Proving AI works for everyone, at the speed AI ships

02 — Strategy

Three layers, one source of truth

I gave every one of the 55 criteria an owner, matched to whoever is best at testing it.

What Layer 2 tests

Guardrails I set before the first test

• Never handle credentials. Claude reuses my signed-in session.
• Never submit forms that create records or send anything. It captures a screenshot and logs the form instead.
• Never guess beyond what its tools can measure, and never overwrite a verdict the scanner already settled.
• Never decide a borderline case alone.

03 — Platform baseline

Seven AI capabilities, one honest picture

Layer 1 ran across the whole platform. For the first time we could see all seven capabilities side by side, and see how much the scanner simply couldn't tell us.

Lighthouse · Fabric / Harmony · Pipelines screen

* Document AI and Engage figures are derived estimates. 573 of the 700 failures are exact counts from five capability workbooks.

04 — Calibration

Earning trust before scaling

Before pointing Layer 2 at our product, I calibrated it against a public benchmark site. Claude ran all 55 criteria on 6 screens, for 330 checks in total, and graded how confident it was in every row.

1. 9criteria flagged for human review after the first pass
2. 3 movesmeasure every gap, cite W3C guidance on judgment calls, and read the real accessibility tree
3. 4 / 330rows still needed a human call (1.2%)

The turning point came when Claude reversed its own call. It had marked a link as meeting the 24px target size without measuring it. On a second check it measured the link at 82.4 × 18.5px and changed its answer to Fail. A re-run days later gave identical results. That's when I knew the method was safe to scale.

05 — Full audit

All 55 criteria on the product's front door

The AI Next landing page is the first screen every buyer and user sees. I audited it with all three layers together, for a full A and AA picture.

• Pass 21
• Fail 13
• Partial 3
• Not evaluated 1 (flash-rate)
• +17 not applicable

edgeverve.com/ai-next · live capture, annotated

{ autoplay:true, autoplayTimeout:5000, loop:true }

outline:none · box-shadow:none

<button><i class="fa-search"></i></button> // name: ""

<input placeholder="E-mail"> // no label

edgeverve.com/ai-next · lead form · live capture, annotated

Where the AI stopped and I stepped in

To test error handling, someone had to submit a live lead form. When Claude tried, the platform's safety layer stopped it, so I submitted the form myself with clearly labeled test values. Claude then inspected what came back:

What a sighted user sees · real submission

What assistive technology receives

<input>

aria-invalid: not set

aria-describedby → nothing

<span> error

role: none · aria-live: none

not linked to its field, never announced

Partial · Errors identified (3.3.1) The errors are specific, but assistive tech can't see them. Pass · Fix suggestions (3.3.3) The messages suggest how to fix the input. Fail · Status messages announced (4.1.3) No live region exists anywhere on the page.

06 — Before → After

What changed

07 — Design system

Fix the system, not the screen

I traced every failure back to the component or token that caused it. Ten fixes close all 16 open findings on the front door. Because they live in shared components, they'll reach every screen that uses them.

08 — Leadership and alignment

Six groups, one plan

An AI-run audit inside an enterprise needs a yes from people who are measuring very different risks. I designed the program around each one.

10 — Business impact

What this is worth to the business

How the effort numbers were calculated

Coverage is measured. The effort and timeline figures come from a planning model, not from timed results. Every input is shown below so you can check the arithmetic, or swap in your own assumptions.

12 — How I work

What this project says about me as a designer

• I reframe before I execute

  The ask was to fill gaps in a spreadsheet. I made it about whether people can use our AI products, and whether we can prove it.Evidence: three-layer strategy, chapter 02

• I design systems, not screens

  I traced 13 failures back to 10 shared fixes, so the impact reaches every capability.Evidence: remediation plan, chapter 07

• I make risk visible and manageable

  I set guardrails before the first test, named the unknowns, and treated the AI's refusal to submit a live form as the system working.Evidence: guardrails and the human-in-the-loop test, chapters 02 and 05

• I lead without authority

  Security, engineering, marketing and leadership each had a reason to say no. I designed the plan so each could say yes.Evidence: stakeholder plan, chapter 08

• I speak the language of the business

  Coverage, effort, timelines and procurement readiness, all quantified and labeled honestly.Evidence: business impact, chapter 10

• I use AI as a multiplier, not a crutch

  Claude does the hands-on testing, which comes to hundreds of hours across the full rollout. I own the method, the trust model and every judgment call.Evidence: calibration, chapter 04

Glossary

Terms used in this case study

WCAG 2.2, Level A and AA

The W3C's international accessibility standard. A and AA are the 55 checks most laws and buyers expect.

Criterion (for example 2.4.7)

One individual check in the standard, such as “Visible focus.”

Screen reader

Software that reads the interface aloud for blind and low-vision users.

Keyboard focus

The highlight that shows which control a keyboard user is on.

Accessibility tree

The browser's description of a page that assistive technology reads: names, roles and states.

Landmark

A labeled region, such as main content or navigation, that lets people jump around a page.

Live region (aria-live)

Code that tells a screen reader to announce updates, like an error message, as they appear.

aria-describedby

Code that links an error message to its field, so it's read out with that field.

Lighthouse

Google's automated audit tool. It can check only part of WCAG.

VPAT / ACR

A Voluntary Product Accessibility Template, or Accessibility Conformance Report. It's the document buyers request to see how well a product conforms.

Design token

A shared design value, like a color or focus style, that one change updates everywhere.