'What Is an AI Agent? Definition, Examples and Limits'
An AI agent is software that takes a goal, works out on its own which steps are needed, uses tools such as APIs and databases to carry them out, and keeps iterating until the goal is reached. The difference to classic automation: you define the outcome, not every step. The agent picks the route.
That is the whole definition. No buzzword needed. The rest of this article makes it concrete: how you spot a real agent, how it differs from a chatbot and a workflow, what an agent looks like in our own operation, and where the limits are.
The four traits of an AI agent
Software is an AI agent when it can do these four things:
1. Understand goals, not just commands. "Process this return" instead of "run steps 1 through 7".
2. Decide on its own which steps are needed and in what order.
3. Use tools: call APIs, look things up in databases, send emails, drive other software.
4. Iterate: check whether it worked and try a different route when something breaks.
Miss one of them and it is not an agent. A language model that only replies is not an agent. A workflow that runs fixed steps isn't one either.
Agent, chatbot, workflow
The three terms get mixed up constantly. The short version:
- Workflow: fixed steps, hard wired. "When an order comes in, send this email, create this label." Every step is predefined. When reality deviates, everything stops.
- Chatbot: answers questions in a conversation. Classically with fixed answer trees, today with a language model. It talks, but it doesn't act inside your systems.
- AI agent: gets a goal, plans on its own, acts in your systems and checks its own result.
In practice you usually need a combination of all three. Which task fits which approach is covered in detail in our comparison of AI agent, chatbot and workflow.
A picture for it: the workflow is an assembly line, the chatbot is an information desk, the agent is a case worker with access to every system.
Where the difference shows up
| Scenario | Classic workflow | AI agent |
|---|---|---|
| Standard case | runs perfectly | runs perfectly |
| Slight deviation | error or stop | adapts |
| Completely new case | can't do it | tries, with risk |
| Traceability | 100 percent | limited |
| Cost per run | very low | higher, depending on the model |
The last row gets forgotten often: an agent calls a language model on every run, sometimes several times. What that means per month is in the article on the cost of automation.
A real example: our support agent
Theory is cheap, so here is an agent from our own operation. We run 3 e-commerce brands with 2 people in day-to-day business. That only works because an agent handles customer service.
This is how a request runs through:
1. A customer writes: "My order hasn't arrived and I'm leaving on holiday on Friday."
2. The agent reads the email and identifies the intent: a WISMO request, meaning "where is my order". Those questions make up around 40 percent of our support volume.
3. It looks up the order through the Shopify API, pulls the tracking status from the carrier and checks the promised delivery date.
4. It writes a reply with the real parcel status and the expected delivery date.
5. Special case detected: the delivery date falls after Friday. The agent offers to change the delivery address and puts the case up for review, because address changes need a human sign-off here.
Steps 1 through 4 run entirely without us. In total, the AI answers around 65 percent of our support requests automatically. The full setup is in the guide to customer service automation.
One thing matters: the agent works inside a frame. It may give information and solve standard cases. Refunds, address changes and angry customers go to humans. An agent without boundaries is a risk, and it stops being a tool.
What AI agents can do today
Reality, no science fiction:
- Answer customer requests, with access to the order system, FAQ and policies
- Pull data from several sources and summarise it
- Create content, based on guidelines and context
- Make decisions, inside defined boundaries
- Write code and fix errors
What they CANNOT do
The list that matters, because this is where most projects fail:
- Be 100 percent reliable. Agents make mistakes, and different ones than classic software. A workflow fails loudly, an agent sometimes fails convincingly and quietly.
- Take responsibility. You are liable in the end. That is why critical decisions belong behind a human sign-off.
- Know what you don't give them. An agent only knows the systems and data you connect.
- Be arbitrarily cheap. Every run costs model calls. For simple if-then cases a workflow is faster, cheaper and easier to trace.
When an AI agent makes sense
Yes, when the task has many variations, when human judgement helps but isn't critical, and when a classic workflow would get too complex.
No, for critical financial processes without control, for legally binding decisions, or when a simple workflow does the job. Our own order processing deliberately runs as a classic workflow: reliability counts there more than flexibility. The skill is knowing which is which.
Frequently asked questions
Is ChatGPT an AI agent?
No, ChatGPT on its own is a language model with a chat interface. It answers, but it doesn't act independently in your systems. An agent only appears when a language model is connected to tools, goals and a loop for checking and correcting.
What does an AI agent cost in e-commerce?
Running costs consist mainly of model calls and hosting, which for our support setup is a low three-digit amount per month across all brands. The build is the expensive part: connecting systems, rules, tests. Budget several days to several weeks, depending on scope.
Does an AI agent replace support staff?
It replaces tasks. The people stay. Here the AI answers around 65 percent of requests automatically, humans take the rest. The time freed up goes into cases that need empathy and judgement.
Do I need programming skills for an AI agent?
For simple agents, tools like n8n with ready-made AI blocks are enough. As soon as the agent has to act across several systems and escalate special cases cleanly, you need technical understanding or someone who brings it.
Wondering where an agent would make sense in your shop? Get in touch, and we will also tell you when a simple workflow is enough.