Marketing operations teams are dropping a language-model node into n8n or Make and calling the result an agent. A form arrives, the model writes a summary, and a HubSpot or Salesforce step updates the record. That is a useful automation. It is not a system that can plan, choose tools, and stop itself. The distinction matters the moment the flow can enroll a sequence, change a lifecycle stage, or create a task a rep will act on. Those writes need a contract, not a prompt that improved over a weekend.
The workflow should own the trigger, the identity of the record, and the fields that are allowed to change. The model should own language: classify the note, draft the follow-up, extract only the fields the workflow already named. When those jobs blur, the flow treats model output as permission to write fields nobody approved. A schema on the model node is not governance if the next step maps whatever came back onto the CRM. Pin the outputs, and refuse the run when a required field is missing instead of guessing.
Before that flow calls anything outside the CRM, prove the tool with the same inputs production will send. An MCP server that looked fine in a notebook fails on auth, empty properties, and rate limits the first time Make runs it against a real record. Use an MCP tester for the happy path and for the empty and duplicate cases. A passing call shows that the tool answers. It does not show that the CRM step should be connected. The write stays behind the workflow's own checks, including suppression and the fields a human still has to confirm.
Give the model a fixed job, not a blank prompt, if it sits inside that flow. A skill states the task, the tools it may call, and the shape of the answer. The workflow invokes that skill. If the skill is vague, the node improvises, and improvisation is how a contact gets a second owner or a stage moves backward. Keep the skill as narrow as the workflow step. A broad skill attached to a CRM write is how a helper becomes an unreviewed editor of pipeline.
Judge the stack by the writes, not by how clever the node looks. Count the records the flow touched, how many a person reversed, and which tool calls failed. If reversal is common, the model is deciding something the workflow should have gated. Hybrid is the durable design: automation for routing and suppression, a tested tool for anything external, a narrow language step, and a person on anything that commits the company. A raw CRM update driven by an open prompt automates the mistake at the speed of the queue.
Test the tool, pin the language step, and only then let the workflow write. Use the MCP tester and the skill builder, then the rest of the tools. We wire hybrid agentic stacks (automation + MCP/skills) into your CRM — Book a Consultation. If you want the write steps reviewed before they run unattended, email hello@nexamartech.com.
Top comments (0)