Generative Artificial Intelligence Supporting Functional Communication and Return to Work in a Person With Wernicke's Aphasia: A Case Study

Abstract

Background. Generative artificial intelligence (AI) is increasingly explored as a support for people with aphasia. Large language models (LLMs) have been used to reconstruct impaired utterances in conversation (Adikari et al., 2025) and to support written expression at the single-case level (Allen et al., 2025), and reviews highlight their potential for assessment and intervention (Zhong, 2024) and for linguistically diverse populations (Imaezue et al., 2025). Emerging frameworks also address the ethical use of generative AI in aphasia management, including its role within augmentative and alternative communication (AAC) and digital accessibility (Bailey et al., 2026). However, evidence on the integration of mainstream generative AI tools into real-life functional contexts, particularly vocational reintegration, remains scarce.

Aims. To describe the implementation of mainstream generative AI tools as compensatory communication supports, within a person-centred, participation-focused speech-language therapy plan aimed at return to work.

Methods. Case study of a 53-year-old woman with Wernicke’s aphasia secondary to meningioma resection with anterior temporal lobectomy (2024), who pre-morbidly held coordination and vice-presidency roles in the social sector. From January 2026, speech-language therapy (50 minutes, twice weekly) focused on vocational reintegration through building and teaching AI-based supports; her initial AI literacy was null. Baseline language assessment (Lisbon Aphasia Assessment Battery; Aachen Aphasia Test, Portuguese version) showed fluent discourse with phonological paraphasias, anomia, impaired auditory and reading comprehension for complex material, and writing marked by lexical, morphosyntactic and agreement errors; neuropsychological assessment showed preserved cognition. Functional communication was measured with the ASHA Functional Assessment of Communication Skills (ASHA-FACS) at baseline (January) and re-assessment (June), complemented by a purpose-built, non-validated self-report scale of perceived usefulness and communicative participation (18 items, five dimensions, 5-point Likert). Supports implemented included: personalised LLM agents (ChatGPT custom GPTs; Claude projects) targeting reading comprehension (summarising, clarifying and simplifying e-mails and documents), text revision, and the organisation of meeting transcripts; custom Claude skills encoding her pre-morbid institutional writing style and her habitual presentation formats; recording and transcription of phone meetings; text-to-speech for auditory verification of writing; NotebookLM and Claude for preparing presentations; and voice cloning (ElevenLabs) for delivering them. All AI outputs were subject to human review, and implementation followed GDPR and EU AI Act requirements. Ethical guidelines as required by the relevant authorities have been followed, and informed consent was obtained.

Results. The participant achieved partial return to work, using the tools semi-autonomously to process e-mails, produce documents and deliver presentations. ASHA-FACS communication independence increased from 5.49 to 6.54 (1–7 scale) and qualitative dimensions of communication from 3.56 to 4.50 (1–5 scale). Overall self-reported scores increased from 1.82 to 4.39 (1–5 scale), with the largest gains in professional participation (1.00 to 4.50), communicative autonomy (1.67 to 5.00), perceived usefulness of the tools (1.33 to 4.67) and general satisfaction (2.00 to 5.00). The emotional-impact dimension remained essentially unchanged (3.50 to 3.25), driven by the two reverse-scored items, which moved counter to all other items; this may reflect the increased salience of communication demands after returning to work, or the known difficulty of negatively worded items for people with aphasia. The participant ranked personalised GPTs and NotebookLM as the most important tools in her daily work. Facilitators included preserved cognition, personal motivation, family support, employer funding of the tools, and clinician AI literacy. Barriers included limited AI literacy among supervisors and colleagues, credibility concerns, and the cost of business-grade accounts with adequate data-protection agreements.

Conclusions. Integrated into a participation-focused intervention plan, generative AI functioned as a compensatory tool, comparable in function to conventional AAC, enabling access to language-dependent professional tasks. Notably, because the agents, skills and voice clone were built from her pre-morbid writing style, her habitual presentation formats and her own voice, the AI did not write for her but as she used to write: beyond functional participation, the supports preserved communicative identity and authorship, directly addressing the authenticity concerns raised in her workplace and aligning with self-determination principles in AAC. Limitations include the single-case design without control, the impossibility of attributing improvement solely to AI, and the retrospective, non-validated self-report scale, whose reverse-scored items may not be reliable in aphasia and support the use of consistently positively worded, aphasia-friendly item formats. The case suggests that mainstream AI tools can operate as emerging AAC supports for adults with aphasia, and underlines the need for clinician training, communication partner education, and robust ethical and data-protection safeguards.