a11y heuristic framework

The temptation in 2026 is to point AI at a product and ask for an accessibility report. That produces output that looks rigorous and isn't — generic, inconsistent with the product's real state, prone to citing WCAG criteria incorrectly.

I designed a method where AI amplified analytical capacity without generating conclusions. Every finding came from human judgment first. AI reviewed, verified, and questioned — never authored.

The 5-layer framework

Layer 1 — Independent manual analysis. Two evaluators worked through 64 criteria per screen separately, no comparison until done. Reduces confirmation bias.

Layer 2 — Debrief and consolidation. Disagreements surfaced and resolved. Ambiguous items flagged for deeper investigation.

Layer 3 — AI-assisted review. Each consolidated finding submitted to Claude for critique: imprecise WCAG citations, missing nuance, weak reasoning. AI questioned the analysis. Humans decided what to revise.

Layer 4 — Technical verification. Claude with browser access ran live JavaScript checks: ARIA attributes, CSS values, element measurements. Findings backed by code, not impressions.

Layer 5 — Triangulation. Four independent reports (analyst 1, analyst 2, AI review, technical verification) cross-referenced. Contradictions resolved with code evidence.

Each layer addresses a limitation of the one before it. The triangulation is what gives the final report confidence.

What AI did well

• Critical review — flagged WCAG citations that were wrong, surfaced inconsistent logic

• Console verification — ran scripts to confirm CSS, ARIA, element dimensions that would take hours manually

• Report cross-referencing — identified real contradictions vs. differences in granularity

• Structured documentation — turned findings into evidence guides and Confluence-ready reports
  
  

What AI didn't do

• Generate initial findings autonomously — that was human judgment

• Decide what counted as a violation when interpretation was ambiguous

• Replace testing with real assistive technology (NVDA, VoiceOver)

• Prioritize fixes — the impact×effort matrix was built with engineering

Naming these limits is part of the method. Without them, AI-assisted analysis becomes AI-generated analysis dressed up as rigor.

From audit to action

The audit produced 192 evaluated points across three screens. With the dev team, I ran a workshop to plot findings on impact × effort and group corrections into four priority waves.

The output wasn't a report. It was a backlog with evidence — each item tagged with the type of evidence captured (code snippet, contrast measurement, ARIA state, screenshot), so engineers could verify and fix without re-running the diagnosis.

A living document tracks correction progress across the waves.

Reflection

The most useful AI in design isn't the one that generates faster. It's the one that questions better.

This method took longer than a traditional heuristic checklist. The trade wasn't speed — it was depth, reliability, and engineering trust. For an audit that needed to become real engineering work, that trade was the whole point.

→ Full method, prompts, and what didn't work: [Read the article on Medium]