A federal agency is about to buy an artificial-intelligence assistant. It will answer questions about benefits, help a veteran find a form, route a worker through human resources or guide a resident through a permit application. The demonstration is smooth. The assistant hears a question, answers promptly and never gets tired.
Then a deaf person tries to use it. Federal law requires comparable access to the information and services such systems provide.
The usual institutional answer is familiar: File a request. Ask for an interpreter. Send an email. Call a help desk, somehow. The agency has converted a service that is immediate for one group into a case-management problem for another — and then calls the complaint channel an accommodation. Section 508 does provide an administrative complaint route; that remedy belongs in the enforcement architecture, not in place of usable technology.
That is the accommodation trap. It mistakes a remedy for a product.
Federal disability law already points in the other direction. Section 508 of the Rehabilitation Act requires federal agencies, when they develop, procure, maintain or use information and communication technology, to provide employees with disabilities and members of the public access comparable to that available to others, absent undue burden. It does not say that comparable access begins after someone has been shut out.
Congress also made Section 508 adaptable. The Access Board must periodically review and, when appropriate, amend the standards to reflect changes in technology; the acquisition rules must then incorporate those changes. That is not an antique obligation being awkwardly applied to a new machine. It is a legal design: technological change should update the means of access, not narrow the promise of it.
The procurement rules are equally plain. The Federal Acquisition Regulation requires agencies acquiring information and communication technology to ensure comparable access and, barring a documented exception or exemption, to meet the applicable accessibility standards. When a conforming commercial product is unavailable, the rule requires the option that best meets the standards and a record of the market research, unmet requirements and rationale.
What is missing is not authority. It is a buying rule for services whose very interface is language.
A chatbot that works only through speech is not made accessible because it can send a transcript after the fact. An automated intake system is not comparable because a user may request a human being after the bot has misunderstood the question. And a video agent does not become usable for deaf and hard-of-hearing people because a form at the bottom of the page invites them to report the failure. Comparable access is the legal benchmark, not the availability of a later grievance.
For many deaf people, signed language is not a decorative add-on to a spoken exchange. It is the route into the exchange. For others, the essential pathway may be reliable text, captioning, visual interaction or tactile feedback. Those needs are not interchangeable, and “deaf and hard of hearing” is not a single test population. The government’s own Section 508 guidance says not to convey information only through audio or voice, calls for alternative communication formats, and tells agencies to identify how users with disabilities will perform the technology’s functions.
Procurement officers should therefore require product evidence before deployment. Not a marketing statement. Not a promise to improve after launch. Evidence.
First, every AI-mediated service that accepts or produces speech should have a signed-language pathway when the service’s users need one. That pathway must let a person complete the same consequential task — not merely watch a translated welcome screen. The relevant test is task completion: Can the person ask, understand, correct, consent, receive a decision and preserve a record on terms comparable to a hearing user’s? That translates the government’s user-needs requirement into an acceptance test for a language-mediated service.
Second, require a visual and haptic twin for every essential voice interaction. If the AI asks a question, signals an error, requests confirmation or gives a deadline by sound, the same information and control must exist through sight and, where appropriate, touch. The government’s guidance already directs agencies not to convey information only by audio or voice and identifies haptic feedback as an alternative means of operating controls.
Third, test with deaf and hard-of-hearing people as distinct user groups, and report the results separately. A single aggregate score can conceal the failure of a signer, a caption-dependent user or a person navigating a noisy service counter. The agency should publish the tasks tested, success and abandonment rates, the types of communication used, and the material gaps that remain. The law now expressly calls for assessment through consultation with people with auditory, tactile and other disabilities, alongside technical conformance testing.
Fourth, put a human fallback inside the service, not behind a complaint form. It should be reachable without voice, available at the point of failure and authorized to finish the transaction — not simply to apologize and open a ticket. The rules retain alternative means of access where an exemption applies; that makes a fallback necessary, not proof that the primary product works.
Finally, make failure modes contract data. Vendors should document where speech recognition fails, where language generation becomes ambiguous, where captioning lags, where a signing pathway breaks and how the system hands off to a person. Updates should trigger retesting. Section 508 procurement guidance already says accessibility belongs at the beginning of the lifecycle, calls for accessibility terms and acceptance criteria in solicitations, and directs agencies to validate claims with testing before award or deployment.
This approach is not a demand that Washington freeze useful technology until it is perfect. It is the ordinary discipline of buying a service that works. Section 508 already frames the procurement task as selecting a solution that meets a well-defined business need while remaining accessible to employees and the public.
The cost argument also has the wrong time horizon. A complaint process shifts the cost of a design failure onto the person excluded: the extra wait, repeated explanation, lost appointment or missed deadline. Pre-deployment testing shifts it back to the moment when the agency still has leverage — before the contract is awarded, the workflow is entrenched and the vendor’s promises harden into someone else’s inconvenience. Federal guidance similarly warns that omitting technical requirements creates remediation and schedule or cost risk after delivery and acceptance.
Accommodation remains a vital residual duty. People will always have individual needs that no general product test can foresee. But residual is the point. It is the backstop after a service has been designed for access, not the front door through which people must beg to enter. That ordering reflects the statute’s comparable-access rule and its separate provision for alternative means when compliance would impose an undue burden.
Call this the buying rule. It is the procurement face of the larger doctrine: equal access in an AI-mediated public service has to be built, tested and purchased before deployment. A complaint process may repair a wrong. It cannot be the product.
%402x.svg){kind=icon}
