Why I Built an Accessibility Engine That Treats AI as a Last Resort

Two months ago, someone on a government compliance call said something that stuck with me:

"We tried three AI tools for PDF remediation. They all hallucinated alt text and broke our table structures."

That's the state of accessibility tooling right now. The ADA Title II deadline is April 24, 2026. Every government entity serving 50,000+ people needs all digital content to be WCAG 2.1 AA compliant. Most have thousands of legacy PDFs that have never been touched. The manual remediation industry charges $7–$20+ per page, and the backlog is enormous.

The obvious move is to throw a large language model at the problem. Upload a PDF, have GPT-4 or Claude "fix" it, ship the output. This is what most tools in this space do, and it's why compliance officers don't trust any of them.

What actually goes wrong when AI drives the whole process

Structure is a parsing problem, not a language problem. A PDF's tag tree – the thing screen readers use to navigate the document – is a deterministic data structure. It has headings, paragraphs, lists, tables, and artifacts. Repairing a broken tag tree doesn't require "understanding" the content. It requires parsing the existing structure, identifying what's wrong, and applying the correct fix from a known set of rules. An LLM doesn't do this better than purpose-built logic. It does it worse, and unpredictably.

Contrast is math, not generation. If foreground text doesn't meet the 4.5:1 contrast ratio against its background, the fix is a color calculation. You don't need a neural network to adjust hex values. You need the WCAG contrast formula and a color space library.

Reading order is spatial logic. PDFs don't store content in the order a human reads it – they store it in the order it was drawn on the page. Fixing reading order means analyzing spatial coordinates, column detection, and flow inference. This is computational geometry, not natural language processing.

But some things genuinely need AI. Generating meaningful alt text for a photograph of a building, a chart with data points, or a complex diagram – that's a vision problem. This is where AI is not just useful but necessary.

Use the right tool for each layer

So I built Foresera with a simple architecture principle: use the right tool for each layer of the problem.

The engine has multiple stages. First, it parses the PDF structure and identifies every element that needs remediation. Then it runs each issue through the appropriate fix layer – deterministic rules for structure, math for contrast, spatial logic for reading order. Only when an issue can't be resolved by these systems does it escalate to an AI model. The most common escalation is alt text generation for images, where vision AI genuinely excels.

After all fixes are applied, the engine runs a full re-audit – 85+ WCAG checkpoints – against its own output. This is the part most tools skip entirely. You don't just get a "fixed" PDF; you get a conformance report proving what was checked, what was fixed, and what passed.

The result: upload a PDF, get it back compliant in under 60 seconds.

What I learned building this

Government buyers care about determinism. When a compliance officer signs off on a document, they need to know the fix is reproducible and auditable. "The AI decided to tag this as a heading" is not an acceptable answer in a legal compliance context. "The heading was detected based on font size hierarchy and structural position" is.

Speed matters more than I expected. The agencies dealing with this problem don't have 10,000 PDFs to fix "eventually." They have a legal deadline. The ability to process a document in 60 seconds instead of sending it to a contractor for a week changes the math on whether compliance is even feasible.

AI-as-fallback is a better product positioning than AI-first. When I stopped saying "AI-powered" and started explaining the actual architecture, conversion rates went up. People in this space have been burned by AI tools that overpromise. Telling them exactly when and why AI is used – and when it isn't – builds trust.