IT managers and small teams receive dozens of vendor advisories, newsletters, supplier notices and system mails per week, and most of them do not matter. A scheduled flow collects them, lets a language model drop duplicates and summarise the rest, and posts a single digest where the team already works. Typical benefit: 20 to 60 minutes of reading per person and week replaced by a five-minute scan, with fewer missed security notices.
Problem & value
- Today: people subscribe to newsletters and RSS feeds, a shared mailbox collects supplier notices, monitoring tools mail alerts. Nobody reads all of it; important items (a patch for the firewall in use, a supplier price change, an expiring certificate) are buried or noticed late.
- AI value: an LLM filters by relevance to your own context ("we run Microsoft 365, DATEV, a Fortinet firewall"), merges duplicates, and writes two lines per item with the original link. The flow itself is classic automation; the AI step only reads and condenses.
- Honest benefit: no reliable independent study quantifies time saved for this specific pattern. Measure your own baseline (minutes per week spent on triage) before and after a pilot.
- Scope boundary: this is a team-level helper. Professional media monitoring is LANG-018, competitive intelligence programmes are LANG-017, regulatory horizon scanning is LANG-020.
Variants
- IT security digest: BSI/CERT-Bund advisories and vendor security bulletins filtered by your software inventory.
- Supplier and customer notices: a shared mailbox such as
lieferanten@condensed into a weekly list of changes (prices, delivery terms, end-of-life). - Trade and authority news: industry association, tax and customs news for management.
- System notification digest: hundreds of "backup succeeded" mails reduced to "3 failures, 1 warning".
- Private use: a morning briefing from favourite feeds, summarised by a local model in LM Studio or Ollama.
How it works
- schedule (daily 07:30 / weekly Monday)
- collect (RSS Read, IMAP/Graph mailbox, API of the monitoring tool)
- clean (strip HTML/footers, drop already-seen items, cap length)
- LLM step (relevance vs. context prompt, dedupe, 2-line summary, category)
- format (Adaptive Card, Markdown, HTML mail)
- post to Teams/Slack/Mattermost channel or send mail
- feedback (thumbs-up/down or reply to tune the prompt)
- Key AI component: one prompt with a short "context card" (systems in use, topics to ignore) and a strict output format (JSON or Markdown list). A small model of 3-8B parameters is enough for summarising; classification of relevance works better with the context card than with a larger model.
- Human in the loop: the digest is information only. Nobody should patch, pay or cancel based on the summary alone - the link to the original is mandatory in every item. A named owner reviews the digest quality weekly in the first month.
- Prompt injection: feeds and mails are untrusted text. The flow must not give the LLM tools, credentials or the ability to act; output is posted as plain text only (GUIDE-013).
Data preparation
Which data: a list of feed URLs, the mailbox or folder with notices, and a one-page context card describing your systems and interests.
Where it lives: public RSS/Atom feeds, a shared mailbox in Microsoft 365 or Google Workspace, monitoring or ticket tools with email or API output, newsletters in individual inboxes.
Typical gaps: newsletters land in personal mailboxes, notices arrive as long HTML with tracking links, the same advisory shows up from three sources, and nobody has a current software inventory to judge relevance.
Preparation steps: 1. Create a shared mailbox or mail rule and ask subscribers to use it for newsletters and supplier notices (half a day, plus asking suppliers to change the address). 2. Write the context card: systems, versions, topics, exclusions (1-2 hours with IT; reuse the asset list if one exists, see PREP-001). 3. Pick 5-15 sources to start, not 100. 4. Add a "seen items" store (n8n data table, SharePoint list or Excel file) so items are never summarised twice. 5. Define what may be sent to an external model: public feeds yes, internal alerts with hostnames or customer names only to EU-hosted or local models (PREP-024).
Tools: n8n, Power Automate, shared mailbox, Excel or SharePoint list. Rough share of effort: under 20 percent of total; see GUIDE-003.
Implementation paths
Which path for whom
- SMB - Path C: n8n runs on a small server or existing VM, can also serve AUTO-007, AUTO-008, AUTO-010 and AUTO-016 (Baukasten effect), and a small local model keeps internal alert text in house. If you live fully in Microsoft 365 and only want a news digest, Path A is a fine alternative.
- Enterprise - Path A: a commodity function; Power Automate with AI actions, Copilot in Teams or Slack AI is already licensed and governed by an existing platform.
- Developer - Path C: a few lines of Python or an n8n workflow with Ollama; full control over prompts, sources and logging.
- Private - Path C: local apps (Ollama, LM Studio) with a feed reader such as Miniflux or FreshRSS; no data leaves the device.
Path A - Buy (off-the-shelf)
- Microsoft 365: Power Automate (cloud flow with RSS connector, "Create text with GPT" AI Builder action or Copilot Studio) posting through the Teams Workflows app; Copilot in Outlook and Teams can also summarise threads and channels (AUTO-002, AUTO-004). The old Office 365 connector webhooks in Teams have been retired (final cutoff reported for May 2026; check the Microsoft 365 Developer Blog), so build new flows on Workflows.
- Google Workspace: Gemini in Gmail and Google Chat summaries; Apps Script for scheduled digests.
- Specialised tools: Feedly (AI feeds), Slack AI channel recaps, Zapier or Make with an AI step. Zapier is US-based; Make and n8n Cloud offer EU data centres - check AVV and region.
- Choose A when: you already pay for the licences, sources are public, and no internal alert text goes into the prompt.
Path B - Build on cloud APIs
- Models: Mistral (EU provider), Azure OpenAI or Anthropic and Google models with EU regions, via n8n or a small script. Cost per digest is typically cents; check current price lists.
- Choose B when: you need better summaries of long German-language documents than a small local model delivers, and the content is not sensitive.
Path C - Self-host (open source) (recommended for SMBs)
- Orchestration: n8n self-hosted (RSS Read and RSS Feed Trigger nodes, IMAP, HTTP, Teams/Slack/Mattermost nodes) [1]. Note that the RSS Read node reads one feed per item, so loop over your feed list [1]. Alternatives: Node-RED, Huginn, Apache Airflow for developers.
- Model: Ollama with a small instruction model (for example Qwen3 or Llama 3.x in the 3-8B range, Mistral Small if a GPU is available). A 4-8B model runs on a CPU server or laptop (HW1), slowly but sufficiently for a nightly batch; a 16-24 GB GPU (HW2) is comfortable.
- Feed reader: FreshRSS or Miniflux as collector with an API.
- Ops effort: 1-2 days to set up, then about 1 hour per month for updates and checking failed runs.
Path D - Custom training
Not worthwhile. Better prompts and a context card solve relevance problems; fine-tuning only pays off for dedicated media-monitoring products (LANG-018).
State of the art (2026)
- Summarisation of short texts is a solved commodity; even small open-weight models produce usable two-line summaries in English and German.
- Limits: small models miss subtle relevance ("this CVE affects only version 7.2, you run 7.4"), may invent details when the feed text is truncated, and can be steered by hostile text inside a feed item. Dedupe across sources is imperfect.
- Best practice: keep the original title and link next to every summary, limit input length, ask for "not stated" instead of guessing, and keep a list of critical sources (e.g. the BSI and your firewall vendor) that are always forwarded unfiltered.
- Platform trend: vendors keep adding built-in recaps (Teams, Slack, Gmail), which makes pure news summarisation a commodity; the value of an own flow lies in your context card and your internal sources.
Shared resources & combinations
- Same stack (workflow-automation + llm-small): AUTO-007 attachment filing, AUTO-008 form submissions, AUTO-010 weekly status updates, AUTO-016 change alerts for websites and documents; one n8n instance and one Ollama server serve all.
- Natural combinations: AUTO-004 (chat catch-up) for people, this use case for machines and external sources; AUTO-002 for threads; an escalation into a ticket when a "needs action" flag appears.
- Upgrade path: if the digest grows into a real monitoring programme, move to LANG-017, LANG-018 or LANG-020.
Effort, cost & KPIs
- Path A: 2-4 hours for a first flow if licences exist; cost mostly seats you already pay.
- Path B: 1-2 days incl. testing; API cost small compared with the hours saved.
- Path C: 1-2 days plus 1 hour per month; server cost can be zero on existing hardware.
- KPIs: minutes per week spent on triage (before/after), share of digest items rated useful, missed critical notices (target zero), cost per digest, number of failed runs.
People & change
Involve: the IT manager or department lead as owner, the few end users who receive the digest, IT for the server and mailbox, and the DPO if mailboxes or tickets with personal data are read (GUIDE-005). Management only needs to approve the tool list under the company AI policy (GUIDE-007).
Works council: usually no. A digest of external feeds and system alerts does not evaluate employee behaviour. It becomes relevant (§87(1) no. 6 BetrVG, §90(1) no. 3) once the source is chat or mail content of staff or it reports who handled which ticket; in that case inform the works council early and exclude individual performance data (GUIDE-004). Public bodies involve the Personalrat.
AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy; a 30-minute briefing on "the digest can be wrong, check the link" is enough here (GUIDE-006).
Communication and pilot: start with a team of 3-10 people and 5-10 sources for four weeks, collect thumbs-up/down, and name one AI champion who tunes the prompt (GUIDE-008, GUIDE-009). Typical concerns are "will I miss something?" (answer: critical sources stay unfiltered) and information overload from yet another channel (answer: one digest replaces the others, so unsubscribe the sources).
Humans in control: the digest informs; decisions and actions stay with named people.
Risks, regulation & compliance
- EU AI Act: minimal risk; no Annex III use. Art. 50 transparency applies to chatbots and synthetic content; a clearly labelled internal summary ("AI-generated") is good practice (GUIDE-014).
- GDPR: public feeds carry no personal data; mailboxes, tickets and supplier mails often contain names and email addresses. Conclude an AVV (Art. 28 GDPR) with every external processor, prefer EU regions or local models, and document the purpose. A DPIA is normally not needed for a feed digest (GUIDE-005).
- Security: treat feed text as untrusted (prompt injection), store API keys in the workflow tool's credential store, give the flow read-only mailbox access (GUIDE-013).
- Copyright: summarise and link; do not republish full articles in channels or newsletters.
- Reliability: feeds disappear or change format; alert on empty digests so silence is not mistaken for "nothing happened" (GUIDE-011).
Getting started
- List the 5-10 notice sources that cost your team the most time and set up a shared mailbox for the mailed ones.
- Write a half-page context card (systems, versions, ignore list).
- Build the flow in your existing tool: Power Automate or n8n with RSS, mailbox, one LLM step and a Teams or Mattermost post [1][2].
- Run it for four weeks next to normal reading; ask the team for thumbs-up/down and review misses weekly.
- Unsubscribe the sources that the digest now covers, then add AUTO-016 or AUTO-010 on the same platform.
Sources
- n8n documentation - RSS Read node - https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-base.rssfeedread
- Microsoft Learn - Create an incoming webhook (Teams Workflows) - https://learn.microsoft.com/en-us/microsoftteams/platform/webhooks-and-connectors/how-to/add-incoming-webhook
- Microsoft 365 Developer Blog - Retirement of Office 365 connectors within Microsoft Teams - https://devblogs.microsoft.com/microsoft365dev/retirement-of-office-365-connectors-within-microsoft-teams/
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Ollama - https://ollama.com