Why every B2B marketer needs to understand n8n
LLMs, MCPs, and AI agents: what each layer actually does, and how n8n ties them into workflows that replace costly human capital where it counts.
Most of the expensive parts of a B2B marketing stack are data transformation, not judgment. You pay a human to look up a prospect, write a version of an email, paste it into a sequence, and log the result. n8n is where that pipeline lives. Here’s what the three layers actually do, and where each one earns its cost.
The three layers
The LLM is the language layer. It reads, thinks, and writes the way ChatGPT or Claude does: drafting email copy, scoring a lead’s intent from a form response, summarizing a prospect’s LinkedIn profile. It understands and generates language, but on its own it only responds. It doesn’t act.
MCP is the connector layer. It’s how an LLM reaches into external tools and data sources, your CRM, your browser, a file system, through one standardized protocol instead of custom glue code for every integration. It’s what lets Claude or GPT pull a live Salesforce record or read a webpage mid-conversation.
The agent is the layer that acts. It wraps an LLM with the ability to plan and take action across multiple steps: decide what to do next, call a tool, check the result, keep going until the job is done. In n8n, an agent node is the orchestrator. It might receive a new inbound lead, look them up through Apollo, research their company, draft a personalized outreach email, and log the result to Salesforce, without you touching any of it.
Matching the model to the job
An n8n workflow can include as many or as few of these layers as the job actually requires. If a step needs real reasoning, a heavier model like Opus costs more in tokens and latency, but it earns that cost: the decision gets weighed properly instead of guessed at. Hand that same judgment to a smaller model to save on API costs, and you’re trading reliability for a lower bill, in a workflow a rep or a prospect is about to see the output of.
The workflows worth trusting are the ones where the model’s power matches what the step actually asks of it, applied to input data that’s actually worth transforming, in the places where a human doing it by hand would cost more than the API ever will.