What is an AI agent?
An AI agent is a system that takes a goal, decides which steps to take, and calls tools or APIs to carry them out, checking its own results as it goes. A chatbot talks. An automation follows a fixed path. An agent picks the path.
Reviewed:
In short
If it cannot choose between two different actions, it is automation wearing the word "agent". The test is not whether a language model is involved. The test is whether the system decides.
| Chatbot | Automation | AI agent | |
|---|---|---|---|
| Chooses its own next step | No | No | Yes |
| Calls tools and APIs | Rarely | Yes, a fixed set | Yes, picks which one |
| Handles an unplanned case | Says it cannot help | Fails or skips the record | Tries another route, then escalates |
| Runs with nobody watching | No | Yes | Yes, with approval steps |
| Typical trigger | Someone types | A schedule or webhook | A goal, from either |
| Typical failure | A wrong answer | A stuck queue | A confident wrong action |
Sources: Anthropic: building effective agents · n8n docs: AI agent node · reviewed: May 28, 2026
Where it shows up in practice
Support triage
A ticket arrives. The agent reads it, checks the order in your system, decides whether it is a refund, a delivery question or a bug, and either answers or routes it. The routing rule was never written down; the agent infers it from the ticket and the order record.
Lead qualification
A form comes in. The agent enriches the company, scores fit against your criteria, and books a meeting or sends a polite no. When the enrichment lookup fails it tries a second source rather than dropping the lead.
Reporting
Every Monday the agent pulls numbers from ads, CRM and invoicing, notices which ones moved, and writes the summary about those. A scheduled report would send the same twelve charts whether or not anything changed.
Where an agent is the wrong answer
Moving a paid order from a shop into an accounting system is one path with no decisions in it. An agent adds cost, latency and a way to be creative about your invoices. Use a plain automation.
How we use it
The word got stretched in 2026 until it stopped carrying information. We have been shown a cron job with an API key and told it was an agent, and we have seen genuinely autonomous multi-step systems described as "just some automation" by the team that built them. Both descriptions are useless to a buyer.
The line we hold is decision-making. If the system can reach the same goal by more than one route and chooses between them, it is an agent, and it needs the things agents need: logs of what it decided, limits on what it can touch, and a human approval step wherever a mistake costs money. If it runs one path every time, it is an automation, and it needs none of that overhead. Most of the projects that land on our desk asking for agents turn out to want two automations and one agent. That is a cheaper, more reliable system, and we say so before quoting. See what we actually build on the AI agents page, or read the buying-side view in what an AI agency actually does.