Many older people live alone, struggle with small touchscreens and feel isolated. A voice-first AI companion talks with them, reminds them of medication and appointments, suggests activities and, with consent, keeps family or carers informed. Pilots report high engagement and lower self-reported loneliness, but evidence is mostly vendor- or programme-reported, and the use case sits close to the EU AI Act's ban on exploiting vulnerabilities (Art. 5).
Problem & value
- Today: loneliness and isolation among older adults are addressed by family visits, volunteer phone circles, day centres and, for emergencies, a button-based home emergency call. Carers are short of time; reminders are paper notes or phone calls.
- AI value: always-available, patient conversation in natural speech; proactive check-ins; cognitive stimulation (reminiscence, quizzes); lightweight routine support; relief for carers.
- Evidence: the New York State Office for the Aging reported, for a programme deploying the ElliQ companion to older New Yorkers, a 95% self-reported reduction in loneliness and very frequent daily interaction [1]. Caveat: this is a state-announced result produced with the vendor, with no independent control group found. Treat all such figures as indicative, not proof of clinical benefit.
Variants
- Private / family purchase: tabletop companion or smart speaker for a parent at home.
- Care-home and day-care deployment: group activities, resident check-ins, staff dashboards.
- Municipal / social-services programmes: loan devices to isolated seniors (public sector).
- Proactive wellbeing calls: a phone voice agent calls at set times (see CONV-002, CONV-008 patterns).
- Dementia-support mode: orientation prompts, reminiscence, caregiver alerts; highest vulnerability, strictest safeguards.
- Assistive input: speech access to video calls, news, radio, smart home for people with limited mobility or vision (see PERS-007).
How it works
- wake word / button / scheduled check-in
- ASR (accent- and impairment-robust, slow speech tolerant)
- dialogue manager: LLM + persona + safety policy + user memory ↳ tools: reminders, calendar, call family, music/radio, smart-home, care-plan lookup
- guardrails (no medical advice, no manipulation, crisis detection)
- TTS (clear, slow, warm voice; consistent persona)
- event log
- consented summary to family/carer; escalation path
- Memory: a small per-person profile (family names, routines, preferences) in a vector or key-value store, visible and editable by the user or a trusted person.
- Proactivity: schedule-driven prompts; the LLM adapts timing and tone, but reminders for medication come from a deterministic schedule, not from model generation.
- Escalation: keywords or patterns indicating a fall, pain, confusion or distress trigger a defined path (ask, call contact, notify carer, call emergency services only with clear rules). A human carer receives alerts and decides.
- Human-in-the-loop: carers set the care plan and thresholds; family reviews summaries; the user can switch off sharing.
Data preparation
The system needs little classic business data. Core inputs are a person profile (preferences, topics, biography), a care plan with medication and appointment schedules entered by carers or family, optional smart-home or wearable events, and family messages and photos. These live in care documentation software, paper care folders, calendars and phones.
Typical gaps: care plans on paper or in free text, no structured medication schedule, missing consent and no clear legal guardian or authorisation, speech data that does not reflect dialect, slow or impaired speech, and unclear access rights for family versus carers.
Preparation steps: digitise the relevant care plan fields into a structured schedule (PREP-008, PREP-011), define who may see and edit what, with a recorded consent per recipient (PREP-025, PREP-024), pseudonymise any data used for tuning or evaluation (PREP-022), and build a test set of realistic utterances, including dialect and unclear speech, with carers (PREP-018). Collect good conversation examples and safe responses for crisis topics (PREP-016). Tools: forms, a simple schedule database, consent templates. Rough share of effort: under 20 percent; consent and care-plan structure matter most.
Implementation paths
Which path for whom
- Private (older adults and families): Path A, because a certified, easy-to-use product with support and emergency escalation is safest for non-technical users.
- Enterprise (care providers, insurers, housing companies): Path B, because the companion must integrate with care plans, alerting and carer workflows, and health-related data needs a controlled EU setup.
- Public (municipalities, social services): Path C or sovereign EU cloud in practice, chosen here as C, because health and social data call for sovereignty, procurement rules and local operation on HW2.
- Developer (assistive-tech builders): Path C, because open ASR, TTS and small models run on-device, protect conversations and can be tuned for dialects.
Path A — Buy (recommended for private households)
- Dedicated companion devices: ElliQ (Intuition Robotics, Israel/US; check language support for EU markets), Pillo-style dispensers, Joy for All robotic pets (non-LLM comfort devices).
- Care-tech platforms with voice: Tunstall, Vitaphone and similar home-emergency-call vendors, plus EU care-robotics and telecare providers (check current offerings per country); smart-speaker assistants such as Alexa+, Google Gemini for Home, Apple Siri on home devices for basic routines.
- Choose A when: a family or municipality needs a working product within weeks and accepts vendor cloud processing. Check EU hosting, German/local language support, DPA, device repair and what happens to data on discontinuation.
Path B — Build on cloud APIs
- Realtime voice: OpenAI Realtime API, Google Gemini Live API, or a pipeline of ASR + LLM + TTS (Deepgram, Speechmatics, ElevenLabs; EU-region options and EU providers such as Mistral Voxtral for ASR, Mistral LLMs, Azure OpenAI / Vertex AI / Amazon Bedrock EU regions, Claude via EU cloud regions).
- Orchestration: LiveKit Agents or Pipecat for streaming, telephony via SIP (EU carriers) for call-based companions; Home Assistant for smart-home tools.
- Choose B when: a care provider wants its own branded service with integration into care records.
Path C — Self-host (open source)
- ASR: Whisper large-v3 / faster-whisper, Voxtral (open weights), NVIDIA Parakeet/Canary. TTS: Piper, Kokoro, XTTS-style models (check licences). LLM: Mistral, Llama, Qwen or Gemma class models via vLLM/Ollama. Local voice stack: Home Assistant Assist with Wyoming protocol; Pipecat / LiveKit self-hosted.
- Hardware: HW2 (one 16–32 GB GPU) serves a care home with a ~8-14B model; single homes can run small models on HW1 mini-PCs with degraded quality.
- Why: privacy-sensitive care settings; data never leaves the building. Ops effort includes updates, monitoring and device support.
Path D — Custom training
Worthwhile for dialect, aphasia or dysarthric speech ASR fine-tuning, and for local-language persona tuning. Needs consented speech data from older speakers (GDPR Art. 9 if health-related) and ethics review. Do not fine-tune on private conversations without explicit consent.
State of the art (2026)
- Realtime speech-to-speech models give natural, interruptible conversation with low latency; open ASR is strong for standard speech but still weaker on elderly, dialect, dysarthric and noisy speech.
- Long-term memory and proactive agents make companions feel continuous, which also increases the risk of dependency.
- Limits and failure modes: hallucinated facts or medication advice, mis-heard speech, sycophancy and emotional over-reliance, confusing a person with dementia, false reassurance in emergencies, voice cloning misuse. Companion chatbots have prompted regulator attention and public-safety concerns, particularly for minors and vulnerable users.
- Research gap: most trials are short, small and self-reported; effects on real loneliness over time and on cognitive decline remain uncertain. Robotic and social-agent reviews call for controlled studies.
Shared resources & combinations
- Voice stack shared with CONV-002 (phone voice agent), CONV-014 (in-car/embedded voice), CONV-015 (personal assistant), CONV-009 (wellbeing companion).
- Care context: HLTH-025 (voice-based nursing documentation), PERS-002 (health and sleep coach with wearables).
- Related ethics: CONV-023 (social companions), CONV-028 (memorial avatars).
- Hardware: the same HW2 voice server can run several of these.
Effort, cost & KPIs
- Path A: days to weeks setup; per-device purchase or monthly subscription; install support for seniors is a hidden cost.
- Path B: 3–6 months for a production service with safety evaluation; cost drivers are realtime voice minutes and support.
- Path C: 4–8 months; GPU, device hardware, maintenance.
- KPIs: daily active use, loneliness scale change (e.g. UCLA-3 or De Jong Gierveld, pre/post), reminder adherence, escalation precision/recall, carer time saved, user-reported trust, opt-out and complaint rates. Measure safety incidents separately.
People & change
Involve the older person first, then family, carers and the department lead of the care service, plus management, the DPO and legal/compliance. Health-related and behaviour data of vulnerable people make a DPIA necessary; consent must be informed and revocable, with guardians where capacity is limited. HR joins when staff use the system.
Works council: depends. For households no; in care facilities or services, staff who use dashboards, alerts or activity summaries can be monitored by the system, so §87(1) no. 6 BetrVG applies, and §90(1) no. 3 covers AI-supported care processes. Public bodies involve the Personalrat.
AI literacy: support carers with training on limits, escalation and recognising errors (AI Act Art. 4; GUIDE-006). Users must be told clearly that they talk to an AI (Art. 50). Pilot with a small group of volunteers who have carer support and a fallback by phone. Typical concerns: replacing human contact, deception and dependency, privacy; address by positioning the companion as a supplement, with no secret monitoring, transparent alerts and a human always able to intervene. See GUIDE-005, GUIDE-008 and GUIDE-009.
Risks, regulation & compliance
- EU AI Act: typically limited risk (Art. 50: users must be told they talk to an AI; synthetic voice marking). Art. 5 prohibitions apply: AI that uses subliminal, manipulative or deceptive techniques, or exploits vulnerabilities due to age or disability to materially distort behaviour in a harmful way, is prohibited (Art. 5(1)(a),(b)) [2][3]. Companions must not nudge purchases, discourage contact with family, or foster emotional dependence for engagement. Inferring emotions of employees or students is banned (Art. 5(1)(f)); applying emotion recognition on care staff would fall here, and emotion inference in care settings needs careful legal review. Systems that act as, or feed, medical devices (diagnosing, monitoring conditions) can become high risk via MDR (Annex I).
- GDPR: voice recordings and conversations reveal health data (Art. 9). Require a legal basis (explicit consent where feasible), data minimisation, local processing where possible, DPIA (Art. 35), clear sharing controls toward family, short retention, and consent capacity checks for people with dementia (involve legal guardians). Always-on microphones need transparent indicators.
- Other: national care law, consumer protection, accessibility (European Accessibility Act, in force 2025), cybersecurity (Cyber Resilience Act for connected devices), scams targeting seniors via voice cloning.
- Mitigations: persona and system prompts that forbid medical advice and dependence-building, red-team tests with older testers, human escalation paths, easy off switch, transparency about AI, independent evaluation.
Getting started
- Define the goal (loneliness, reminders, safety) and involve seniors, family and carers; write a short DPIA.
- Trial 2-3 Path A products with 5-10 real users for 8 weeks; record outcome scales and incidents.
- Define safety rules: escalation triggers, forbidden topics, disclosure of AI nature.
- If building, prototype with a speech pipeline in Pipecat or LiveKit and test ASR with older speakers in your language.
- Plan support: installation, repair, a human contact number, and an exit plan for data.
Sources
- NYSOFA — ElliQ rollout shows 95% reduction in loneliness (Aug 2023) — https://aging.ny.gov/news/nysofas-rollout-ai-companion-robot-elliq-shows-95-reduction-loneliness
- Regulation (EU) 2024/1689 (AI Act), Arts. 5, 50 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- European Commission — Guidelines on prohibited AI practices (Feb 2025) — https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-prohibited-artificial-intelligence-ai-practices-defined-ai-act
- GDPR (Regulation (EU) 2016/679) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- Pipecat open-source voice agent framework — https://github.com/pipecat-ai/pipecat
- Home Assistant Voice (Assist) — https://www.home-assistant.io/voice_control/
- OpenAI Realtime API docs — https://platform.openai.com/docs/guides/realtime
- Whisper — https://github.com/openai/whisper