n8n · SERVICE
n8n Workflow Automation: routine out, minds free
Self-hosted n8n on your server plus ready-made workflows. Invoice triage, lead routing, mail classification, daily summary – documented and handed over.
Researched & fact-checked by: DuneDive LLC · As of: 2026-05
What is n8n workflow automation?
n8n is an open-source platform for workflow automation. It connects existing systems – email, CRM, accounting, databases, LLM APIs, webhooks – via a visual interface into flows that run without manual clicks. Founded in Berlin in 2019, n8n has been stable in production since version 1.0 (2023) and is used in over 60 languages as of May 2026. Unlike Zapier or Make.com, n8n runs on your own server – no data at the vendor, no per-execution limits, no fees per 1,000 tasks.
As a service from us: we install n8n via Docker-Compose on your server, connect it to your sources, build the workflows, document each step and hand it over. You do not get a black box but a readable workflow graph that your IT team can extend. Workflows include LLM nodes where classification or extraction is needed – coupled to your LiteLLM gateway so model choice can later be changed without altering the flow.
The difference from classic script automation: n8n makes processes visible. An accountant sees what happens, when it breaks, where it hangs. She can adjust conditions herself – without writing TypeScript.
Why it matters
In every fiduciary or SME operation there are 5 to 30 recurring processes that each consume 5–15 minutes of staff time per case. Inspect incoming invoice, extract supplier name, draft booking, approval email. Pull lead from website, create in CRM, send confirmation, assign to sales. These are not intellectually demanding tasks – they are friction. n8n removes that friction without anyone having to learn to code.
Two effects are measurable. First: 60–80 % time saved on the automated steps – a 5-person practice typically gets 4–8 staff-hours back per week. Second: fewer errors. Manual steps produce typos, forgotten copies, missing fields. A workflow runs deterministically or not at all – and "not" is visible in Loki and Telegram.
Add the LLM lever. Workflows that used to be rule-based only ("if sender = X, then Y") can now use an LLM node to classify mail, extract amounts from PDFs, judge urgency. That opens automation for tasks only humans could handle before 2024.
How we set it up
The service follows a clear path: workshop, install, build, test, handover.
In the workshop (half a day) we jointly identify the 3–5 processes that cause real pain. We talk to the people who run them today – not management alone. That avoids the classic mistake of automating what looks important from the outside instead of what hurts inside.
Installation runs via Docker-Compose with n8n, Postgres as the workflow database, and Redis for queue mode (parallel execution). Access through a reverse proxy with SSL, authentication via SSO or basic auth. Backups run daily to S3-compatible storage.
Workflow construction is iterative. We build version 1, let it run two weeks on real data, see where it breaks, harden it. Every workflow has: trigger (cron, webhook, IMAP, database polling), processing nodes (LLM, code, HTTP), output (email, CRM API, database insert), error branch (what happens when a step fails – retry, Telegram alert, escalation email).
LLM nodes point at your multi-LLM gateway, not directly at OpenAI. That lets you change the model later – from GPT-4o to Claude-Sonnet to Mistral – without changing the workflow. Routing by sensitivity (local model for PII content) also lives in the gateway, not the workflow.
Handover means: workflow exported as JSON into the repository, documentation per workflow (what it does, where it breaks, who gets alerted), training for 1–2 internal power users (3 hours).
From workshop to handover
- 01Workshop (half a day): identify 3–5 pain processes together with the people who run them, prioritise, document.
- 02Install: n8n via Docker-Compose on your server, Postgres persistence, Redis queue mode, SSL via reverse proxy, daily backup.
- 03Connections: credentials for the involved systems (mail server, CRM, accounting, LLM gateway, cloud storage) – encrypted in the n8n vault.
- 04Workflow build v1: 3 or 5 flows with trigger, processing, output, error branch and Telegram alert. LLM nodes point at LiteLLM, not directly.
- 05Test run (2 weeks): real data, watched in n8n execution list plus Loki logs. Document edge cases and fold them back into the workflows.
- 06Documentation: one Markdown page per workflow with purpose, trigger schedule, exception handling, on-call contact. Export as JSON into the Git repo.
- 07Handover: 3-hour training for 1–2 internal power users, plus a 30-day adjustment guarantee.
When to use
n8n is the right choice when (a) the process repeats daily or weekly, (b) the involved systems have APIs or at least IMAP/SMTP/SFTP, (c) business rules are stable enough that a change per quarter is sufficient.
Concrete cases we built in the past 12 months: invoice triage (PDF arrives by mail, LLM extracts supplier/amount/date, compares with open orders, sends booking suggestion to bookkeeper); lead routing (website form or LinkedIn DM lands in Postgres, LLM classifies by industry and urgency, assigns sales rep, sends personalised confirmation); mail classification (inbox of an info@ address is read, sorted into 6 categories, each routed to the right team); daily summary (cron runs at 07:00, gathers KPIs from 4 systems, generates a Markdown report via LLM, sends to management); document ingestion (new file on OneDrive/Drive triggers workflow, OCR, metadata extraction, indexing into Qdrant for RAG).
Our two packages: "Setup + 3 Workflows" for CHF 1,200 covers the entry. "Workflow pack with 5 custom flows + AI triage" for CHF 3,500 is the typical step for fiduciary offices with 5–20 staff.
When not to use
n8n is the wrong choice when the task is one-off – a data migration, a report that runs exactly once a year. A script is cheaper than a continuously maintained system.
It is also wrong when the underlying business rules are not yet clear. If accounting itself does not know "how is a supplier invoice handled today", n8n automates an imprecise rule – faster and therefore more harmful. Clarify the process first, then automate.
Be careful with strict compliance requirements lacking an audit trail. n8n logs every execution with inputs and outputs – good for debugging, but depending on data class it can itself become a compliance problem. In sensitive areas (client privilege, AML suspicion cases) a log-retention policy is needed, plus possibly pseudonymisation inside the workflow.
n8n is not suited to real-time use cases with sub-100 ms latency – the architecture targets the second-to-minute range. Anyone needing a chatbot with instant reply does not build it in n8n but as a dedicated service (see Bots & Voice Agents).
Trade-offs
STRENGTHS
- Self-hosted: data does not leave your infrastructure, executions are unlimited
- Visual graph: flows are readable, adjustments possible without code
- LLM nodes embeddable in workflows – mail classification, receipt extraction, summary
- Error branches with Telegram alerts: breaks are visible, not silently swallowed
WEAKNESSES
- Self-hosted also means operations: updates, backups, monitoring need someone (or managed service)
- With very many parallel executions, queue mode with Redis must be configured cleanly
- Logs contain workflow inputs – for sensitive data a retention policy is needed
- Not for sub-100 ms latency; workflows live in the seconds-to-minutes range
FAQ
Why self-hosted and not n8n.cloud?
n8n.cloud starts at USD 24/month for the Starter tier with 2,500 executions, USD 60 for Pro with 10,000. A fiduciary workflow with mail polling quickly hits 50,000 executions per month – the cloud gets expensive. More importantly: client data and receipts pass through the workflow. We recommend this data does not leave your infrastructure. Self-hosted on Hetzner costs between CHF 5 and 15 per month for the server share, executions are unlimited.
What happens when a workflow breaks at night?
Every workflow has an error branch. If a node breaks, the case is not silently dropped – it lands in an error queue, a Telegram alert goes to on-call, the next workflow run retries automatically. The n8n execution list shows every failure with input and stack trace. With Pro managed service we respond within 4 hours, with Plus 24/7.
Can our bookkeeper build new workflows herself?
After the 3-hour training, trained power users can make simple changes (change recipient, add a condition, shift the schedule). New, more complex workflows with LLM nodes and error handling we typically build together in a 2-hour pair-programming session. That keeps the learning curve gentle and avoids someone without knowledge deploying a critical automation to production.
Which LLM models do we use in the workflows?
Through the multi-LLM gateway we route by sensitivity and task. For PII content (client name, receipt content) a local model (Mistral 7B or Llama 3.1 8B on Ollama). For general classification Claude-Haiku or GPT-4o-mini (cheap, fast). For complex extraction with reasoning Claude-Sonnet or GPT-4o. The choice lives in the gateway routing YAML, not in the workflow – you can change the model without touching the workflows.
Related topics
Sources
- n8n.io – Self-hosted documentation · 2026-04
- n8n.io – Cloud pricing (Starter, Pro, Enterprise) · 2026-05
- n8n.io – Queue mode and Redis configuration · 2026-03
- n8n GitHub – Release notes 1.x · 2026-05