After a holiday, a day of meetings or a busy project week, employees face hundreds of unread messages in many channels. A built-in AI assistant condenses a long thread or a whole channel into a few lines: what was decided, what is still open and what concerns me. It saves reading time, but it reads chat content, which makes works council, permissions and data protection the real work, not the technology.
Problem & value
- Today: people scroll through channels, search for their @mentions or ask a colleague "what did I miss?". Important decisions get overlooked; some staff mute channels and then miss relevant news.
- AI value: a summary on request, limited to what the user could read anyway. It is most useful for long threads, project channels and return from absence (see AUTO-028).
- Evidence: Microsoft documents "Summarize thread" for Teams channel threads and chat recaps via Copilot, with limits (no images, Loop components or shared files; by default only messages of the last 30 days are referenced) [1]. Slack offers channel recaps and thread summaries in its AI features, depending on plan [2]. We found no independent, sourced time-saving figures, so measure your own baseline.
Variants
- Suite-built-in: Copilot in Teams, Slack AI, Google Chat with Gemini.
- Self-hosted chat: Mattermost with the Agents plugin or Matrix/Rocket.Chat with a bot.
- Workflow digest: a scheduled flow posts a daily channel digest (AUTO-009, AUTO-010).
How it works
- user opens channel / thread (or asks "what did I miss since Friday?")
- platform fetches only messages the user may read (permission check)
- LLM summarises: decisions, open questions, mentions, deadlines
- summary shown privately to the user, with links to source messages
- user opens the originals for anything important
- Key AI component: a general LLM with a fixed prompt; no training or labelled data needed.
- Human in the loop: the summary is a reading aid, not a record. Decisions and tasks are confirmed in the original thread.
Data preparation
Which data: the chat messages themselves, nothing else needs labelling.
Where it lives: Teams (Microsoft 365 tenant), Slack workspace, Google Chat or a self-hosted server.
Typical gaps: the real risk is permissions. The assistant sees what the user can see, so private channels that are open to too many members, guest accounts, and old public channels with sensitive content become visible via summaries.
Preparation steps: 1. Review channel membership, guests and external sharing (1-2 days for an SMB tenant; see PREP-023). 2. Define which content must never go to an AI tool, and mark such channels (PREP-024). 3. Check retention settings for chat and AI interaction history.
Rough share of effort: about 10 percent of the project; almost all of it is permission clean-up. See GUIDE-003.
Implementation paths
Which path for whom
- SMB - Path A: Teams or Slack is already in use, the feature is a standard add-on and works in days; nothing to build.
- Enterprise - Path A: this is a commodity function; use the built-in feature and govern it centrally. Build (B) only if one shared internal AI platform should serve chat summaries too.
- Public - Path C: chat often contains personnel or case information, and sovereignty rules favour self-operated chat such as Mattermost with a local model.
Path A - Buy (recommended)
- Microsoft 365 Copilot (Teams): summarise a channel thread or ask Copilot about a chat or channel [1]. Needs a Copilot licence per user; check current Microsoft pricing and the EU Data Boundary documentation for your tenant before budgeting. A Copilot Business offer for smaller tenants exists; prices change, so verify [3].
- Slack AI: thread summaries and channel recaps; the higher-tier Business+ plan adds recaps and more [2]. Slack offers EU data storage regions (Frankfurt, Paris, Stockholm), but we could not confirm in official sources that AI processing stays in the EU. Ask the vendor in writing (GUIDE-010).
- Google Workspace: Gemini in Google Chat; check edition and region settings.
- EU-hosted options: Mattermost (self-hosted or EU cloud), Nextcloud Talk with an AI app, Element/Matrix.
- Choose A when: you already use Teams/Slack, have fewer than a few hundred users and accept the vendor's data processing terms.
Path B - Build on cloud APIs
- Pattern: n8n or Power Automate reads a channel via the Graph API or Slack API on a schedule, sends the text to an EU-hosted model (Azure OpenAI in an EU region, Mistral API, Aleph Alpha-class provider) and posts a daily digest.
- Effort: 2-5 days for the first flow. Limit bot permissions to selected channels, otherwise the bot becomes an oversharing risk.
Path C - Self-host (open source)
- Mattermost Agents plugin with Ollama or vLLM via an OpenAI-compatible endpoint: thread and channel summaries stay on your server. Thread summarisation requires a Mattermost licence; check the current admin guide for the server version [4].
- Models: small open models (about 8B-30B, e.g. Qwen, Llama, Mistral) suffice.
- Hardware: HW2 (one 24 GB GPU) serves a team of a few hundred users; for a handful of users a CPU server with a small quantised model works but is slow.
Path D - Custom training
Not worthwhile. Prompting is enough; tune the prompt for your terminology instead (PREP-015).
State of the art (2026)
- Chat summarisation is a built-in standard feature in the main collaboration suites; quality is good for decisions and action items in plain text.
- Limits: sarcasm, internal jokes, missing context, abbreviations and mixed languages lead to wrong or flat summaries; images and attachments are often ignored; very long histories are cut off (Teams: 30-day default reference window [1]). Summaries can omit the one critical message. Vendor claims on data residency for AI processing differ from storage residency and must be checked.
Shared resources & combinations
- Same capability (office-suite-ai): AUTO-002 mail catch-up, AUTO-003 meeting briefs, AUTO-024 daily planning and AUTO-028 absence handover use the same licence and settings; one Copilot or Gemini rollout serves all.
- Meeting content: SPEECH-001 handles transcripts; this use case covers written chat only.
- Search across tools: KNOW-001.
- Workflow engine and HW2 model server shared with AUTO-009, AUTO-010.
Effort, cost & KPIs
- Path A: 2-5 days including permission review and pilot; main cost is per-user licences.
- Path B: 2-5 days per digest flow; low API cost.
- Path C: 3-7 days setup plus GPU server and chat licence.
- KPIs: weekly active users of the feature, self-reported minutes saved per week (survey), share of summaries rated useful, number of permission issues found and closed, incidents.
People & change
Involve early: management (policy, licences), IT and the information security officer (permissions, logging), the DPO (legal basis, AVV with Microsoft/Slack, records of processing; a full DPIA is usually not required for a pure reading aid, but check if the tool logs per-person usage) and, if one exists, the works council (GUIDE-005, GUIDE-007).
Works council: yes. Chat tools and their AI features log who uses what and when, so they are objectively suited to monitor behaviour or performance and trigger co-determination under §87(1) no. 6 BetrVG, according to the Federal Labour Court's case law. §90(1) no. 3 requires timely information when AI is planned in work processes; §80(3) allows the works council to call in an AI expert. A short works agreement should rule out performance evaluation from chat or AI usage data. Public bodies involve the Personalrat (GUIDE-004).
AI literacy: Art. 4 AI Act requires providers and deployers to take measures that support staff's AI literacy. A 45-minute session suffices: what the summary can miss, never forward summaries of private content, always open the original for decisions (GUIDE-006).
Communication and pilot: 10-20 volunteers in two or three project teams for four weeks, one AI champion; state clearly that the feature does not evaluate people. Typical concern: "is my chat being read?" Answer with the works agreement (GUIDE-008, GUIDE-009).
Humans in control: summaries are never an official record; decisions are confirmed in the thread.
Risks, regulation & compliance
- EU AI Act: minimal risk for a pure summarising aid; Art. 4 literacy duty applies. If summaries or usage data were used to rate employees, the classification would change; exclude this in the works agreement. Art. 50 transparency applies since 2 Aug 2026 mostly to chatbots and synthetic content; mark AI-generated digests as such (GUIDE-014).
- GDPR: chat messages are personal data; AVV (Art. 28) with the provider, purpose limitation and retention rules; check transfer terms for US providers and prefer EU data boundary options.
- Security: prompt injection through messages (an external guest posts hidden instructions) and oversharing; limit bot permissions and exclude guest-heavy channels (GUIDE-013).
Getting started
- Check whether your existing licences already include the feature; list who really needs it.
- Clean up channel membership and guest access, mark sensitive channels (2 days).
- Talk to the DPO and the works council; agree a one-page rule set before the pilot.
- Run a four-week pilot with 10-20 users and a short survey.
- Decide: roll out, limit to some teams, or move to a self-hosted option (GUIDE-002).
Sources
- Microsoft - How to use Microsoft Copilot in Teams chats and channels - https://support.microsoft.com/en-us/teams/copilot/how-to-use-microsoft-365-copilot-in-teams-chats-and-channels
- Slack - Security for AI features in Slack - https://slack.com/help/articles/28310650165907-Security-for-AI-features-in-Slack
- Microsoft - Microsoft 365 Copilot Business pricing - https://www.microsoft.com/en-us/copilot/pricing/business
- Mattermost - Agents plugin documentation - https://github.com/mattermost/mattermost-plugin-agents
- Slack - Slack updates and changes (data residency regions) - https://slack.com/help/articles/115004846068-Slack-updates-and-changes
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj