AI Coding Accessibility Audits: Calculate Cost per Verified Page

An AI accessibility audit is not finished when it produces a list of warnings. A useful workflow discovers pages and states, runs deterministic scanners, inspects visual and semantic behavior, proposes fixes, executes regression tests, and preserves human judgment for issues automation cannot establish.

Cost per warning is a bad metric because one missing label and one broken keyboard workflow have different remediation effort and impact. Use cost per verified page-state or cost per accepted issue fixed. A page-state includes route, viewport, authentication state, theme, language, and any modal or interaction required to expose the interface.

Inventory the Audit Pipeline

Direct costs can include model tokens, screenshots or visual tokens, browser runtime, static-scanner time, CI, preview deployments, and storage for evidence. Human costs include triage, keyboard and screen-reader checks, design decisions, and review of generated code. Scanners catch many mechanical issues cheaply, so send structured findings to the model instead of asking it to rediscover them from a screenshot.

Separate discovery from remediation. An audit pass should produce stable issue IDs with rule, element, page-state, evidence, severity, and confidence. A fix pass can group issues that share a component. Without stable IDs, reruns create duplicate findings and make it impossible to measure closure.

A Worked Page Budget

Suppose 50 page-states require 1M model input tokens, 150K output tokens, 90 browser minutes, and $4 of CI and previews. At $1/M input and $6/M output, model cost is $1.90. Browser runtime at $0.04 per minute adds $3.60, bringing machine cost to $9.50, or $0.19 per audited state.

Human verification takes four hours at $80 per hour, adding $320 and lifting total cost to $6.59 per state. This is normal: accessibility depends on meaning and actual interaction, not only syntax. The agent's value is reducing repetitive discovery and patch work so experts spend their time on judgment.

Reduce Duplicate Work

• Audit shared components first. Fixing a button, dialog, or navigation primitive can close findings across many routes.
• Deduplicate by DOM signature. Group the same component and rule rather than paying to explain every occurrence.
• Capture focused evidence. Crop screenshots and retain the relevant accessibility tree instead of full-page noise.
• Test changed states. Re-run the affected component and its consumers before a full-site regression pass.

Keep Human Verification in Scope

Automated and model-assisted checks cannot prove that alternative text conveys the right meaning, focus order matches the task, announcements are understandable, or a complex interaction works well with assistive technology. Budget manual sampling according to severity and novelty. Reusable, well-tested components can earn a lower sample rate over time.

Track false positives, accepted fixes, reopened issues, pages verified, component reuse, reviewer minutes, and regression escapes. A model that reports more issues can be worse if reviewers discard them. Quality-adjust the cost by accepted findings and durable fixes.

Prioritize by Reach and Severity

Multiply issue severity by component reach and user-path importance. A focus defect in the shared navigation can affect every route, while a contrast issue on an internal prototype has narrower exposure. This scoring directs agent and reviewer time toward fixes that improve the most experiences, without claiming that low-severity issues never matter.

Maintain an exception record when a finding cannot be fixed immediately. Include owner, reason, compensating behavior, target date, and affected states. Re-evaluate exceptions when components change. Otherwise repeated audits spend tokens rediscovering the same known issue and reviewers repeat the same decision.

Build Prevention into CI

Turn durable fixes into component tests, lint rules, design tokens, or browser assertions. Prevention has a higher initial cost but avoids paying the audit pipeline on every occurrence. Keep CI checks deterministic; use the model to explain and remediate failures, not to decide whether a stable rule passes.

Sample production pages after deployment because build-time fixtures may not include real content, localization, feature flags, or third-party widgets. Compare new findings with stable issue IDs and attribute regressions to the introducing change. This closes the loop from discovery to prevention and gives cost per verified page a quality dimension.

Forecast recurring coverage separately from a one-time baseline. The first audit pays to inventory routes, build fixtures, and establish evidence formats. Later runs should reuse that foundation and focus on changed components and risk-based samples. Report setup amortization explicitly so stakeholders do not interpret a cheaper second month as a sudden model improvement.

Bottom Line

Use deterministic scanners for cheap facts, agents for synthesis and remediation, and people for semantic and interaction judgment. Attribute all work to stable page-states and issue IDs, then divide total spend by verified outcomes. The cheapest audit is the one that fixes shared causes and prevents regressions, not the one that emits the longest report.