student who cannot see the figure the lecturer drew.

Earlier systems could automate pieces of the problem: extract text, flag a missing title, detect structural defects, generate limited descriptions. Frontier vision-language models changed the feasible boundary. A model could look at a rendered instructional slide, recover text and visual relationships, and often describe what the figure was trying to communicate. That capability made a new class of automation plausible. It did not make a production accessibility system.

The missing work was engineering. A model that can interpret a slide does not thereby know which parts of an output may change, how accessibility findings map to standards, which transformations preserve document semantics, what evidence should justify admission, or when a decision still requires human judgment. Those obligations live in the system around the model. Commodity intelligence supplied a powerful new component; modeling and alignment made that component usable inside a consequential system.

DocAble also crossed an uneven capability surface. The agents handled ordinary web and cloud infrastructure fluently and struggled much more with PDF and OpenXML internals. Whatever portion of that difference came from training-data fit, implementation complexity, or both, the engineering consequence was observable: in unfamiliar and consequential regions the project needed more explicit representation, stronger evidence, and more deliberate human judgment. The remainder of Part V reconstructs how the system responded to that uneven capability surface.