Automating business processes with n8n (and where AI comes into play)
n8n automates processes self-hosted in the EU. Where AI helps, where it gets fragile, and when real engineering is needed.
This translation was produced automatically using AI. The German version is the editorially reviewed original.
n8n is interesting for the Mittelstand for one reason above all: you can run it on your own server in the EU and automate processes without data leaving the building. That is the GDPR lever most US SaaS tools do not offer.
The honest part comes after. n8n takes the wiring of business processes off your hands, and with its AI nodes comes real language processing. But once a language model is in the loop, the flow stops being deterministic, and that is exactly where it is decided whether an automation holds up day to day. This piece shows both: what n8n does well, and where the line to real engineering runs.
01. What n8n actually is
n8n is a node-based platform for workflow automation. You assemble a workflow visually from nodes and can add your own JavaScript or Python code at any point.
🔸 Self-hosted or cloud. The decisive point for the Mittelstand: n8n runs as a Docker container on your own infrastructure. Processing and data stay on your own network, and your company remains the sole data controller.
🔸 Over 400 integrations, plus an HTTP node for any REST interface. Whatever isn't available as a ready-made node, you connect yourself.
🔸 Fair-code, not classic open source. The Sustainable Use License permits internal business use free of charge. What it prohibits is reselling n8n as a hosted product to third parties. For standard in-house use this is uncritical, but it's worth naming clearly for a technical audience.
If you're still deciding whether AI should run locally at all, you'll find the trade-off in On-prem vs. cloud LLM.
02. How n8n automates business processes
The structure is always the same: a trigger starts the workflow, then nodes run step by step.
🔸 Trigger: webhook, schedule, an app event such as a new email or a new CRM entry, a submitted form.
🔸 Nodes: app integrations, HTTP calls, database queries, logic such as IF and Switch, data transformation.
In practice, we see the same patterns again and again in the Mittelstand:
| Use case | n8n mechanics | Where AI helps |
|---|---|---|
| Lead routing | Form trigger to CRM to Teams notification | Agent qualifies and summarises the lead |
| Invoice intake | Email trigger to extraction to ERP | LLM reads fields from the PDF, identifies the document type |
| CRM reconciliation | Schedule to bidirectional API nodes | LLM normalises and deduplicates free text |
| Reporting | Schedule to database to email | LLM writes the management summary |
This is the same logic we used to build our own Event Scout micro-app, before we decided what deserved real engineering.
03. Where AI comes into play
n8n didn't bolt AI on. It integrated it as its own node family built on LangChain.
🔸 AI Agent Node. A full-fledged agent with tool calls, memory and an iterative reasoning loop, not just a prompt field. The agent gets other nodes or workflows as tools.
🔸 RAG over your own documents. Since version 1.74.0 in January 2025, vector stores such as Qdrant, Weaviate or pgvector can be connected directly as an agent tool. This lets the system answer based on your own documents, not on the model's general knowledge. We explained exactly what RAG is in What is RAG.
🔸 Local models. Via the Ollama connector, inference runs on your own hardware. Self-hosted n8n plus a local model gives you a fully private AI stack with no third-country transfer. What that hardware realistically costs is covered in Local LLM in the enterprise.
What works well: classification, extraction, routing, RAG, simple agents. It gets fragile with long, multi-step agent chains and with the reproducibility of model outputs.
04. Limits, and when it tips toward engineering
A few points we raise with clients before they start:
🔸 Determinism and debugging. An AI step does not necessarily produce the same output for the same input. Debugging long agent runs is tedious.
🔸 Versioning. Workflows live in the database. Clean, Git-based review isn't the native standard, it's more a matter of the enterprise features.
🔸 Scaling. Single-server mode hits its limits with many parallel webhooks. In production you need queue mode with separate workers.
🔸 Maintenance. Self-hosted means updates, security patches, backups and monitoring are on you.
It tips toward real engineering as soon as business-critical, deterministic logic, high transaction volumes, strict service levels or sustained multi-agent operation are required. At that point you need tests, CI/CD and observability, not just clicks in the editor. The n8n 2.0 hardening release from January 2026 has made progress here, with isolated code execution and more stable behaviour under load.
05. How we place n8n at iiterate
We like to use n8n for exactly what it's good at: automating internal workflows quickly, plugging in AI wherever it genuinely makes a step better, and running the whole thing GDPR-compliant, in-house.
The mistake we see most often is confusing a working workflow with a production-ready system. For internal, well-tolerated processes, n8n is often the whole answer. Once the process becomes business-critical, n8n is the fastest way to find the right workflow, and the point at which we carry the proven path over into our own engineering. We've summarised where AI overall makes sense for the Mittelstand in AI solutions for the Mittelstand.
