6,200 PDFs, One ADA Deadline: How a State HHS Agency Cleared Its Backlog

This case study describes an AI-assisted PDF remediation deployment at a state health and human services agency serving residents across a state with a population of roughly 4.2 million. The agency's public-facing document library, benefit program guides, provider forms, public health notices, and administrative rules going back over a decade, had grown to more than 6,200 PDFs, the large majority of them scanned or exported without any accessibility tagging. The agency has requested anonymity for this case study; the accessibility coordinator quoted below has agreed to be identified by role only.

The Situation Before Remediation

The agency's document library had accumulated the way most large government PDF libraries do: over more than a decade, published by dozens of different staff members across 14 program areas, using whatever tool happened to be on hand. Some PDFs were exported from Word with reasonable structure intact. A large share, roughly 60% by the agency's own estimate, were scanned paper forms or reports with no underlying text layer at all, meaning a screen reader encountering them would announce nothing but a blank image.

Two staff members held accessibility remediation as part of a broader digital compliance role, alongside website accessibility, procurement review, and public records requests. Neither was remediating PDFs full time. At the pace they were able to sustain, roughly 15 to 20 documents a month between other duties, the existing backlog alone represented over 25 years of work at current staffing. And the backlog was still growing: the agency's program areas were publishing new benefit guides and provider notices faster than the two-person team could clear old ones.

What Was Actually in the Backlog

"6,200 PDFs" undersells how varied the backlog actually was, and that variety mattered for prioritization. Benefit program guides, the eligibility rules and application instructions for programs residents depend on to eat, get medical care, or keep housing, made up roughly 1,800 documents. Provider enrollment and billing forms for contracted clinics, hospitals, and community organizations accounted for another 1,400. Public health notices, outbreak advisories, boil-water alerts, immunization schedules, added close to 900, many time-sensitive and republished with minor updates year after year. The remainder split across administrative rules, internal policy manuals posted publicly by mistake or habit, audit reports, and a long tail of meeting minutes going back to the early 2010s.

That mix meant the backlog wasn't one problem. A scanned 2013 board packet nobody has opened in years carries almost no accessibility risk today. A 2026 benefit eligibility guide carries a great deal, because a resident with a screen reader who cannot read it cannot apply for the program it describes. The compliance office had never broken the backlog down this way before. Doing it first is what let prioritization target the documents that mattered, instead of working front-to-back through a folder.

The Cost the Agency Was Already Paying

The staffing math wasn't the only pressure. The compliance office tracked three accessibility complaints filed against the agency's document library in the eighteen months before the engagement began, two through the state's own ADA grievance process and one referred from a regional Department of Justice office after an advocacy organization flagged a batch of unreadable benefit forms. None had escalated to litigation. All three consumed weeks of staff time to investigate, respond to, and remediate the specific documents named, work that came directly out of the same two people who were supposed to be clearing the broader backlog. The agency's own playbook for handling that kind of complaint, developed after the first one caught the team flat-footed, is covered in the government ADA complaint response playbook.

There was a quieter cost too. Supervisors estimated that a meaningful share of benefit-line calls, the coordinator's working guess was 5% to 8%, came from callers who had tried and failed to use a PDF form online. Nobody had built a way to tag those calls as accessibility-driven, a gap the coordinator later flagged in the agency's own data. "We were pretty sure the PDFs were pushing people to the phone," she said. "We just couldn't prove it with a number, and you need a number to get budget attention."

The Deadline Pressure

The April 2024 Department of Justice rule under ADA Title II set a hard compliance deadline of April 24, 2026 for state and local government entities with populations over 50,000 or budgets over $50 million, requiring web content, including PDFs, to conform to WCAG 2.1 Level AA. This state's HHS agency, publishing benefit eligibility guides, provider enrollment forms, and public health notices used disproportionately by residents with disabilities and older adults, was squarely in scope, and squarely at risk.

"We weren't naive about the deadline. We'd been flagging the PDF backlog in our annual compliance reporting for three years running. But flagging a problem and having the staffing to fix it are two different things, and two people cannot manually remediate 6,000-plus documents in any timeframe that matters. We needed a different approach or we needed to explain to leadership why we were going to miss a federal deadline."

, Accessibility Coordinator, state health and human services agency

The Manual Remediation Cost Baseline

Before evaluating AI-assisted remediation, the agency's compliance office built a cost baseline for fully manual remediation, both to understand the scale of the problem and to make the business case to leadership for a different approach. Industry-typical manual remediation runs 20 to 30 minutes per document for a moderately complex PDF: tagging headings and reading order, adding alt text to images, marking table structure, fixing form field labels, and validating with a screen reader. Complex documents, long reports with nested tables or nonstandard layouts, run considerably longer.

The 25-plus year clearance timeline at current staffing was the number that moved leadership. A dollar figure is an abstraction; a timeline longer than most staff members' remaining careers is not. The agency's compliance office used this baseline to justify evaluating an AI-assisted alternative rather than requesting a large permanent headcount increase, which was unlikely to be approved in the state's current budget cycle.

The AI-Assisted Deployment

The agency deployed ADA Audit to triage, remediate, and validate the 6,200-document backlog. The engagement ran in three phases over five months, structured to prioritize the highest-risk and highest-traffic documents first rather than working through the backlog in arbitrary order.

Phase 1: Triage and Prioritization (Weeks 1–3)

Every document in the backlog was scanned and scored for accessibility severity (missing tags, no text layer, broken reading order, missing alt text, unlabeled form fields) and cross-referenced against page traffic data to identify the highest-impact documents. Benefit eligibility guides and provider enrollment forms, the documents residents with disabilities were most likely to need and least likely to be able to use, were prioritized into the first remediation batch regardless of document age.

Phase 2: AI-Assisted Remediation (Weeks 3–16)

The AI platform handled the bulk of remediation work automatically: OCR and text layer generation for scanned documents, heading structure and reading order tagging, table structure markup, and draft alt text generation for images and charts. The two-person accessibility team shifted from doing remediation by hand to reviewing and approving AI-generated output, a fundamentally different and much faster workflow. Documents with straightforward structure moved through review in minutes rather than the 20 to 30 minutes manual tagging previously required. Complex documents, multi-column public health reports with embedded charts, still required meaningful human review time, but even there, the AI-generated first pass eliminated the majority of the manual tagging work.

Phase 3: Validation and Sign-Off (Weeks 16–20)

Every remediated document was validated against WCAG 2.1 AA and PDF/UA criteria before being marked complete, including a screen reader spot-check for a statistically representative sample from each of the 14 program areas. Documents that failed validation were routed back for a second remediation pass rather than being marked done on the first attempt. The agency's compliance office signed off on the full backlog in month five.

Systems and Connectors

The document library lived across three sources: the public-facing CMS, a document management system where program staff drafted and versioned forms before publication, and a shared network drive holding years of archived public health notices never migrated into either system. Connecting all three mattered. A remediation project touching only the live website would have missed roughly 900 archived documents still linked from search engines and old bookmarks. The platform pulled all three sources into a single triage queue, so the two-person team worked from one prioritized list instead of three.

The Human Review Workflow

Review and approve is a simple phrase for a process the team built almost from scratch. Each reviewer worked from a queue sorted by risk score, opening the AI-generated remediation alongside the original PDF in a side-by-side view. Reviewers checked heading structure against the document's actual logical flow, since AI-suggested headings occasionally flattened a nested list or missed a section break that wasn't visually obvious. Alt text got the closest scrutiny. Draft alt text was accurate more often than not, but reviewers rewrote roughly one in five image descriptions, mostly for charts where the AI's summary was technically correct but missing the specific number a caseworker would need.

Approved documents moved automatically into the validation queue. Rejected ones went back into remediation with reviewer notes attached, so the second pass wasn't starting from zero. Tracking rejection rate by document type became a useful early-warning signal: a spike in rejected provider forms in week six traced to a nonstandard table layout the AI was misreading, a pattern fix that prevented the same error across another 200 similar documents still in the queue.

Staff Training

Neither staff member had reviewed AI-generated accessibility tagging before this engagement, and the training curve was real, if shorter than expected. Week one involved joint review sessions where both staff and the vendor's implementation team worked through the same 40 documents together, comparing notes on what the AI got right and where human judgment differed. By week three, the team had written its own internal reviewer guide, a one-page reference covering edge cases specific to this agency: multi-language forms, merged-cell tables, scanned documents with handwritten annotations. That guide, not anything from the vendor, is what new reviewers now get handed on day one.

Results

Cost and Time Compared

The agency's total cost for the AI-assisted engagement, including platform cost and the two staff members' review time across the five-month project, came to approximately $25,400 for the full 6,200-document backlog, an average of $4.10 per document against the $17.50 manual estimate. Measured in staff time alone, the two-person team's review-and-approve workflow consumed roughly 640 hours across five months, against a manual-remediation estimate of 2,583 hours for the same volume, a reduction of about 75% in direct staff time.

What Changed Operationally

Clearing the existing backlog solved half the problem. The other half was preventing a new backlog from forming. The agency built a new-document workflow requiring accessibility validation before any PDF is published to the public website, using the same AI-assisted tagging and validation process on new documents as they're created rather than letting them accumulate for a future remediation project. The two-person accessibility team now spends the majority of its time on this new-document review rather than working through historical debt.

"The number that mattered to me wasn't the dollar savings, though that helped make the case internally. It was that we went from a backlog we couldn't honestly promise to ever clear, to zero, in less time than it takes to plan a single legislative session. And we're not accumulating a new one, because accessibility review is part of how we publish now, not a project we do once and hope holds."

, Accessibility Coordinator, state health and human services agency

How Should Other Agencies Prioritize a Backlog Like This?

Prioritize by traffic and program impact, not by document age or alphabetical folder order. The accessibility coordinator's clearest piece of advice for peer agencies: don't start remediation with whatever document happens to be easiest to find. Pull page-view data first, even rough data from a basic analytics tool is enough, and start with the forms residents actually download most. A benefits eligibility guide with ten thousand annual downloads matters more than an archived policy memo from 2014 that nobody has opened in years, and a sequencing plan that ignores traffic data tends to spend early project momentum on the wrong documents.

Lessons for Agencies Facing the Same Deadline

Three things from this deployment generalize well beyond one state HHS agency. First, size the actual backlog before making any staffing or budget request. The agency's compliance office had been flagging "a PDF accessibility problem" for three years without a hard document count attached, and vague problems get vague, insufficient budget responses. A specific number, 6,200 documents, a specific manual-remediation cost estimate, and a specific timeline changed the conversation with leadership within a single budget cycle.

Second, build the human review step into the plan from day one rather than treating it as a fallback. Roughly a quarter of this agency's backlog needed meaningful human attention, complex multi-column reports, documents with legally mandated formatting, forms in languages other than English. An approach that promised full automation with no review queue would have either missed those documents' actual accessibility needs or bottlenecked the entire project waiting for someone to notice the automation wasn't sufficient for that subset.

Third, fix the publication pipeline, not just the backlog. Clearing 6,200 documents and then continuing to publish new, untagged PDFs the old way just rebuilds the same problem on a delay. The agencies still exposed after their compliance deadline passed are disproportionately the ones that treated remediation as a one-time project rather than a permanent change to how documents get published.

What Surprised the Team

Ask the accessibility coordinator what she'd tell a peer starting the same project, and the honest answer isn't on the vendor's feature list. Two things caught the team off guard, one in a good way and one that took real adjustment.

The good surprise was how much the reviewers' own judgment improved, and how fast. By month three, the two staff members were catching structural problems in the source documents themselves, not just in the AI's tagging of them: a benefits form where two questions shared a single answer field, a defect that predated the accessibility project entirely. "We started this project as taggers," the coordinator said. "We ended it thinking like editors. That's probably worth more long-term than the backlog number."

The adjustment was cultural, not technical. Program staff who had been publishing PDFs their own way for years did not universally welcome a new validation gate sitting between their draft and the public website. A few program areas treated the first weeks of the new workflow as a bottleneck rather than a safeguard, and the compliance office spent more time than expected explaining why a form that looked fine on screen could still fail for a screen reader user. That friction eased once program staff saw their documents pass review faster than the old manual process ever had. It didn't disappear on day one, and the team underestimated how much of the project would be about people rather than PDFs.