AI widens the door to work for people with disabilities. Employers decide how far

In 2025, 8.3 percent of people with a disability in the US labor force were unemployed, against 4.1 percent of people without one. That gap is a loss of talent employers can see in every open role they struggle to fill.

AI is lowering some of the daily barriers behind it. The decisions that open or close the door still belong to employers.

The tools already remove daily friction

Some of the most useful AI at work was built for people with disabilities first. Microsoft’s free Seeing AI app narrates text, documents, and scenes for blind and low-vision users.

Be My Eyes offers Be My AI, which describes images on request. The company notes that human volunteers are on hand for tougher requests. That pairing is the right model: AI for speed, a person when it matters.

General workplace assistants help too. When the UK Department for Business and Trade piloted Microsoft 365 Copilot↗ with 1,000 licenses in late 2024, neurodiverse staff were significantly more satisfied than other users. Staff reporting any health condition or disability were more likely to recommend it.

In interviews, one employee with ADHD said the tool had “levelled the playing field”. An employee with a hearing disability said it let them relax in meetings without fear of missing key details.

The same technology can close the door

AI that screens résumés learns from human decisions, including biased ones. University of Washington researchers asked GPT-4 to rank a résumé against the same résumé with disability-related awards added. The model ranked the enhanced résumés lower and justified it with ableist reasoning.

Written instructions reduced the bias in that study. They did not remove it.

The legal standard is clear. The US Department of Justice’s guidance says employers violate the ADA if hiring technology unfairly screens out a qualified person with a disability. The guidance names voice and facial analysis, and online interviews a screen reader can’t use, as examples.

Access is part of the tool

An AI assistant only helps if everyone can use it. A chat window that a screen reader can’t navigate, or a voice-only interface, excludes the people who might gain the most.

WCAG 2.2, the W3C’s accessibility standard, is the yardstick to ask vendors about. Ask for conformance evidence, then test with the people who use assistive technology every day.

Disability information stays private

Employees often explain a disability to get the right tool. Under EEOC guidance, medical information an employee volunteers must be treated as a confidential medical record.

That rule applies to AI too. Accommodation notes should not flow into prompts, shared transcripts, or analytics that managers browse. Decide where that information lives, and who can see it, before a single license goes out.

Why this matters to Ludia

Lucas Diaz, Ludia’s founder, serves on the Board of Directors of Relay Resources in Portland, Oregon. Relay began in 1951, when parents of children with intellectual and developmental disabilities refused to place them in institutions.

Today Relay employs nearly 900 team members across four lines of business. Its abilIT program, run with Melwood, prepares people with disabilities for careers in IT and cybersecurity. Its stated vision is plain: “All disabled people belong. Everywhere.”

Relay’s work is Relay’s. It shapes Ludia’s view of AI: the technology does the legwork, and people make the calls. That stance runs through Ludia’s own AI work and our Dynamics for Good program.

Where the evidence is still thin

The strongest workplace evidence so far is early and self-reported. The UK pilot covered one department for three months, and its authors describe it as an evaluation without a counterfactual.

Its vision and hearing findings came from interviews; the survey tests for those groups were not significant.

The same study found Copilot helped most with written tasks such as drafting and summarizing. For hands-on field and facilities work, there is not yet much published evidence.

One more caution from that pilot: when licenses were removed, readjusting was harder for users with disabilities. A tool that becomes an accommodation should not disappear with a budget cycle.

What leaders should do now

Test every hiring tool for accuracy and for screen-out. Ask the vendor how it was tested with disabled applicants. Offer a clear accommodation path, and have a person review automated rejections.

Put disabled employees in the pilot group. They find the barriers first, and they often gain the most.

Buy accessible AI, and prove it. Ask for WCAG 2.2 conformance evidence, then test with screen readers, keyboard-only use, and captions.

Keep disability information out of prompts and dashboards. Treat it as a confidential medical record, with named owners and access rules.

Treat a working tool as an accommodation. If someone relies on it to do their job, don’t let the license lapse at the end of a pilot.

How we know this

Unemployment figures: US Bureau of Labor Statistics, Persons with a Disability: Labor Force Characteristics, 2025 (March 3, 2026). Seeing AI: Microsoft, Seeing AI. Be My AI: Be My Eyes, Be My AI. Copilot pilot: UK Department for Business and Trade, The Evaluation of the M365 Copilot Pilot (August 2025). Résumé study: Glazko, Mohammed, Kosa, Potluri and Mankoff, Identifying and Improving Disability Bias in GPT-Based Resume Screening, ACM FAccT 2024. Hiring guidance: US Department of Justice, Algorithms, Artificial Intelligence, and Disability Discrimination in Hiring (May 12, 2022). Accessibility standard: W3C, Web Content Accessibility Guidelines (WCAG) 2.2. Confidentiality: US EEOC, Enforcement Guidance on Disability-Related Inquiries and Medical Examinations of Employees. Relay Resources: About us and abilIT, relayresources.org.