For most SMBs, AI-powered automation starts with a SaaS platform: Make.com, Zapier, or a similar tool connected to a cloud language model via API. Every email summary, every support ticket classification, every CRM update costs a fraction of a cent, and those fractions accumulate into three-figure monthly bills. More importantly, every document you automate travels through infrastructure you do not control.
In 2026, the self-hosted AI community has settled on a practical alternative: n8n as the workflow automation layer, and Ollama as the local LLM runtime. Both run entirely on your own hardware. The result is AI-powered workflows that generate no per-token API costs, require no third-party data processing agreements, and keep every processed document on your own network.
What n8n Is and Why It Matters for SMBs
n8n is an open-source workflow automation platform with a visual builder, comparable to Zapier or Make.com, but self-hosted and with native AI node support. It connects to hundreds of services through pre-built integrations: email inboxes, CRM systems, ERP exports, calendars, ticketing platforms, Slack, RSS feeds, and more.
The n8n team publishes a Self-Hosted AI Starter Kit on GitHub that bundles n8n, Ollama, PostgreSQL, and Qdrant (for vector search) in a single Docker Compose stack. Ollama exposes an OpenAI-compatible REST endpoint on port 11434 that n8n consumes without additional configuration. Community reports suggest the stack is production-ready within an hour from scratch.
The practical implication: AI-driven workflows that run exclusively on your own infrastructure, with no route for business data to reach external systems.
Use Cases That Work Well for SMBs
Email Triage and Prioritisation
A workflow reads incoming emails, sends them to the local model, and extracts priority, topic, and a suggested response. Results land in a kanban board or CRM before a team member opens the inbox. With Llama 3.1 8B, the model most commonly recommended by practitioners for real-time workflows, latency is low enough for near-instant processing.
Document Routing and Classification
Invoices, contracts, delivery notes, job applications: n8n can read incoming documents, classify them, and route them to the correct folder or inbox without file contents ever reaching an external API. For industries handling confidential material, legal, accountancy, healthcare, this is a structural compliance advantage.
Support Ticket Pre-Processing
Incoming support requests are summarised, categorised, and enriched with a suggested reply. The agent reviews and decides; the model handles the groundwork. Community practitioners report reductions in average ticket handling time, though exact numbers vary by use case and model choice.
Contract and Document Data Extraction
A Qwen2.5 7B model can extract structured data from contracts, parties, durations, payment terms, termination clauses, and write directly to an ERP or database. Combined with a locally running Whisper model for transcription, the same infrastructure can turn meeting recordings into structured minutes automatically. See our overview of data sovereignty with local AI for more on the architecture.
CRM Enrichment and Lead Scoring
When a contact form is submitted, n8n can route the text to a local model, generate an industry classification and urgency score, and update the CRM before a salesperson sees the lead. No prospect data passes through a third-party AI service.
Model Recommendations for Automation Workflows
Not every automation task requires a frontier model. For most SMB workflows, smaller models deliver the necessary quality with low latency:
- Llama 3.1 8B (Meta, Apache 2.0): The community default for time-sensitive workflows. Reported throughput on a Mac Mini M4 with 16 GB unified memory: 30-60 tokens/second.
- Qwen2.5 7B (Alibaba, Apache 2.0): Strong multilingual performance, well-suited to German, Spanish, and French business documents. Similar hardware requirements to Llama 3.1 8B.
- Mistral 7B (Mistral AI, Apache 2.0): Reliable structured output and JSON generation, which n8n can consume directly in subsequent nodes.
- Llama 3.3 70B (Meta, Apache 2.0): For complex analysis, long documents, or tasks requiring deeper reasoning. Requires 64-96 GB unified memory; recommended on Mac Studio M3 Ultra or a dedicated GPU server.
All models are managed through Ollama and available via a single pull command.
Cost Comparison: Cloud Automation vs. Local Stack
A realistic comparison for an SMB running roughly 500 AI-powered automation steps per day:
Cloud stack (Make.com Business + GPT-4o mini via API)
- Make.com Business: approximately €29/month
- OpenAI API at 500 calls/day, approximately 600 tokens each: per community cost calculations, roughly €80-130/month
- Three-year total: approximately €3,900-5,800
Local stack (n8n self-hosted + Ollama + Mac Mini M4)
- Mac Mini M4 16 GB: approximately €900, one-time
- Power (~40 W, 10 hours/day): approximately €2.50/month across EU markets
- n8n Community Edition: free
- Three-year total: approximately €990
At moderate workload, the payback period is under twelve months based on community-reported comparisons. As automation volume grows, the advantage compounds: API costs scale linearly, hardware costs do not.
GDPR Compliance by Architecture
Cloud-based automation platforms route customer data, contract text, and internal documents through US-based infrastructure. That creates an obligation chain: data processing agreements, transfer impact assessments under GDPR Article 46, and reliance on a vendor's sub-processor list, which can reach dozens of entities for a major SaaS platform.
With a local n8n + Ollama stack, that chain does not exist. GDPR Article 25 (data protection by design) is satisfied architecturally: there is no network path by which data could reach a third party. This is not a contractual promise that depends on a vendor's compliance posture; it is a physical constraint imposed by the infrastructure itself.
For EU AI Act Article 50 transparency obligations, applicable from 2 August 2026, a local stack makes documentation straightforward: every model invocation, input, and output can be logged to a system you own and audited at any time without requesting records from a cloud provider. For a full overview of EU AI Act deployer obligations, see our article on local AI and regulatory compliance.
For SMBs operating in Spain or Germany specifically, this architecture also removes the need to navigate national supervisory authority guidance on international data transfers, a frequently shifting landscape that adds compliance maintenance overhead to cloud-based AI use.
What You Need to Get Started
Hardware
- Mac Mini M4 (16 GB): approximately €900, adequate for Llama 3.1 8B, Qwen2.5 7B, Mistral 7B
- Mac Studio M3 Ultra (192 GB): recommended for Llama 3.3 70B and parallel heavy workloads
- Existing Linux server with a recent NVIDIA GPU (12+ GB VRAM): also a viable option with vLLM or Ollama's GPU backend
Software
- Docker Desktop (free for individual use and qualifying SMBs, check current licence terms)
- n8n Self-Hosted AI Starter Kit (MIT licence)
- Ollama (MIT licence)
Setup time Community reports consistently place initial setup at 1-3 hours to first productive automation. n8n maintains a growing library of Ollama workflow templates that serve as starting points.
First model pull
ollama pull llama3.1:8b, approximately 4.7 GB download, then runs fully locally.
Where to Start
If you have been treating AI-powered automation as too expensive, too privacy-sensitive, or too technically demanding to deploy, the local stack addresses all three concerns simultaneously. Costs reduce to near zero in ongoing terms, GDPR compliance is structural rather than contractual, and the n8n starter kit and template library lower the setup barrier considerably.
The natural starting point is identifying your highest-volume, most repetitive document-handling task, email triage, invoice routing, support classification, and building a single workflow around it. The rest follows from there.
If you would like to map your specific workflows to a local AI stack, we are happy to walk through it with you, no commitment required.