Building an AI Agent Workflow: A Real Example with n8n

2025's fashionable word: the AI agent. By the definition everything is an agent, while in practice what most people build is ordinary automation. Knowing the difference matters, because half of those wanting to build an ai agent in truth need not an agent but a simple flow; and the other half lets the agent loose without control and runs into trouble. This article both seats the concept in its place and builds a real example on n8n.

The plan: the agent-versus-automation difference, n8n's place, the step-by-step real scenario (processing incoming enquiries) and the control rules.

What is an agent, and how does it differ from automation?

Classic automation is a rigid script: "when A happens, do B"; every step is written in advance. An agent is a flow with a decision layer: the model gets a goal and tools, and picks the steps itself by the situation ("this message is a complaint → create a record + write to the manager; this is a sales enquiry → prepare an answer from the price table"). The simple formula: automation = the route, the agent = the driver. The rule that follows: if the process is fully predictable (invoice arrived → save to the folder), an agent is excess: the plain flow is both cheaper and more reliable; if decision variability exists (a different road by the message type), the agent layer adds value. Recall the LLM's nature: the model predicts; every degree of freedom you hand the agent is a unit of both power and risk.

n8n: why this tool?

n8n is a visual workflow builder: you connect boxes with lines, the code drops to a minimum. The reasons for the choice: it is open-source (it can run free on your own server; a cloud plan exists too), the AI nodes are ready (the model call, the agent layer, the memory get added like a box) and the integration base is wide (WhatsApp, Telegram, spreadsheets, CRMs, email: hundreds of ready nodes talking through APIs). Alternatives exist (Make, Zapier: simpler, subscription-based), but in AI-agent scenarios n8n's flexibility and its self-hosting option (the data control) make it this article's example tool.

The real scenario: the enquiry-processing agent

This structure's beauty is its staging: in the first version step 5 passes everything through (all of it approved); as the system proves itself, the low-risk answers (the opening hours, the address) get released automatically. The engine of the bot+human division from the bot article looks exactly like this.

The control rules: the agent's leash

• The human gate is the standard: every step touching money, obligations or the customer relationship passes through approval; the fully autonomous agent is the exception, not the start.
• The log is compulsory: what the agent decided, why (the classification result), what it did: all of it gets logged; when a problem appears, the "what happened" question must not go unanswered.
• Set the boundary: keep the agent's tool list minimal (only the needed integrations); "an agent with access to everything" is a design mistake.
• The cost ceiling: an agent caught in a loop (self-repeating) can burn the API cost; the step limit + the monthly budget ceiling is the technical insurance (the cost management).
• The test set: the periodic check with 20–30 typical messages; the agent version of the hallucination control.

Sources and further reading

Where to verify the source

The tool documentation:

• The n8n documentation: the nodes and the AI capabilities

Continuing the topic

The automation line's neighbouring articles:

• The automation map
• The answer systems
• The knowledge base layer
• The API foundation
• Other articles on this topic

The first step this weekend: open an n8n account and build the simplest flow: "when a new email arrives, let the AI summary drop into Telegram". That two-hour experiment is the agent world's door; the big systems are born of exactly these small boxes being connected.