Someone in the company checks a supplier's terms page, a funding call or a shared price list "every now and then" and usually forgets. A small flow checks these sources on a schedule, compares the new version with the last one and lets an LLM write two or three sentences: what changed and whether it looks relevant. One employee can set it up in a day; the benefit is that nothing important is missed and nobody scrolls through diffs.
Problem & value
- Today: price lists, general terms, tender portals, funding pages, authority notices and competitor pages change without notice. People check manually, rely on newsletters, or find out when an invoice is higher than expected.
- AI value: classic change detection already says that a page changed, but raw diffs are noisy (cookie banners, dates, counters). The LLM filters noise, summarises the real change in plain language and can rate it against a short relevance instruction ("only tell me about price, delivery time and notice-period changes").
- Honest limit: this is a lightweight watch set up by one employee, not a monitoring programme. For systematic regulatory change management see LANG-020, for structured competitive intelligence see LANG-017. No sourced benefit figures exist for this micro-use-case; measure your own hours saved.
Variants
- Public pages and RSS/Atom feeds: supplier terms, funding calls, authority notices; feeds are more stable than scraping.
- Internal documents: price lists or policies in SharePoint, Nextcloud or Google Drive, triggered on file change (see AUTO-005).
- Private use: rental listings, product prices, municipal notices.
How it works
- schedule (daily/weekly) or file-modified event
- fetch page text / feed / document (HTTP request, RSS, file connector)
- strip noise (navigation, cookie banner, dates) and extract main text
- compare with stored previous version (hash first, text diff second)
- if changed: LLM summary + relevance rating against a short instruction
- send message (email, Teams, Slack) with summary, link and diff
- human decides; store new version as baseline
- Cheap first, AI second: a content hash decides whether anything changed at all; the LLM is only called on real changes. This keeps API costs at cents per month.
- Relevance step: a small model is enough ("price, deadline, legal terms: yes/no, one sentence why").
- Human in the loop: the flow only informs. Nobody should accept new terms or change an order based on the summary alone; the link to the original page is always part of the message.
Data preparation
Which data: a list of sources (URL or file location), the responsible person per source, a one-line definition of "relevant", and a stored baseline version of each source.
Where it lives: in the head of the person who checks manually, in browser bookmarks, in a shared Excel list, or in SharePoint/Nextcloud folders for internal documents.
Typical gaps: JavaScript-loaded pages, login-protected portals, constantly changing elements (dates, counters), unclear ownership of alerts.
Preparation steps: 1. Collect sources in a simple sheet: URL, owner, why it matters, check frequency. 2. Check each source: is there an RSS feed, a PDF or an API? Prefer these over scraping the layout. 3. Define per source what counts as relevant, in one sentence. 4. Name a substitute for every owner so alerts do not die with holidays. 5. For internal files, clean up the folder and rights first (who may see what the flow reads).
Rough share of effort: low, 10 to 20 percent of one to three days. See GUIDE-003.
Implementation paths
Which path for whom
- SMB — Path A: mature change-detection products and the automation tools already licensed (Power Automate, Zapier-style tools) cover the standard case at low cost; no own infrastructure.
- Enterprise — Path B: teams that want the same pattern across many sources, with central logging, access control and an approved LLM endpoint, build it as flows on an internal automation platform.
- Private — Path A: hosted watch services or a free app are enough for personal pages.
Path A — Buy (off-the-shelf)
- Dedicated change-detection services: e.g. Visualping, Distill Web Monitor, Wachete; check where the vendor hosts, whether an AVV (Art. 28 GDPR) is offered and whether AI summaries are included. Visualping documents AI change summaries (vendor-reported); we found no confirmed AI summary feature for Distill or Wachete (Wachete offers digest reports), so verify on the vendor page before buying.
- Already in your stack: Power Automate has an RSS connector and an HTTP action; Microsoft 365 Copilot or Copilot Studio can add the summary step. Google Workspace users can use Apps Script with Gemini. SharePoint and Nextcloud can alert on file changes natively; a Copilot or Nextcloud Assistant step adds the summary.
- Choose A when: a few sources, public data only, nobody wants to run a server.
Path B — Build on cloud APIs
- Orchestration: n8n (cloud region in the EU or self-hosted), Power Automate or Make. A flow of Schedule Trigger → HTTP Request → Code/Compare → LLM → Email/Teams is typically 6 to 10 nodes.
- Models: an EU-hosted endpoint is preferable: Mistral (EU provider), Azure OpenAI in an EU region, or other EU providers such as IONOS AI Model Hub or STACKIT. A small, cheap model is enough.
- Choose B when: you want your own relevance rules, many sources, or login-protected sources handled with your own credentials.
Path C — Self-host (open source)
- changedetection.io (Docker, self-hosted) does the watching, offers a Chrome-based fetcher for JavaScript pages and sends notifications through Apprise to email, Teams or Slack. Combine with n8n (self-hosted) and a small Ollama model (7 to 8B class) for the summary.
- Hardware: HW1 is enough; a small VM or an existing NAS/server runs it, a CPU-only small model is slow but fine for a handful of changes per day.
- Choose C when: internal documents or confidential supplier terms must not leave your network, or the same n8n/Ollama platform already serves other everyday automations (Baukasten effect with AUTO-007, AUTO-009 and AUTO-012).
Path D — Custom training
Not needed. A prompt with two or three examples of "relevant" and "not relevant" changes does the job.
State of the art (2026)
- Diffing is solved, relevance is new: hashing, text diff and headless fetching are mature open source; LLMs add summaries and filtering. Small models suffice.
- Failure modes: the LLM may misread a changed number in a table; sites block bots or change layout so a watch silently breaks; cookie walls and logins defeat simple fetchers. Always include link and raw diff, and add a heartbeat check.
- Legal limit: respect terms of use and robots.txt, poll gently (daily or weekly).
Shared resources & combinations
- Same stack (workflow-automation + llm-general): AUTO-007, AUTO-009, AUTO-010 and AUTO-012 run on the same n8n or Power Automate setup.
- Combinations: AUTO-005 (detailed version comparison), DATA-004 (structured extraction). Scale-up to LANG-020 or LANG-017.
Effort, cost & KPIs
- Path A: about 1 to 3 hours per source group to set up; hosted watch tools are typically priced by number of monitored pages and check frequency (check current vendor pricing).
- Path B: one to two days for a first flow including testing; LLM costs are small because the model only runs on real changes.
- Path C: one to three days for changedetection.io plus n8n plus a small model, then roughly an hour per month for updates and broken watches.
- KPIs: relevant alerts versus false alarms (target: most alerts are read and acted on), time from change to awareness, missed changes found in spot checks, broken watches per quarter.
People & change
Involve early: the department lead who owns the topic (procurement, finance, compliance) and the future recipients; IT for hosting or flow permissions; IT security for credentials on login sources. The DPO is needed only if people-related pages (e.g. staff pages of competitors) or personal data are watched; an AVV with the tool vendor is required whenever a processor is involved (GUIDE-005, GUIDE-010).
Works council: no, in the usual setup. Watching public pages and company price lists does not monitor employees. It becomes relevant (§87(1) no. 6 BetrVG) if the flow logs who opens or acts on alerts, or if recipients' reactions are evaluated. The works council should still be informed in good time when AI is part of work processes (§90(1) no. 3 BetrVG); a short note usually suffices (GUIDE-004). Public bodies inform the Personalrat.
AI literacy: Since 27 Jul 2026, Art. 4 AI Act requires providers and deployers to take measures that support the development of staff AI literacy (no guaranteed level). A 30-minute introduction to what the summary can and cannot do is enough (GUIDE-006).
Communication and pilot: start with three to five sources and two recipients for four weeks; add sources once alert quality is good (GUIDE-008, GUIDE-009, GUIDE-007).
Typical concerns: alert fatigue (fix with a stricter relevance sentence and weekly digests), "is this allowed?" (public pages, gentle polling), and trust in summaries (always link the original).
Humans in control: the flow informs, people decide. No automatic orders, acceptances or replies.
Risks, regulation & compliance
- EU AI Act: minimal risk. No Annex III category; Art. 4 literacy measures apply. No Art. 50 duty arises when summaries are for internal use only (GUIDE-014).
- GDPR: mostly no personal data; it can appear in documents or on pages (names, contact persons). Use an AVV with SaaS and LLM providers, prefer EU hosting, and do not send confidential contracts to consumer chat tools (GUIDE-007).
- Website terms: check terms and robots.txt; store excerpts and diffs, not full copies.
- Security: keep credentials in the tool's secret store with read-only accounts; fetched pages can contain prompt injection, so give the model no tools or write access (GUIDE-013).
Getting started
- List 5 to 10 sources that have hurt you in the past (price changes, missed deadlines) and name an owner each.
- Replace scraping with RSS, PDF or API wherever the source offers it.
- Build one flow for one source in Power Automate or n8n, or add it to a watch tool; write the relevance sentence.
- Run for four weeks, review false alarms and misses, tighten the instruction.
- Roll out to further sources and document owners, fallbacks and the cost in a one-page note.
Sources
- changedetection.io project (self-hosted website change detection, Docker, Apprise notifications) — https://github.com/dgtlmoon/changedetection.io
- Microsoft Learn — RSS connector for Power Automate (trigger, polling interval) — https://learn.microsoft.com/connectors/rss
- n8n documentation — Schedule Trigger node — https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-base.scheduletrigger/
- n8n documentation — HTTP Request node — https://docs.n8n.io/integrations/builtin/core-nodes/n8n-nodes-base.httprequest/
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj