Dana runs talent acquisition at a 1,200-person regional health system that hires roughly 600 people a year across clinical and administrative roles. Last quarter, a strong candidate for a data analyst role withdrew after the one-way video interview, and the exit note said only that the process did not work for them. Dana followed up and learned the candidate had a stutter that the timed, recorded format made unbearable. That candidate would have been a strong hire. Dana realized her new AI screening stack had quietly built a wall, and she had no idea how many other candidates had hit it and simply vanished from the funnel without a word. This lesson is the audit she ran next, and the process she rebuilt as a result.

What Accessibility Actually Means in Recruiting

Accessibility is not a niche concern bolted onto recruiting at the end. It means your recruiting process works for people across a wide range of abilities, and it is the difference between a talent pool that includes the roughly one in four working-age adults who report a disability and one that silently filters them out. When Dana makes her process accessible she is not only meeting a legal obligation; she is recovering candidates her competitors are losing. Her audit began by mapping the categories of disability her process could exclude, and the concrete design response for each, because a general commitment to inclusion changes nothing until it becomes a specific change to a specific screen.

Mobility Differences

Candidates with limited ability to use a mouse or keyboard need keyboard navigation, voice-control compatibility, and large click targets. An application form that requires precise drag-and-drop ordering of work history will block these candidates before they ever reach a human reviewer, and it will do so invisibly, since an abandoned application produces no signal that anything was wrong. Dana treats any interaction that can only be completed with fine pointer control as a defect rather than a design preference.

Vision Differences

Candidates who are blind or have low vision rely on screen readers, alternative text for every meaningful image, and sufficient color contrast. A resume upload box that says to drag your file here with no keyboard fallback is a hard stop, and so is a CAPTCHA with no audio option. These are among the easiest barriers to find and the most common to leave in place, because the people who build and test application flows are rarely the people who navigate them with a screen reader.

Hearing Differences

Candidates who are deaf or hard of hearing need captions on any video content, a transcript option in place of a phone screen, and written communication channels. A recruiter who only schedules phone screens has excluded this group by default, without ever making a decision that looks like exclusion. Offering a written or video-relay alternative at the point of scheduling, rather than waiting for someone to ask, is what converts the policy into practice.

Cognitive Differences and Neurodivergence

Candidates with ADHD, autism, dyslexia, or related profiles benefit from clear instructions, extra time, reduced sensory load, and a genuine choice between written and spoken interaction. This is the group most often harmed by AI assessment, and the one Dana's video tool failed. It is also the group whose exclusion is hardest to see, because the barrier is usually not a locked door but a format that quietly costs them more than it costs anyone else to get through.

Mental Health Conditions

Candidates managing anxiety, depression, or similar conditions are served by accommodating communication and a process that does not manufacture unnecessary stress. A surprise recorded interview with no preparation time manufactures exactly that kind of stress, and it measures a candidate's tolerance for it rather than their ability to do the job. Predictability, advance notice of format, and the option of a live conversation cost the employer almost nothing and remove a real barrier.

Accessibility sits on three legs at once: a legal requirement, an ethical obligation, and a business case. Dana found the business case the most persuasive to her CFO, though she was careful never to let it stand in for the other two. At 600 hires a year, even a one-percent improvement in conversion from a previously excluded group is real headcount filled at no added sourcing cost. The legal obligation, though, does not become optional when the business case is weak; it is the floor beneath the whole conversation.

The Legal Floor: ADA, WCAG, and Local Bias-Audit Law

Dana works in the United States, so her baseline is the Americans with Disabilities Act. Title I of the ADA prohibits discrimination against qualified individuals with disabilities in the hiring process and requires reasonable accommodation unless it imposes undue hardship on the employer. That is a legal duty, not a courtesy, and the phrase that carries the weight is "qualified individual": someone who can perform the essential functions of the job with or without reasonable accommodation. A process that screens out a person who could do the job with an accommodation is the exact scenario the statute is written to prevent.

The ADA does not name a specific technical standard, but courts and the Department of Justice have repeatedly pointed to the Web Content Accessibility Guidelines as the practical benchmark for whether a digital application is accessible. WCAG 2.1 Level AA is the version most legal and procurement teams now require, covering keyboard operability, text alternatives, contrast ratios, and predictable navigation. For organizations operating in Ontario, the Accessibility for Ontarians with Disabilities Act makes WCAG conformance an explicit statutory requirement for many employers rather than a recommended benchmark. Dana's health system has a clinic across the border, so her procurement checklist references both.

Two adjacent laws shape the AI side specifically. New York City Local Law 144 requires an independent bias audit of any automated employment decision tool used on candidates in the city, plus advance notice to candidates and an opportunity to request an alternative process. While Local Law 144 centers on protected-class bias rather than disability as such, the request-an-alternative provision is exactly the accommodation hook accessibility depends on, and Dana treats the two obligations as one workflow rather than two compliance projects. Separately, when an AI assessment screens out neurodivergent candidates at a disproportionate rate, that can trigger EEOC scrutiny under both the ADA and disparate-impact analysis.

The EEOC issued guidance in 2022 warning specifically that algorithmic hiring tools can violate the ADA when they screen out people who could do the job with reasonable accommodation. Dana reads that guidance as the single most important sentence in her compliance file, because it closes the gap a lot of teams try to slip through: the tool being automated, purchased from a vendor, and applied uniformly to everyone does not make its exclusionary effect lawful. Uniform application of an inaccessible format is precisely how disability discrimination happens without anyone intending it.

The four-fifths rule gives Dana a concrete tripwire she can run on her own data. If candidates who request an accommodation pass a given stage at less than 80 percent of the rate of candidates who do not, she has a measurable adverse-impact signal worth investigating, regardless of whether anyone intended harm. Intent is not the test. The rate is observable, the comparison is arithmetic, and it turns a question that usually gets answered with reassurance into one that gets answered with a number.

Where AI Creates New Barriers

Dana's audit found that her AI stack introduced four specific accessibility risks her old manual process did not have. The point is not that AI is uniquely hostile to disabled candidates. It is that AI tends to standardize and automate the format of an interaction, and a standardized format is exactly what removes the informal flexibility a recruiter used to provide without thinking about it.

One-Way and AI-Analyzed Video

Tools that score candidates on recorded video often read non-verbal behavior as signal. For a candidate with a stutter, a tic, atypical eye contact, or anxiety on camera, that signal is noise that penalizes a disability rather than measuring job capability. This critique has been consequential enough in the market that at least one major video-interview provider publicly retired its facial-analysis scoring. Dana's rule is simpler than any vendor's response: video is always optional, with an equivalent written or live-conversation alternative offered up front rather than only on request, because requiring a candidate to ask is requiring them to disclose.

Timed Assessments

Many skills platforms impose a countdown. Candidates with processing-speed differences, dyslexia, or ADHD need either extended time or an untimed version, and this is one of the most established accommodations in the entire field. Dana's vendor allowed a per-candidate time multiplier, but it was buried three menus deep and off by default, which meant the accommodation technically existed and practically did not. She turned it into a standard option surfaced at scheduling. A capability nobody can find is not a capability.

Personality and Culture-Fit Inference

Tools that infer personality or cultural fit from writing samples or video are frequently biased against autistic and other neurodivergent candidates whose communication style differs from the training norm. The safest posture is to avoid personality inference altogether where the decision can be made another way, and at minimum to make it optional rather than gating. Dana removed culture-fit inference from her stack entirely. Where a hiring manager wanted a values signal, she replaced it with a structured, job-related question scored against a written rubric, which is both more defensible and more predictive.

Inaccessible Assessment Platforms

A platform can be substantively fair in what it measures and still be unusable with a screen reader, which is a distinct failure mode from bias and requires a distinct check. Evaluate platform accessibility before you adopt, not after a candidate reports a problem. Dana now requires a Voluntary Product Accessibility Template and a WCAG 2.1 AA conformance statement from every assessment vendor before signing, and she tests the candidate-facing flow with a screen reader herself, because a conformance statement describes the vendor's assessment of their product and her own five minutes describes the candidate's actual experience.

Worked Example: Dana's Accommodation Audit

Dana ran one quarter of data through a simple adverse-impact check rather than waiting for another candidate to withdraw. Of 2,400 applicants for non-clinical roles, 168 requested an accommodation. She compared pass-through rates at the assessment stage, which was the stage her instincts pointed at and the stage she had the cleanest data for. The exercise took an afternoon, which is worth stating plainly, because the reason this check goes undone is almost never that it is hard.

Candidates with no accommodation request passed the timed assessment stage at 62 percent, or 1,389 of 2,232. Candidates who requested an accommodation passed at 44 percent, or 74 of 168. The ratio is 44 divided by 62, which is 0.71. That is below the four-fifths threshold of 0.80, a clear adverse-impact flag and, more importantly, a specific number she could take to a vendor and to her own leadership rather than a general concern about fairness.

Digging in, Dana found that 90 of those 168 requests were for extended time, but the extension had only been applied correctly to 51 of them, because the toggle defaulted off and depended on a coordinator remembering to set it. This is the ordinary shape of an accessibility failure: not a refusal, but a default. After making extended time the default for anyone who requested it and adding an untimed alternative, the next quarter's accommodated pass-through rose to 57 percent, lifting the ratio to 0.92, comfortably above the threshold.

The cost was zero in licensing and roughly four hours of coordinator time to fix the default and retrain on the request flow. The return was eight additional qualified candidates advancing who would otherwise have been screened out, in a quarter where Dana had 22 open non-clinical reqs and a 38-day average time-to-fill she badly wanted to shorten. The lesson she drew was not that accessibility is cheap in general, because sometimes it is not. It was that the highest-value fixes are frequently configuration errors sitting inside tools the organization has already paid for.

Building an Accommodation Process That Works

Dana's accommodation process has four properties she insists on, and each one exists because its absence had failed somebody. It is offered proactively, with a plain-language line at every stage stating that alternatives are available and how to ask, so candidates do not have to disclose more than they choose in order to find out whether asking is even possible. The point of the proactive offer is that it shifts the burden of initiating a conversation about disability off the candidate, who is the person in the interaction with the least power and the most to lose.

It is flexible, treating each request as a conversation about what the candidate needs rather than a match against a fixed list of pre-approved accommodations. It is fast, with a named owner and a 48-hour response commitment, so that an accommodation request never silently stalls a candidate out of the funnel while a requisition moves on without them. And it is documented, so the team can show its work if a candidate or a regulator ever asks, and so that patterns in requests feed back into process design instead of being handled one at a time and forgotten. Documentation here serves the candidate as much as the employer.

The last point matters most over time. When Dana saw a cluster of requests for written alternatives to phone screens, she did not just grant them one by one. She made a written asynchronous option a standard path for the screening stage, available to everyone who wanted it. That is universal design: the accommodation, generalized, improved the experience for candidates who never would have asked, including candidates in noisy homes, candidates in different time zones, and candidates who simply write better than they speak on demand. Designing for the edge reliably improves the middle.

Anti-Patterns

Accessible but not really. Dana's careers site passed an automated scanner: screen-reader compatible, keyboard navigable, good contrast. Yet the application was a 47-field form across six screens with no save-and-resume. A candidate with a cognitive disability could technically operate every field and still be defeated by the sheer load. This happens when a team meets the minimum technical requirement without considering actual usability, and the result is a claim of accessibility alongside exclusion in practice. The fix is testing with real users rather than only passing a checker. Dana ran three paid usability sessions with disabled candidates from a local advocacy group and cut the form from 47 fields to 14.

One-size-fits-all accommodation. Many teams offer exactly one accommodation, usually extra time, and nothing else. When a candidate with anxiety asked Dana's team to do the screen by email instead of by phone, the answer was that they do not do that. This happens because accommodations feel risky or inconvenient, so teams limit them to a predetermined menu, and the result is that anyone whose needs fall outside the menu is excluded by a policy that looks accommodating on paper. That is exclusion dressed as policy. The fix is a flexible request process rather than a fixed list.

Accessibility added late. You design the workflow and select the AI tools, then discover afterward that they are not accessible, and now you are retrofitting. Retrofitting is expensive and never fully succeeds, which is why this pattern is worth naming as a design failure rather than a scheduling one. Dana's original stack was chosen on price and speed and then patched for accessibility afterward at roughly three times the cost of building it in. Her new vendor scorecard makes accessibility a gating criterion weighted alongside price and integration, so it can never again be an afterthought.

Practice Prompts

• Audit your application flow. Walk your current application end to end and answer specifically: can someone complete it with a screen reader, is every step keyboard navigable, are the instructions clear, and is there a save-and-resume path? Note where you had to guess rather than verify.
• Inventory your accommodations. List every accommodation you currently offer candidates with disabilities, then list the ones a candidate might plausibly need that you do not offer. The gap between those two lists is your exclusion surface.
• Review your AI tools for accessibility. For each AI assessment tool in your stack, ask whether it captures relevant information fairly across candidates, whether its format penalizes any disability, and whether the vendor can produce a Voluntary Product Accessibility Template and a WCAG conformance statement.
• Design the request process. Write out exactly how a candidate would ask for what they need: where the offer appears, who owns the response, what the response time commitment is, and how the request is recorded without contaminating the evaluation record.
• Run the four-fifths check. Pull one quarter of data and compare pass-through rates at each stage for candidates who requested an accommodation against those who did not. Divide the accommodated rate by the non-accommodated rate and see where it falls relative to 0.80.
• Test from the other side. Move through your own process as a candidate with a disability would, using a screen reader, keyboard only, or under a countdown. Write down every point where something breaks, slows, or requires you to disclose.

Reflection

• Have you or someone close to you ever faced a barrier in a recruiting process because of a disability or a neurodivergent profile? What would have removed it?
• What is the biggest accessibility gap in your current recruiting process, and how would you know if you were wrong about which one is biggest?
• How would you explain to leadership why accessibility matters for business outcomes without letting the business case become the only argument?
• What is one change you could make this month, at low or no cost, that would remove a real barrier?
• Who in your organization owns accessibility in recruiting today, and if the answer is nobody in particular, who should?

Glossary

• Accessibility. The design principle that systems work for people with a wide range of abilities and disabilities, evaluated by whether a person can actually complete the task rather than by whether a standard was nominally met.
• Reasonable accommodation. A change to the process, format, or environment that enables a qualified individual with a disability to participate, required under the ADA unless it imposes undue hardship on the employer.
• ADA (Americans with Disabilities Act). US law requiring equal opportunity in employment, including recruiting. Title I covers the hiring process and the duty of reasonable accommodation.
• WCAG (Web Content Accessibility Guidelines). The technical standard courts and procurement teams treat as the practical benchmark for digital accessibility. Level AA of WCAG 2.1 is the level most commonly required.
• AODA. The Accessibility for Ontarians with Disabilities Act, which makes WCAG conformance an explicit statutory requirement for many employers operating in Ontario.
• Neurodiversity. The concept that cognitive differences such as ADHD, autism, and dyslexia are natural variations rather than deficits.
• Assistive technology. Tools such as screen readers, voice control, switch devices, and magnification software that candidates use to interact with your systems. Your process must work with them, not merely alongside them.
• VPAT (Voluntary Product Accessibility Template). A vendor-produced document describing how a product conforms to accessibility standards. Useful as evidence of effort, not as a substitute for testing the candidate-facing flow yourself.
• Four-fifths rule. An adverse-impact screen: if one group passes a stage at less than 80 percent of the rate of the comparison group, that disparity warrants investigation regardless of intent.
• Universal design. Designing the standard path so that it works for the widest range of people, which turns many individual accommodations into features available to everyone.

Related Lessons

• Bias and Fairness Risks in AI-Assisted Recruiting for the broader fairness picture that disability adverse impact sits inside.
• Auditing AI-Assisted Decisions: Sampling Methodology and Fairness Metrics for the measurement methods behind the four-fifths check.
• Transparency and Disclosure: Telling Candidates About AI Use for how the proactive offer of an alternative fits into candidate notice.
• Understanding Candidate Experience: What Matters in Hiring for the experience frame this lesson applies to disabled candidates specifically.
• DEI and Culture Alignment: Using AI to Advance Inclusion for how accessibility connects to wider inclusion goals.

Closing

Inclusive recruiting starts with accessible recruiting, and accessible recruiting starts with looking. Dana's whole turnaround began with one withdrawn candidate and a follow-up question she did not have to ask. The candidates your process excludes will almost never tell you; they will simply stop, and the funnel will show nothing but a number that looks normal. That is the argument for auditing on purpose, testing with real users, defaulting the accommodation on instead of off, and treating accessibility as a gating criterion in procurement rather than a patch applied later. Make it a core principle, and the process improves for everyone who touches it.

Key Takeaways

• Accessibility is a talent-pool decision as well as a legal duty. Roughly one in four working-age adults reports a disability, and a process that excludes them shrinks your funnel silently, without ever showing up as a rejection.
• The ADA sets the floor. Title I prohibits discrimination against qualified individuals with disabilities in hiring and requires reasonable accommodation unless it imposes undue hardship. This is a legal requirement, not a best practice.
• Treat WCAG 2.1 Level AA as your technical benchmark. The ADA names no standard, but courts and the Department of Justice have pointed to WCAG, and AODA makes conformance an explicit statutory requirement for many employers in Ontario.
• The EEOC has warned that algorithmic tools can violate the ADA. Its 2022 guidance addresses tools that screen out people who could do the job with reasonable accommodation, and disproportionate screening of neurodivergent candidates can draw scrutiny under the ADA and disparate-impact analysis.
• AI introduces four specific new barriers. One-way and AI-analyzed video, timed assessments, personality and culture-fit inference, and inaccessible assessment platforms. Make video optional with an equivalent alternative offered up front, default extended time on, drop culture-fit inference, and evaluate platform accessibility before adopting.
• Run the four-fifths check on accommodated candidates. If they pass a stage at less than 80 percent of the rate of non-accommodated candidates, you have a measurable adverse-impact signal. Dana's 0.71 ratio traced back to a default-off toggle and was fixed at zero licensing cost.
• Test with real users, not just scanners. A 47-field form can pass an automated checker and still defeat a candidate through sheer complexity. Paid usability sessions with disabled candidates find what scanners cannot.
• Make accommodation proactive, flexible, fast, and documented. Offer alternatives at every stage so candidates need not disclose to ask, treat each request as a conversation rather than a menu, name an owner with a 48-hour response commitment, and record requests separately from evaluation data.
• Build accessibility in from the start. Retrofitting is expensive and never fully succeeds. Make accessibility a gating criterion in vendor selection, weighted alongside price and integration.
• Generalize accommodations into universal design. When a cluster of candidates asks for the same alternative, make it a standard path for everyone. Designing for the edge improves the experience for all.

Frequently Asked Questions

Can we ask candidates whether they have a disability so we can plan accommodations?

Ask about accommodation needs, not about disability. The practical and lawful pattern is to state at each stage that alternative formats are available and to give a simple way to request one, without requiring the candidate to name a diagnosis or provide documentation to start the conversation. That framing keeps the exchange focused on what the person needs to participate rather than on protected medical information, and it means a candidate can ask for a written alternative to a phone screen without disclosing anything about their health at all.

Does making video interviews optional put us at a disadvantage in evaluating candidates?

Only if the video was measuring something job-related, which is the question worth testing. Tools that score non-verbal behavior on recorded video read a stutter, a tic, atypical eye contact, or on-camera anxiety as signal when it is noise attached to a disability. Offering an equivalent written or live-conversation path up front costs you nothing in signal you should have been relying on, and it removes a barrier that was screening out candidates who could do the job. Structured, job-related questions scored against a rubric outperform format-based inference in any case.

Where should accommodation requests be recorded?

Separately from the evaluation record, with restricted access. An accommodation request is a disability disclosure, and under the ADA that information is protected and must be kept apart from hiring decisions. Storing it in the same notes field as interview scores puts protected information in front of the people making the decision, which is both a privacy failure and a discrimination risk. Keep a documented log for compliance and for spotting patterns, but hold it in a channel that decision-makers do not see.

Is a vendor's accessibility conformance statement enough?

It is evidence, not assurance. A Voluntary Product Accessibility Template and a WCAG 2.1 AA conformance statement should be a gating requirement before you sign, because a vendor unwilling to produce either has told you something useful. But a conformance statement describes the vendor's own assessment of their product. Test the candidate-facing flow yourself with a screen reader and a keyboard, and, where you can, with real disabled users. A platform can be substantively fair in what it measures and still be unusable with assistive technology.

What if extended time makes the assessment unfair to other candidates?

Extended time for a candidate with a processing-speed difference, dyslexia, or ADHD does not confer an advantage; it removes a penalty unrelated to the job. Under the ADA, reasonable accommodation is required unless it imposes undue hardship, and a time multiplier that a platform already supports is difficult to characterize as hardship. If the countdown itself is not measuring anything the role requires, the better answer is often to offer an untimed alternative to everyone, which is universal design rather than exception handling.