Intro and scope
Chat, voice and avatar assistants that talk to people and, where permitted, act through back-office APIs. Four patterns: - CONV.SERVICE: text-first assistants answering from a knowledge base and doing simple transactions (customer, employee, citizen, patient, candidate, in-app). - CONV.VOICE: telephone and embedded voice agents (inbound, outbound, drive-thru, in-car, interview, emergency). - CONV.SALES: pre-sales and shopping assistants. - CONV.COMPANION: personal, wellbeing, social, child and memorial companions.
Scope boundaries: - CONV-001 is the generic baseline; CONV-011, -012, -013, -004 and -026 are domain or channel specialisations that reuse it. - CONV-002 (inbound), CONV-008 (outbound) and CONV-003 (booking-only) share the voice stack; CONV-007 assists a human instead of replacing one. - CONV-017 only communicates with candidates; CONV-019 scores them (high risk, Annex III employment); CONV-020 is research, not selection. - CONV-009 is self-help; CONV-010 is clinical intake (medical-device and high-risk implications). - CONV-018 is only the avatar front end over a CONV-001/-002 brain. - CONV-015 is a personal assistant (delegated access); CONV-026 is a vendor-embedded copilot. - Knowledge retrieval itself belongs to KNOW; speech components to SPEECH; agents to AGENT. - AI Act: Art. 50 disclosure applies to all (users must know they talk to an AI); high-risk where the output steers access to essential services, employment or emergency dispatch. Emotion or minors' vulnerabilities (CONV-009, -016, -023, -024) need extra safeguards.
Overview
| ID | Title | Pattern | Maturity | Path | HW tier | AI Act risk |
|---|---|---|---|---|---|---|
| CONV-001 | Customer service chatbot (website, messaging, FAQ) | CONV.SERVICE | mainstream | A | HW3 | limited |
| CONV-002 | AI phone agent / voicebot for inbound calls | CONV.VOICE | early-adopter | A | HW3 | limited |
| CONV-003 | Appointment booking assistant (practices, salons, workshops) | CONV.SERVICE | mainstream | A | HW2 | limited |
| CONV-004 | Internal IT helpdesk assistant | CONV.SERVICE | mainstream | A | HW3 | limited |
| CONV-005 | Sales assistant & lead qualification chatbot | CONV.SALES | mainstream | A | HW2 | limited |
| CONV-006 | Conversational commerce & shopping assistant | CONV.SALES | early-adopter | A | HW2 | limited |
| CONV-007 | Real-time agent assist for contact centres | CONV.SERVICE | mainstream | A | HW3 | depends |
| CONV-008 | Outbound voice agent (reminders, surveys, collections) | CONV.VOICE | early-adopter | A | HW3 | limited |
| CONV-009 | Mental health & wellbeing companion | CONV.COMPANION | early-adopter | A | HW3 | depends |
| CONV-010 | Patient intake & symptom triage assistant | CONV.SERVICE | early-adopter | A | HW3 | high |
| CONV-011 | Citizen services chatbot (public administration) | CONV.SERVICE | early-adopter | A | HW3 | limited |
| CONV-012 | Travel & hospitality concierge | CONV.SERVICE | mainstream | A | HW2 | limited |
| CONV-013 | Banking & insurance customer assistant (transactional) | CONV.SERVICE | early-adopter | A | HW3 | limited |
| CONV-014 | In-car & embedded voice assistant | CONV.VOICE | early-adopter | A | HW1 | limited |
| CONV-015 | Personal AI assistant (email, calendar, tasks) | CONV.COMPANION | early-adopter | A | HW3 | limited |
| CONV-016 | Companion & assistive conversation for elderly people | CONV.COMPANION | early-adopter | A | HW2 | limited |
| CONV-017 | Recruiting chatbot & candidate communication | CONV.SERVICE | mainstream | A | HW2 | depends |
| CONV-018 | Digital human / avatar for customer interaction | CONV.SERVICE | early-adopter | A | HW3 | limited |
| CONV-019 | AI-conducted screening interviews & candidate assessment | CONV.VOICE | early-adopter | A | HW3 | high |
| CONV-020 | AI-moderated customer & user research interviews | CONV.SERVICE | early-adopter | A | HW3 | limited |
| CONV-021 | Emergency & non-emergency call triage assistant (112, police, control rooms) | CONV.VOICE | early-adopter | A | HW3 | high |
| CONV-023 | Social companion & character chat (entertainment, role-play) | CONV.COMPANION | mainstream | B | HW3 | limited |
| CONV-024 | Conversational AI toys & child-directed voice companions | CONV.COMPANION | emerging | B | HW1 | depends |
| CONV-025 | Drive-thru & restaurant voice ordering | CONV.VOICE | early-adopter | A | HW1 | limited |
| CONV-026 | In-product AI copilot (embedded assistant in SaaS & apps) | CONV.SERVICE | early-adopter | B | HW3 | limited |
| CONV-027 | Live-camera visual assistance & remote guidance (smartphone, smart glasses) | CONV.SERVICE | early-adopter | A | HW3 | limited |
| CONV-028 | Digital legacy & memorial avatars (griefbots, interactive testimony) | CONV.COMPANION | early-adopter | A | HW2 | limited |
Shared stack
Common building blocks: LLM gateway with guardrails and logging, embedding model + reranker + vector DB (shared RAG index with KNOW), ASR/TTS for voice, workflow/agent layer with tool APIs (CRM, calendar, ticketing, POS), human-handover queue, AI-disclosure and consent/retention layer.
HW1 (edge, device-class) for CONV-014, -024, -025 (and offline parts of -016): - Small LLM 1-4B (Qwen/Gemma/Phi class, INT4) via llama.cpp or ONNX/TensorRT on SoC or Jetson-class NPU. - Streaming ASR (whisper.cpp / small Whisper or Parakeet), lightweight neural TTS (Piper/Kokoro), wake-word and VAD on device. - Cloud fallback for hard queries; local intent classifier; no raw audio leaves the device by default.
HW2 (single GPU / small server, 16-24 GB VRAM) for CONV-003, -005, -006, -012, -016, -017, -028: - vLLM or Ollama serving a 7-14B instruct model (INT4/FP8) plus small model for classification. - bge-m3 / multilingual-e5 embeddings, bge-reranker, Qdrant or pgvector, Postgres. - Whisper-large-v3-turbo or faster-whisper and Piper/XTTS for voice; n8n or Temporal for workflows; Langfuse for tracing. - Shares the RAG index and gateway with KNOW-002 and the ASR/TTS services with SPEECH use cases.
HW3 (multi-GPU server / EU private cloud, 2-8 x 40-80 GB) for CONV-001, -002, -004, -007, -008, -009, -010, -011, -013, -015, -018, -019, -020, -021, -023, -026, -027: - vLLM/SGLang cluster with a 30-120B model (Mistral, Llama, Qwen, gpt-oss class) plus reasoning model for scoring/triage (CONV-010, -019, -021). - Dedicated streaming ASR with diarization (CONV-007, -021), low-latency TTS or speech-to-speech model, SIP/WebRTC gateway (LiveKit, Asterisk/FreeSWITCH) for telephony. - VLM endpoint for CONV-027, avatar renderer (GPU) for CONV-018/-028. - Qdrant/Weaviate cluster with reranker, Keycloak SSO, policy engine/guardrail service, audit log; EU-hosted, DPA and no training on customer data.
Path A vs B: Path A (managed EU-region APIs behind the shared gateway) is the default; Path B (self-hosted) is pursued for CONV-023, -024 and -026 where data control, cost or latency demands it. All tiers should expose one OpenAI-compatible gateway so use cases swap models without code changes.