Intro & scope
Low-complexity, mostly commodity-to-mainstream use cases for ordinary employees and small IT teams: AI inside the tools people already use, small automated flows, AI-written scripts and personal productivity habits. Patterns: AUTO.OFFICE (AI features in mail, chat, spreadsheets and documents, AUTO-001..006), AUTO.FLOW (small trigger-based flows with an AI step, AUTO-007..016), AUTO.SCRIPT (AI-written macros, scripts, queries and rules, AUTO-017..021) and AUTO.PERSONAL (own assistants, prompt libraries, planning, voice and photo capture, tutoring, handover, AUTO-022..028).
Scope boundaries: - AUTO covers the personal or small-team level with no or little IT project; industrial-grade pipelines belong to DOC (document extraction, DOC-001, 008, 014), KNOW (enterprise retrieval), AGENT (autonomous agents) and CODE (professional software development). - AUTO-007 (filing attachments) and AUTO-008 (form entries to rows) are light flows; DOC-008 and DOC-003 are the full mailroom and form-processing versions. - AUTO-002, 003, 004 and 028 (mail, meeting, chat, absence) overlap in sources but differ in moment: thread, meeting, channel, leave; AUTO-024 plans the day from the same sources. - AUTO-001, 017, 018, 019 and 021 form a ladder from formula to macro, script, Power Query and regex; AUTO-020 is the IT-admin variant. - AUTO-022 (own assistant) vs AUTO-023 (prompt library) vs AUTO-027 (tutoring); KNOW-002 is the governed company-wide version. - AUTO-012 and AUTO-025 handle voice with personal data and default to path C for small firms and individuals (local transcription); AUTO-014 is the only member with limited AI Act risk (personalised mass mailing).
Overview
AI Act note: all members are minimal risk except AUTO-014 (limited, transparency duties for AI-generated mass communication). Keep a human review step before anything is sent or filed, and avoid feeding HR or personal decisions through these tools.
Shared stack
Common layers: the AI features already licensed (Microsoft 365 Copilot, Google Workspace Gemini, ChatGPT/Claude team plans with EU data residency and DPA), one low-code workflow engine (Power Automate, n8n or Zapier-class) with a human-approval step and run log, one OpenAI-compatible gateway (LiteLLM in front of a managed EU API or a local model) for flows, a shared prompt library (AUTO-023), and simple guardrails: no real personal data in prompts for path A tools without DPA, test with invented sample data, logs of what the flow sent.
- HW1 (CPU/laptop): 3-8B quantised model via Ollama/llama.cpp, Whisper small/medium for voice, Tesseract or a small VLM for photos, n8n. Covers AUTO-001, AUTO-005, AUTO-006, AUTO-007, AUTO-008, AUTO-009, AUTO-012, AUTO-013, AUTO-014, AUTO-015, AUTO-016, AUTO-019, AUTO-021, AUTO-022, AUTO-023, AUTO-024, AUTO-025, AUTO-026, AUTO-028.
- HW2 (single 24-48 GB GPU or EU-hosted API): 7-32B instruct and code model via vLLM, Whisper large, n8n/Power Automate, optional vector index for how-to help. Covers AUTO-002, AUTO-003, AUTO-004, AUTO-010, AUTO-011, AUTO-017, AUTO-018, AUTO-020, AUTO-027 (mail, chat, status, approvals, macros, admin scripts, tutoring).
- Path guidance: A = use the AI built into the existing office suite or a managed EU service first (most AUTO cases); B = build a small flow on a workflow engine with an API or self-hosted model when volume or data sensitivity requires it; C = local or self-hosted processing when voice or customer data are involved (AUTO-009, 012, 025 for small firms). Share one gateway, one workflow engine and one prompt library across the family.