n8n + AI Automation

In n8n the choice is a chat-model sub-node plugged into an AI node, so it is cheap to change: swap the Anthropic, OpenAI, or Mistral Cloud model node and re-enter credentials, and the rest of the workflow stays as it is. Claude is the default because classification and extraction steps reward careful instruction-following, and long context helps when a step reads a full email thread or a multi-page document. OpenAI makes sense when you rely on community nodes or templates built around it. Mistral is the privacy and cost pick: its open-weight models can run on your own hardware through n8n's Ollama model node, so a self-hosted n8n plus a local model keeps every document inside your network. Mixing is common too: a small, cheap model for routing and a stronger one only for the steps that write text a person will read.

Start with the built-in nodes. The AI Agent, Basic LLM Chain, Text Classifier, Information Extractor, and Summarization Chain nodes cover most of this stack's use cases, handle credentials and chat-model swaps for you, and produce output the next node can read without hand-written parsing. The HTTP Request node is the fallback for a provider feature n8n's nodes don't expose yet, such as a newly released model parameter or a batch endpoint. The cost of going that route is that you own the request body, the response parsing, and any change the provider makes to its API, which is exactly the maintenance the built-in nodes absorb.

This is the gotcha that bites most AI workflows: an LLM returns text, but the next node (a database write, an IF branch, a CRM update) needs predictable fields. Use the Information Extractor node with a defined schema, or attach a Structured Output Parser to a chain or agent node, rather than asking for JSON in the prompt and hoping. Then plan for the occasional miss: enable Retry On Fail on the AI node, and set an Error Trigger workflow that parks failed items somewhere a person will see them instead of letting a half-parsed record flow downstream. For anything customer-facing, such as drafted replies, keep a human approval step before the send.

Token spend, almost always. n8n itself is flat: self-hosted execution is unlimited, and n8n Cloud counts one execution per workflow run no matter how many nodes it contains, so a 12-step AI workflow costs the same as a 2-step one there. The LLM bill instead grows with every item processed and every token in the prompt. The main levers are filtering before the AI step (an IF node that skips emails that don't need classifying), trimming what gets sent (the relevant fields, not the whole record), and using a cheaper model tier for routing and tagging steps. The Pricing section covers the per-provider figures.

This stack is headless: workflows start from events (an incoming email, a webhook, a schedule) and end in other tools, with no user interface of its own. The n8n + Streamlit AI Agent stack adds a Python front end where people type a request and get an answer back, with n8n running the logic behind a webhook. Pick this one when the AI step should happen in the background without anyone asking for it; pick the Streamlit variant when people need to query the workflow on demand. The n8n workflows themselves carry over either way, so starting here and adding a front end later is a small step.