Automatic captions in video calls, text read aloud on request, or voice control for work software: artificial intelligence can genuinely reduce barriers in everyday professional life. However, the benefit does not arise from a company activating as many AI features as possible. It arises when a specific obstacle is identified, an appropriate aid is selected, and the affected person is involved from the outset.
This guide shows how AI and accessibility in the workplace in Austria can meaningfully work together. It is aimed at employees, applicants, managers, HR, works councils and IT managers. The focus is on practical use cases, a safe pilot rollout, funding opportunities and the question of when human support is more important than automation.
The most important things in 60 seconds
- The starting point is always the specific barrier, not the newest AI tool.
- Real-time captions, read-aloud functions, text simplification, image descriptions and voice control can make everyday work easier.
- AI outputs remain error-prone. Dialects, technical terms, names and numbers must be checked particularly carefully.
- Health data and information about a disability should not be submitted to public AI services without verification.
- Accessible basic design remains mandatory: an AI add-on does not replace accessible software, suitable work equipment, or personal assistance.
- A small, measurable pilot is better than a company-wide rollout without a clear need.
- In Austria, workplace adaptations, technical aids, and training may be eligible for funding; the assessment should begin before procurement.
Accessibility begins with the barrier, not with the tool
The same AI feature can be very helpful for one person and irrelevant for another. An automatic summary is of little use if the underlying conversation was incomprehensible because of poor captions. An image description only helps if the software is also operable with a keyboard and screen reader. And voice control offers no benefit if it consistently fails in a noisy open-plan office.
Therefore, the first question should be:Which specific task is unnecessarily difficult today or not possible to perform independently? Good answers are observable, such as "reliable subtitles are missing in the weekly video call," "scanned documents cannot be read aloud," or "frequent typing causes pain." Only then does the selection of a solution follow.
Accessibility is also not a state that a single piece of software automatically confirms. The Web Accessibility Initiative of the W3C emphasizes that tools support testing, but an informed human assessment remains necessary. For the workplace, this means: automated tests and AI aids supplement the feedback of the affected person; they do not replace it.
Five areas of application where assistive AI can help
1. Making language visible
Real-time subtitles and speech-to-text can make meetings, training sessions, and phone calls more accessible for deaf and hard-of-hearing people. People who do not use German as their first language also sometimes benefit from an additional text layer. However, quality and the ability to make corrections are crucial. Austrian dialects, similar-sounding names, or industry-specific abbreviations quickly lead to incorrect transcripts.
For important conversations, it should be clarified in advance whether an automatic transcript is sufficient or whether professional sign language interpreting or speech-to-text interpreting is required. In the case of safety briefings, personnel interviews, or binding decisions, an unnoticed mistranslation must not become the sole source of information. The rules for recording, storage, and deletion also apply just as they do for other AI note-takers; the jobspot.at guide is relevant here.When AI listens in on a meeting.
2. Reading aloud, explaining, and structuring texts
Text-to-speech can read out emails, instructions, and documents. AI can also outline long paragraphs, explain difficult phrasing, or convert work instructions into clear, individual steps. This can support people with visual impairments, dyslexia, learning difficulties, concentration problems, or temporary exhaustion.
However, simplification must not change the content. "Can be done after approval" must not become "must be done immediately." Legal texts, technical limits, dosages, deadlines, and safety rules therefore require a comparison with the original. For frequently used documents, it is better to create a well-structured and accessible original version from the start, rather than preceding it with an AI repair every time.
3. Describing images and interfaces
AI can create drafts for alternative texts, roughly describe diagrams, or capture the content of an inaccessible scan. This saves time if a human subsequently checks the description. For complex graphics, however, the purpose is decisive: a blind colleague does not need every color in a sales chart, but rather the message, time period, axes, and notable trends. A generic image description can miss this work context.
Here, too, the better way is an accessible source. Semantic headings, real tables, labeled form fields, and keyboard operability help more reliably than a subsequent AI layer. The W3C provides guidance on text alternatives ready, which is just as useful for websites as it is for internal documents.
4. Facilitating input and operation
Speech recognition, word prediction, intelligent correction, and customizable commands can relieve the burden on people with motor impairments. An employee can dictate a report, trigger recurring steps via voice, or generate an initial email draft from keywords. However, the productivity gain should not be confused with a performance target: if an aid makes work possible independently in the first place, it must not automatically lead to higher target figures.
Before implementation, the environment and fallback level are important. Does the speech recognition work with the individual's way of speaking? Is there a quiet room? Can critical commands be confirmed or undone? Does the application remain operable without AI if the service fails?
5. Preparing communication
AI can organize meeting notes, formulate potential follow-up questions, or bring a text into a clearer tone. This can support neurodivergent employees and people with communication barriers. A helpful use would be, for example, to derive three core questions from an agenda before a customer meeting. It becomes problematic when the system is intended to diagnose emotions, resilience, or personality. Such attributions are unreliable and can be discriminatory.
The International Labour Organization therefore describes AI as a double-edged tool: assistive functions such as text-to-speech and subtitles can facilitate employment, while automated selection and evaluation systems can disadvantage people with non-standard speech, facial expressions, or movement.
Where AI creates new barriers and risks
Errors do not affect everyone equally
A misrecognized technical term is just an annoyance for some. For a person who relies entirely on the transcript, it can change the central message. Companies should therefore not only measure average accuracy, but also critical errors: names, numbers, negations, warnings, responsibilities, and deadlines. A solution is only practical if it has been tested in the real Austrian linguistic and work context.
Confidential information needs protection
The search for a suitable adaptation can involve sensitive health and employee data. This information does not belong in just any public chatbot. Before a pilot project, data flow, provider, storage location, access rights, deletion, and potential model usage must be clarified. Where possible, test data should be anonymized or synthetic. For the secure handling of non-approved services, the article Shadow AI at work offers an additional assessment framework.
Automated personnel selection can discriminate
AI that sorts applications, evaluates video interviews, or assesses employees is not an assistive aid for the person concerned. Such systems can interpret disability-related differences as supposedly poor performance. The European AI Act classifies certain AI systems in employment and human resources management as high-risk and explicitly highlights the risk of historical discrimination, including on the grounds of disability.
For HR, this results in a clear distinction: an individually requested tool to support a task is something different from a system that classifies people. Accessible alternatives, human review, and a comprehensible appeals process must be planned before implementation.
Austria: Clarify workplace adaptation and counseling early on
Assistive AI is rarely the entire solution. It is often part of a package consisting of an ergonomic workplace, adapted software, input devices, training, and organizational change. The Social Ministry Service mentions, among other things, technical work aids, software, workplace adaptation, and training as possible areas of support. The Chamber of Labour also refers to subsidies for technical aids and disability-friendly adjustments.
The order is important: document the need, seek advice, check funding requirements, and only then place the order. Anyone who buys an annual subscription first may miss out on potential funding or a more suitable overall solution. Employees and companies can also use the Work Assistance in the Vocational Assistance Network as a point of contact. The jobspot.at guide Use work assistance explains access from the perspective of job search and employment.
Funding eligibility and responsibility depend on the individual case. An AI license is not automatically eligible for funding just because it reduces a barrier. What matters are individual necessity, specific implementation, and the applicable requirements.
In seven steps to a robust pilot project
1. Precisely describe the activity and barrier
Note the situation, frequency and impact. Example: "In two one-hour video conferences per week, decisions are lost because the automatic captioning incorrectly recognizes technical terms." This description is more measurable than "We need better accessibility".
2. The person concerned decides along with
No solutions over the person's head. Clarify desired help, previous strategies, data protection concerns, and an acceptable alternative. The person does not have to disclose their entire diagnosis to describe a functional hurdle.
3. Also check non-technical alternatives
Perhaps a clearer agenda, a quiet workplace, more processing time, or an accessible template will solve the problem more reliably. AI is useful if it provides an additional, demonstrable benefit.
4. Clarify data protection, IT security, and co-determination
What data does the system process? Who has access? How long are inputs and outputs stored? Is a works agreement or the involvement of the works council required? The answers belong in a short pilot approval before the first real data set.
5. Test with a limited scope
Choose one activity, a few weeks, and clearly named participants. Test typical dialects, technical terms, file formats, and failure situations. For transparent AI outputs, information obligations also apply: The RTR AI Service Center weist darauf hin, dass Informationen in klarer, erkennbarer und barrierefreier Form bereitgestellt werden müssen.
6. Measure success and harm separately
Suitable metrics are time required, steps completed independently, number of critical corrections, the burden and the affected person's satisfaction. Also document malfunctions, data protection issues and situations in which human assistance was necessary.
7. Decide, adjust, or stop
After the pilot there are three legitimate outcomes: rollout with clear rules, adjustment and a new test, or termination. A previously paid subscription is not a reason to keep an unsuitable aid. Also record who will re-check, after updates, whether data flow or functionality have changed.
Decision aid for typical situations
- Video call with familiar colleagues: Test real-time captions, prepare a glossary of technical terms, and define error correction and an emergency fallback.
- Safety briefing: AI only as a supplement; ensure accessible original materials and demonstrable understanding.
- Confidential personnel meeting: No spontaneous cloud transcription; clarify human support and data protection in advance.
- Hard-to-access PDF: AI can provide a working draft; the source should subsequently be recreated to be structurally accessible.
- Painful keyboard input: Test dictation and voice commands in the real environment, ensure undo function and a manual alternative.
- Automated applicant ranking: Do not treat as an accessibility measure; assess risks, alternatives and the human decision separately.
Procurement check: Ten questions before a license or contract
- Which specific barrier should the function reduce?
- Has the affected person tested the solution themselves?
- Does it work with screen readers, keyboard, magnification and existing assistive tools?
- How well does it handle Austrian dialect and the company's technical terminology?
- Which personal or confidential data is transmitted to the provider?
- Can training on company data and unnecessary storage be excluded?
- How are errors detected, corrected, and documented?
- What accessible alternative is available in case of an outage or refusal?
- Are consulting, technical assistance, or training eligible for funding?
- Who reviews the solution after updates and at the latest by the agreed deadline?
FAQ on AI and accessibility in the workplace
Must a disability be disclosed in order to use AI assistance?
Not every helpful feature requires a diagnosis. However, for individual support measures or formal workplace adaptations, documentation and a defined procedure may be necessary. In the workplace, only the information that is actually required for assessing needs, safety, and processing should be collected.
Can an employer designate automatic captions as a substitute for interpreting?
Not categorically. Whether captions are sufficient depends on the task, the conversation, the required accuracy, and individual needs. For important or complex content, professional assistance may be necessary. The affected person must be involved in the assessment.
Is an AI-generated alt text automatically accessible?
No. It is a draft. A human must check whether the description is correct and conveys the purpose of the image in the specific context. Purely decorative images often do not require a content description at all.
Who is responsible if the AI makes a mistake?
A company cannot delegate responsibility to a tool. Responsibilities for review, approval, correction, and escalation must be defined before deployment. For safety-critical or legally relevant content, human final review is indispensable.
What is the best first step for a small business?
Describe a specific barrier with the affected person and contact an advisory service before purchasing a product. Afterwards, a small test with clear success criteria can show whether AI truly helps.
Conclusion: Assistive AI must increase independence
Good assistive AI makes a specific task more accessible without creating new dependencies or unnecessary data traces. It gives the affected person more control and complements an accessible baseline design. Poor implementation, by contrast, starts with a license, ignores individual needs, and shifts quality control onto the person who is already struggling with the barrier.
The practical path for Austrian businesses is therefore clear: name the barrier, involve the person, check advisory services and funding, protect data, test on a small scale and make measurable decisions. This way AI does not become a label for inclusion, but a tool that is useful in real everyday work.
