Chatbot, agent and n8n are not the same thing. This note is how to pick and what to ask for: a bounded case, a perimeter, logs and who operates it.
Three different things
- Chatbot / RAG. Question and answer over documents. It does not write to the CRM. It does not send email. It is right when the problem is finding and citing.
- Agent. A model that picks tools: query, classify, open a ticket, call an API. It needs a perimeter and a log of each tool.
- n8n or Make. A flow: if this happens, do that. The model, if it appears, is a step. Not the owner of the process.
If the work is “when this email arrives, create a record and notify”, n8n or Make are usually the core. If the work is “the team asks the wiki”, the core is RAG. If the work is “classify and, by type, run a flow”, an agent or a model inside n8n. Mixing all three in the same case usually leaves a system nobody can operate.
What to ask for
- A bounded case: one mailbox, one document type, one channel. Not “the company”.
- Input and output: where it reads, where it writes, what it must not touch.
- A log: each run with input, tools, result and error.
- A visible failure: retry, alert, and what the human does if the flow stops.
- An owner: who deploys it and who changes the prompt or the graph.
- Human review where the flow sends, charges, deletes or writes to the CRM. See n8n on email or CRM.
What has to be written down
- It runs with a company account, not only the vendor account or a laptop.
- Answers carry a source, or the agent has a tool list.
- n8n lives where the team can edit it. Secrets sit in the environment store.
- There is a log of each run and an owner for the deploy.
Python, FastAPI, Node, LangChain or LangGraph come in when the flow does not fit in n8n. They are not chosen to look more like AI. Madrid. If you want me to set it up: RAG and agents.