An AI agent is software that works towards a goal by choosing its own steps, using tools such as your inbox or CRM, and adjusting based on what it finds. A chatbot answers. An agent takes action. It only belongs in your business with clear limits on what it may do.
What makes software an agent?
The defining feature is who decides the next step. Anthropic draws the line this way: in a workflow, the steps are fixed in code. In an agent, the language model directs its own process and its own use of tools. Everything else follows from that.
In practice an agent has four parts:
- A goal. For example: handle this incoming request and leave the customer record up to date.
- Tools. The systems it may read from and write to, such as your mail, CRM or document store.
- A loop. It acts, looks at the result and decides what to do next, instead of running once.
- Limits. Permissions, stopping conditions and moments where a person has to approve.
How do you tell if something is really an agent?
Ask three questions of any product that carries the label. Vendors have reason to use the word loosely. Gartner has a name for relabelling assistants as agents: "agentwashing".
| Question | Agent | Probably something else |
|---|---|---|
| Who picks the next step? | The model, within limits | A fixed flow someone designed |
| Can it change something in another system? | Yes, through its tools | No, it only writes text |
| Does it continue after the first instruction? | Yes, until done or stopped | No, it waits for your next prompt |
Failing this test doesn't make a product bad. A fixed flow is often the better choice. You just want to know what you're buying. See AI agent vs AI assistant for the most common mix-up.
When does your business need one?
Less often than the market suggests. Anthropic's own advice is to find the simplest solution possible and add complexity only when needed. It also notes that agents often trade cost and latency for better task performance.
Our rule: an agent earns its place when you can't write the process down without saying "it depends". The input is free text, the next step depends on the content, and the answer sits across several systems. If you can write the steps as a plain list, build an automation instead.
What can go wrong?
- Errors stack up. Each step builds on the one before, so an early mistake travels.
- Too many permissions. Give it the access this one job needs, not a general login.
- Nobody looking. Without logs you can't see why it did what it did.
- Untrusted input. An agent reading outside mail or documents can meet hidden instructions. OWASP lists this as prompt injection.
That's why a good agent isn't the most autonomous one. It's the one with a person in the right place in the process.
What does it take to build one?
Start with one process and a definition of done, then choose which process comes first. At 5DAYS a Business Agent starts from €7,950, excluding VAT, and custom agents start with a technical Discovery of €950. External AI and usage costs come on top: see what AI automation costs.
