Include disabled and neurodivergent employees in choosing, testing and changing the AI-supported work they are expected to do. Keep effective existing support available, and judge whether the person can complete the whole task, including checking and sharing the result. A promising feature can still sit inside an unusable workflow.
For an adoption lead or manager, the starting point is the employee's experience of a particular task. People have different preferences and access needs; disability and neurodivergence are not interchangeable identities. Ask what helps the person work rather than assigning a tool or training format from a diagnosis. The broader people and change guide explains how to arrange support around an actual attempt.
Start with the work the person wants to finish
Invite the employee to choose a recurring piece of work and describe where it becomes difficult. They may want help turning spoken ideas into a draft, following a discussion, finding information or organizing a set of instructions. They may also prefer a familiar non-AI method. Establish the required outcome before offering a feature.
A May 2026 study of workplace needs among 20 autistic adults in Singapore found concerns about unclear expectations, communication and workplace understanding. An autistic researcher helped shape the workshops and interpretation. The paper analyzed focus-group discussions, including desired AI support; it did not establish that an AI intervention improved work. Its local, spoken-English workshop format also limits whose experience it captures.
That distinction matters when designing a trial. An assistant might reorganize a vague instruction, but the manager still needs to explain what the work requires. Do not make the employee infer a deadline or an acceptance standard that the team has never agreed.
Ask these questions separately:
- What should the finished work contain? Name the output and the person who relies on it.
- Where does the current method become difficult? Listen for a specific barrier rather than assuming lack of skill.
- What support already works? Include assistive technology, accessible documents, colleague support and working arrangements.
- What would the employee like to try? Agree the format, pace and way to give feedback, including a private conversation when needed.
A written response, a demonstration and a conversation can reveal different things. Offer a workable way to participate rather than making a fast group discussion the only route into the trial.
Useful assistance still needs checking
Laura H. described difficulty turning incomplete meeting notes into a clear record. In her September 2026 post, she explained that she now uses transcripts and AI to draft decisions, open items and owners. She also corrects what the draft gets wrong.
But editing a draft is a different job to writing one from nothing.
Laura described her own experience of dyslexia and meeting records. This is an individual account, not a measured comparison of tools or a result for every employee.
That is a useful possibility to investigate. The employee's expertise remains visible in deciding what the record should say and correcting it. A trial should capture both the help and the effort required to make the result dependable.
There is broader, qualified evidence of useful experiences. EY's 2024 Microsoft-commissioned Copilot report describes employees valuing help with writing, organizing thoughts and reviewing meeting information. Participants were existing Copilot users recruited through their organizations, and the study used self-reported experiences. It supports exploring these uses; it does not establish that the same tool will suit a new employee or task.
The surrounding team can still create a barrier. In a May 2026 Reddit discussion, a Reddit user described repeatedly revising AI-assisted writing to preserve their own meaning, while a coworker rejected the work because AI had been used. In a follow-up, they explained that a previous colleague had helped turn their technical thinking into documents. They had not yet resolved the current situation with their manager. The employer's rules and the work itself were not available to inspect.
A manager needs to examine an actual concern about accuracy, permitted inputs or the handoff. Dismissing an employee's work without identifying the problem leaves both the work standard and the support question unsettled.
Test the complete workflow
Check access beyond the prompt box. Can the employee find the source, tell when the answer is ready, inspect what changed, correct an error and pass the result to a colleague? Include the assistive tools and account settings the person actually uses.
In an April 2026 X discussion about screen-reader access, Devin Prater identified response-ready feedback and headings as meaningful improvements. Replies also discussed changed reading behavior. These dated reports do not certify current product compatibility. They give a concrete testing question: does the person know an answer is available, and can they navigate it?
The September 2026 CACM article on accessible work provides a contrasting account. In a workshop of 44 invited experts, an anonymous blind employee described joining a team that used an AI-enhanced digital whiteboard. Screen-reader editing and keeping track of the shared work created burdens despite the tool's promised benefits. This was a qualitative workshop account, not a product benchmark or estimate of how often the problem occurs.
Use the employee's actual task to look for these failures:
| Part of the task | What to observe | What a failure means for the trial |
|---|---|---|
Open and control the tool | The person can enter instructions, navigate controls and recover from an error using their normal setup. | The workflow needs an accessible route before routine use. |
Read the source and answer | The original material and generated response can both be found and navigated. | A fluent answer is insufficient if the person cannot compare it with the source. |
Correct and verify the output | The person can identify and change a wrong date, omitted condition or altered meaning. | Arrange a usable checking method; do not hide the extra work. |
Share the finished work | The recipient receives an accessible, accurate document in the agreed format. | Investigate the handoff as well as the assistant. |
An accessibility barrier needs someone with authority to fix it. The UK's Business Disability Forum guidance recommends involving disabled people throughout the AI lifecycle and maintaining compatibility with existing assistive technology. Translate that into a named technical or workflow owner and a follow-up date. Repeating a demonstration cannot fix a control the employee cannot reach.
Let the employee shape one task trial
Suppose Alex Morgan, a publisher's production editor, prepares an internal brief for a copyeditor from approved author notes and the publication schedule. The brief needs to preserve the manuscript scope, unanswered author questions and agreed deadline. Alex wants to try dictating rough notes and asking an approved assistant to organize them.
Keep the existing brief-writing method available. Alex chooses how to participate and stays in control of the work while a colleague helps investigate the workflow.
- Agree the task and boundaries. Identify the approved tool and inputs, the copyeditor's required brief and the support that remains available. Use a permitted sample if the real material cannot enter the tool.
- Follow the attempt through checking. Alex compares the generated deadline with the publication schedule and checks that an unanswered author question remains open. Observe whether the source, draft and editing controls are usable.
- Ask about the full effort. Did dictation or restructuring help? What became harder? Include correction, fatigue, interruptions and help from others, using the employee's own account rather than guessing.
- Agree the next arrangement. Keep useful assistance, change the workflow or return to the existing method. Name who will resolve a barrier and review the next brief after the change.
Keep a short record of the task, observed barrier, proposed change, owner and review point. Record relevant work observations, not a diagnosis in a shared team note. A useful next trial might change the input format or checking route; it need not involve a different model.
Keep support and private needs separate from adoption targets
Do not remove effective support on the strength of an AI demonstration. A transcript may help an employee review a meeting afterward while leaving the live discussion difficult to follow. Clear agendas, accessible materials and a workable way to contribute still need attention.
Where an employee requests a formal adjustment, use the organization's confidential accommodation process and the responsible specialists. A general adoption trial cannot settle an individual's needs or the legal requirements that apply. Avoid putting medical histories or diagnostic details into shared prompts or demonstration materials.
Review the finished work and the conditions needed to produce it. Include the employee's account of whether the arrangement is sustainable. Do not treat declining an unsuitable tool, reporting a barrier or needing continued support as a failed adoption score. The guide to performance systems that support useful AI adoption explains why judgment and output quality need attention alongside use.
Questions and answers
Can AI help every disabled or neurodivergent employee?
AI may help a particular employee with writing, organizing information or reviewing a discussion, but a diagnosis does not predict whether a tool will be useful. The person's preferences, assistive setup, task and ability to check the result all matter. Existing-user reports of benefits are a reason to investigate, not a guarantee.
Invite the employee to choose a recurring task and an acceptable way to try it. Compare the complete attempt with their current method, including correction and effort. The publisher's task trial shows what to agree and observe.
How should a company test an AI workflow for accessibility?
Test an AI workflow with the employee's actual setup from entry through the final handoff. Observe whether the person can control the tool, find the source, navigate the answer, correct errors and share usable work. Technical accessibility checks and the employee's experience answer related questions; neither a vendor statement nor a polished demonstration establishes that the whole task works.
For a publication brief, the employee must be able to compare the generated deadline with the schedule and preserve an unanswered author question. Give failures a fix owner and review the changed workflow. Use the complete-task checks to structure that observation.
Should AI replace existing workplace adjustments?
An AI trial should not remove an effective adjustment simply because a feature appears to offer similar help. The trial may reveal useful additional support, extra checking or a new barrier. Keep the existing route available while evaluating the person's complete work and experience.
Any proposed change to a formal adjustment belongs in the appropriate confidential process with the employee and responsible specialists. Requirements depend on the individual's circumstances and the applicable jurisdiction. Separate support decisions from adoption targets before changing the arrangement.
Does an employee need to share a diagnosis in an AI trial?
A task-focused AI trial can discuss the barrier, useful support and preferred format without placing a diagnosis in a shared prompt or team record. Ask what the employee needs to finish the work and how they want to give feedback. Do not infer a medical condition from difficulty using a tool.
If a formal accommodation process needs personal information, route that discussion through the organization's confidential process and responsible specialists. The trial record should contain the work observation, change and owner needed to act, with access limited appropriately. Start with the task questions.
