A retailer's onboarding course was rebuilt with an AI avatar presenter and AI narration in three days, and it looked polished enough to ship. Then the accessibility reviewer ran it and failed it on six counts: auto-generated captions that misheard the product names, an avatar whose lip-sync drifted from the audio, color contrast on the AI-designed slides below the threshold, and three interactive screens a screen reader could not navigate. The fix took longer than the build. The next course was built differently: every accessibility requirement was a generation instruction, not a review note, and the conformance check ran as a build gate, not a final inspection. It passed on the first review, and it shipped with a conformance report a procurement team could read. That shift, accessibility produced and verified inside the build instead of after it, is Stage 3 of the pipeline.

Accessibility Is a Gate, Not a Polish Step

This is Stage 3 of the source-to-certified-course pipeline. Stage 1 made the content true and sourced. Stage 2 made the assessment valid. Stage 3 makes the experience usable by everyone, including the learners who use a screen reader, cannot hear the narration, cannot see the screen clearly, or cannot use a mouse. The bright-line rule of the program states the stakes plainly: an AI-generated experience that fails WCAG 2.2 AA does not ship. Accessibility is a gate, not a polish step, full stop.

Two terms anchor the stage. WCAG 2.2 AA is the Web Content Accessibility Guidelines, version 2.2, at conformance level AA, the W3C standard published as a Recommendation on 5 October 2023 and the practical conformance target for learning content. Why you care: AA is the level most organizations and regulations require, and a course that misses it can be legally non-compliant, unusable for part of your workforce, and a procurement blocker. Section 508 is the US federal requirement that incorporates WCAG by reference, so for many employers WCAG conformance is not a nicety but a legal obligation; note that the version Section 508 legally incorporates can lag the latest WCAG, so you confirm which version applies to you. The reason accessibility must be a gate rather than a polish step is economic and ethical at once: retrofitting accessibility onto finished AI media is the slow, expensive rework the opening scene describes, and every screen that ships inaccessible is a learner locked out of training their job requires.

AI changes the accessibility picture in a specific way that this stage exists to manage. AI media tools generate captions, voices, avatars, images, and interactions at speed, and each of those is a fresh opportunity to fail conformance silently. An auto-caption mishears a term. A synthetic voice mispronounces a drug name so the audio and the transcript disagree. An AI-designed slide picks a trendy low-contrast palette. An AI-generated interaction is built for a mouse and ignores the keyboard. The speed that makes AI media attractive is exactly the speed at which accessibility failures multiply, which is why the check has to live inside the build.

There is a structural reason AI media is so prone to silent accessibility failure, and understanding it changes how you supervise the tool. Most AI media generators are optimized to produce output that looks good to a sighted person clicking through with a mouse, because that is what their training data and their demos reward. They are not optimized for the screen-reader user, the keyboard-only user, or the low-vision user, because those experiences are invisible in a quick visual review. The result is a tool that is genuinely excellent at the surface a designer sees and genuinely indifferent to the surfaces a designer does not see unless they deliberately go looking. This is not a flaw you can prompt away entirely; it is a bias in what the tool optimizes for, and the only reliable countermeasure is to make the unseen surfaces a deliberate requirement of the build and a deliberate object of the check, rather than trusting that polished output is accessible output. Polished and accessible are correlated for sighted mouse users and uncorrelated for everyone else.

AI does not make media accessible faster. It makes inaccessible media faster. The conformance gate is what turns the speed into an asset instead of a liability.

Building Conformance In, Not Bolting It On

The core move of Stage 3 is to convert every accessibility requirement from a review note into a generation instruction, so the AI produces conformant media from the first pass. Here are the load-bearing requirements and how each becomes a build instruction rather than a rework ticket.

Captions and Transcripts

Every piece of audio and video needs accurate captions and a transcript, so a learner who cannot hear can still access the content. AI auto-captions are a draft, not a deliverable: they mishear proper nouns, technical terms, and product names, and an inaccurate caption is an accessibility failure that also teaches the wrong word. The build instruction is to generate captions from the verified narration script, not from speech recognition of the audio, and to have a human correct the terms against the Stage 1 source. The transcript is generated alongside, including descriptions of meaningful visuals, so the audio-only learner loses nothing.

Alt Text That Carries Meaning

Alt text is the text description of an image that a screen reader announces, so a learner who cannot see the image still gets its meaning. AI can draft alt text, but it tends to describe what is literally in the image rather than what the image means in the lesson. A diagram of a circuit gets "a diagram with lines and boxes" when the instructional point is "the breaker sits between the panel and the load." The build instruction is to generate alt text tied to the instructional purpose of the image, grounded in the objective the image supports, and to mark purely decorative images so the screen reader skips them rather than reading noise.

Contrast and Visual Design

Text and meaningful graphics must meet WCAG 2.2 AA contrast ratios against their background, so a learner with low vision can read them. AI slide and image generators routinely pick aesthetic palettes that fail contrast, and color is often used as the only way to signal meaning, which fails learners who cannot distinguish colors. The build instruction is to specify conformant contrast and a non-color signal (a label, a pattern, an icon) in the generation prompt, and to verify the ratios rather than trust the look.

Keyboard and Screen-Reader Operability

Every interaction must work with a keyboard alone and be navigable by a screen reader, so learners who cannot use a mouse are not locked out. AI-generated interactions and drag-and-drop activities are frequently built for the mouse and the eye, with no keyboard path and no programmatic labels. The build instruction is to require keyboard operability and proper labels in the interaction spec, and to test the actual navigation, because this is the failure least visible to a sighted designer using a mouse and the one a screen-reader user hits first.

Synthetic-Media Disclosure

When a presenter is an AI avatar or the narration is an AI voice, learners and often your legal team expect a synthetic-media disclosure, a clear statement that the presenter or voice is AI-generated. Why you care: undisclosed synthetic media erodes trust and, increasingly, runs into disclosure expectations and policies; it is cheap to state and expensive to omit. The build instruction is to include the disclosure as a standard element of any AI-presented media, decided once as policy rather than per course.

The Conformance Check and the ACR

Building conformance in does not remove the check; it makes the check pass. Stage 3 ends with a conformance check that is a build gate: the media does not advance to publication until it passes, exactly as a unit test gates code. And the check produces an artifact that proves it ran.

That artifact is the VPAT or ACR. A VPAT (Voluntary Product Accessibility Template) is the standard document an organization fills in to report how a product conforms to accessibility standards; the completed VPAT is called an ACR (Accessibility Conformance Report). Why you care: the ACR is what a procurement team, a legal reviewer, or a customer asks for to prove a course meets WCAG 2.2 AA, and producing one from a course built with the conformance gate is straightforward because the evidence was generated as you went. A course built without the gate cannot produce an honest ACR, because nobody recorded whether each criterion was met.

It is worth being precise about what "passing the gate" means, because it is easy to confuse two very different things: automated checks and human verification. Automated accessibility scanners can catch a meaningful share of failures, missing alt-text attributes, contrast ratios below threshold, missing form labels, and they are worth running because they catch the mechanical errors fast. But automated tools cannot judge whether alt text actually conveys the instructional meaning of an image, whether captions are accurate to the spoken terms, whether a keyboard path is not just present but usable, or whether color is the only carrier of a meaning. Those judgments require a human, and they are exactly the failures AI media produces most. So the gate is a combination: run the automated scan to clear the mechanical issues quickly, then have a human verify the judgment-dependent criteria that no scanner can settle. Treating the automated scan as the whole gate is itself a failure mode, because a course can pass every automated check and still be unusable for a screen-reader learner whose alt text is technically present and instructionally empty.

A Worked Example: The Onboarding Rebuild

Watch the same onboarding course built two ways.

Before, accessibility as a final inspection. The team generates an AI avatar presenter, AI narration, AI-designed slides, and three drag-and-drop interactions in three days. The course looks finished, so it goes to accessibility review as the last step before launch. The reviewer fails it on six counts: auto-captions misheard two product names, the transcript was missing, the slide palette failed contrast, color alone signaled the correct workflow path, two interactions had no keyboard route, and the AI avatar carried no disclosure. Each failure is now a separate rework against finished media: re-caption, write the transcript, recolor and re-export every slide, rebuild the interactions for keyboard, add the disclosure. The rework takes longer than the original build, and the launch slips. When procurement asks for an accessibility conformance report, the team cannot produce a credible one, because the course was never built to a standard, only inspected against one at the end.

After, accessibility as a build gate. The same team starts Stage 3 with accessibility as generation instructions. Captions are generated from the verified narration script and a human corrects the product names against the Stage 1 source. Alt text is generated tied to each image's instructional purpose, decorative images are marked. The slide generation prompt specifies conformant contrast and a non-color signal for the workflow path. The interaction spec requires keyboard operability and screen-reader labels, and the designer tests the navigation. The avatar carries the standard synthetic-media disclosure by policy. Then the conformance check runs as a gate before publication, criterion by criterion, and the media passes on the first review because it was built to pass. The course ships on time, and an ACR is assembled from the evidence captured during the build. When procurement asks, the answer is the report, handed over without a scramble.

Both courses used the same AI media tools and produced the same polished look. The difference was when accessibility entered: as a final inspection that turned speed into rework, or as a build gate that kept the speed and added a conformance report. The first was three days of build and a week of rework. The second was the genuine fast-and-conformant win, because a course that fails WCAG 2.2 AA does not ship, and the only cheap way to pass is to build to the standard from the first generation.

Accessibility retrofitted is the most expensive media work in L&D. Accessibility built in is nearly free, and it is the only version that ships on time with a report attached.

The Human Owns Conformance

Stage 3 lands the iron rule on accessibility. AI can draft captions, alt text, contrast-aware designs, keyboard-operable interactions, and the disclosure, fast and at scale. It cannot certify that the experience conforms, because conformance is a claim with legal and ethical weight, and a claim is a human's to make and to answer for. AI assists, the human verifies conformance, the human owns the decision to ship, and "the AI generated it accessibly" is never a defense to an accessibility auditor.

In practice this means the conformance check is run and signed by a human, not asserted by a tool that generated the media. An AI caption generator's claim that captions are accurate is the model marking its own homework; a human confirms the terms against the source. An AI design tool's claim that contrast passes is checked with a contrast measurement, not taken on faith. The keyboard path is tested by actually navigating it, because the failure is invisible to a mouse user. And the ACR is signed by the person who verified each criterion, because when an auditor or a locked-out learner raises a barrier the course shipped with, the question "who confirmed this was accessible" is answered by that human, with the conformance evidence they recorded, not by pointing at the avatar. Build to the standard at AI speed, run the gate before publication, sign the report, and own the decision to ship, because the model generated the media but the human certifies that everyone can use it.

Key Takeaways

• Stage 3 of the pipeline produces and verifies media to WCAG 2.2 AA inside the build, because the bright-line rule is that an AI-generated experience that fails WCAG 2.2 AA does not ship; accessibility is a gate, not a polish step.
• WCAG 2.2 AA (W3C Recommendation, 5 October 2023) is the conformance target; Section 508 incorporates WCAG by reference, so for many employers conformance is a legal obligation, and the legally incorporated version can lag the latest WCAG.
• AI does not make media accessible faster, it makes inaccessible media faster: auto-captions mishear terms, alt text describes literally not instructionally, AI palettes fail contrast, and AI interactions ignore the keyboard.
• The core move is converting every accessibility requirement into a generation instruction: captions from the verified script, alt text tied to the objective, conformant contrast plus a non-color signal, keyboard operability, and a synthetic-media disclosure.
• Generate captions from the verified narration script rather than speech recognition, and have a human correct technical terms and product names against the Stage 1 source, because an inaccurate caption also teaches the wrong word.
• The conformance check runs as a build gate before publication, like a unit test gates code, and produces a VPAT or ACR (Accessibility Conformance Report) that procurement, legal, and customers ask for as proof.
• A course built without the gate cannot produce an honest ACR, because nobody recorded whether each criterion was met; the after-version shipped on time with the report assembled from build evidence.
• The iron rule lands on conformance: AI can draft accessible media, but a human verifies conformance, signs the ACR, and owns the decision to ship, because 'the AI generated it accessibly' is never a defense to an auditor.