01What an AI agent is, in practice

In a business, an AI agent is a system that does not stop at the answer: it takes a bounded goal, chooses the steps and uses the tools it has been allowed to use. It may read an order, check stock, draft a reply or leave a note in the CRM. What defines it is not a product name, but that it plans and acts inside a perimeter someone has agreed in writing.

That perimeter matters as much as the model. Without it, the system improvises with data it should not see, or takes a step nobody authorised. With it, the agent works inside a fence: these tools, this data, these actions, and this approval queue when something is sensitive. The value is not in «letting the AI decide», but in bounding what it can and cannot do.

You also do not need an agent for everything. Many tasks in a small business are solved with a fixed-rule workflow, or with an assistant that answers and hands over to a person. Before commissioning an agent, it is worth naming the process and asking whether steps really need to be planned, or whether a clear rule is enough. That judgement is part of the diagnosis I do in the AI integration service.

Fig. 1 · How it differs from a chatbot

02How it differs from a chatbot

A chatbot —or a conversational assistant— is built for dialogue. It answers a question, follows a script or a knowledge base, and, if the query falls outside what was planned, it should know how to hand over to a person. That is already serious work: you have to decide what it may say, when the conversation is cut, and how you make it clear that the other party is speaking to a system, not to a human.

An agent, by contrast, does not stop at the reply. If the goal is «prepare the next step for this order with a problem», it can check the status, draft a notice and leave it ready for someone to send. The chatbot finishes when it has answered; the agent finishes when it has carried out, or prepared, the steps that fit its perimeter.

The confusion comes from using «agent» as a synonym for a more capable chatbot. It is not. Adding a newer model to a WhatsApp assistant does not turn it into an agent. If what you need is an assistant on that channel, with a handoff to a person and what Article 50 of the EU AI Act requires, I cover it in the guide to AI assistants on WhatsApp for business. This guide stays on the difference in approach, not on that channel’s flow.

03Chatbot, automatic workflow and agent: three different things

Besides the chatbot and the agent, a third kind of system often gets mixed up with the other two: the automatic workflow. A workflow does not converse or plan. It follows a rule someone has written: if an order arrives in a given state, notify this inbox; if an appointment is tomorrow, send a reminder. It does not choose the next step: it runs it because it is in the rule.

The three can coexist, and there is no hierarchy that says the agent is always «better». A booking reminder rarely needs an agent. A website assistant can handle repeated questions and pass the rest to a person. An agent makes sense when you have to chain steps, query more than one tool and leave a person to approve what leaves the fence. In the guide to automating orders, bookings and CRM I go into how to pick the first process; here it is enough not to mix the three pieces.

What it solvesWhat it does not doWhen it makes sense
Chatbot or assistantAnswer questions and, if needed, hand over to a person.It does not plan a chain of actions or write into your systems on its own.Repeated queries on the website or a messaging channel, with a clear handoff.
Automatic workflowRun a fixed rule when a condition is met.It does not choose a different step if the case falls outside the rule.Alerts, reminders and predictable handoffs, with little room for interpretation.
AgentPlan and carry out steps with tools, inside a perimeter.It does not replace judgement and should not send anything sensitive without agreement.Internal or operational tasks where you need to look something up, prepare it and leave the next step ready.
Fig. 2 · What an agent can do in a business (and what it cannot)

04What an agent can do in a business (and what it cannot)

In practice, a well-bounded agent usually prepares internal work. It can summarise an email thread, propose a draft reply, leave a record in the CRM, flag a problem with an order, or gather the information a person needs in order to decide. That saves the back and forth between tools, not the decision.

What it should not do, unless it is agreed explicitly and the risk is low, is speak to a customer in your name, change an order, apply a discount or close a deal. Those actions have consequences that are not improvised in a prompt. The starting point I use is that the agent prepares and a person approves anything sensitive.

It is also not a substitute for a procedure that does not yet exist. If nobody in your team knows what to do with an order exception, the agent will not invent that judgement. First you describe the process; then you decide which part a system can prepare. Without that order, you automate the mess.

05How it is bounded: perimeter, permissions and supervision

Bounding an agent is a design job, not a setting you add at the end. In the diagnosis you write down its goal, which tools it may use, which data it may read, which actions it may leave ready, and what stays blocked. You also define the approval queue: who reviews, in what time, and what happens if the case leaves the perimeter —for example an order with an amount or a customer type that was not planned for.

Permissions matter as much as the instructions. An agent that can read the entire customer history, or write to any field in the CRM, is a risk even if the prompt is cautious. You start from the minimum needed, you test with real cases —not only with tidy examples— and you correct it when it fails. Failure is not improvised in production: it is rehearsed first.

Supervision is not decoration. You need to see what the agent did, with which data and why it chose a step. If something goes wrong, someone in the team must be able to stop it and correct it. That record, and the agreement on who maintains the system after delivery, are part of the scope closed in writing. I do not certify the legal compliance of a design; I do leave the perimeter, the permissions and the handoff to a person documented.

If it is a fit, the next step is a 30-minute meeting or an email to hello@diegosanjuan.com, and a written diagnosis before the quote. Contact is on the contact page. Scope is not improvised on the day the system goes into use.