Commercial prostheses still rely largely on single-modality electromyography (EMG) control. This entails unreliable signals and forces users to perform mode switching and calibration. Advances in artificial intelligence (AI) techniques, specifically LLMs, promise further advances for prosthetic devices. 13 semi-structured interviews with clinicians, researchers and manufacturers reveal three converging needs: (i) reliable multi-modal intent recognition beyond EMG, (ii) AI-guided socket-design and real-time socket fit optimisation, and (iii) intuitive, personalised, low-cognitive-load human-prosthesis interaction, potentially via Large Language Models (LLMs). Synthesizing these findings with recent literature, we outline the current state of AI adoption and barriers that cause a research-to-practice gap. Multi-modal control remains confined to laboratory settings, while commercial adoption is hindered by fears of over-complexity. We highlight human-in-the-loop continuous calibration games and prioritize speaker-dependent voice commands with LLM reasoning. These findings provide an agenda for engineers, clinicians and researchers to translate AI techniques, specifically LLMs, into user-centric and comfortable prosthetic devices.