Types of AI agents and which one your business really needs

Blai Lloveras Blai Lloveras
Updated:
elegir el tipo de agente de IA que necesita una empresa

One vendor offers you “an AI agent” to answer customer emails, another pitches a multi-agent system for the same job, and a third tells you an automation will do. All three use the word agent, and none of them explains what type they’re selling you or why.

The five types of AI agents the textbooks describe

The classification that almost every article repeats comes from Stuart Russell and Peter Norvig’s university textbook, Artificial Intelligence: A Modern Approach, in its chapter on intelligent agents. It ranks agents by how they make decisions, from least to most sophisticated.

The simple reflex agent responds to what it perceives at that moment with a fixed rule (if this comes in, do that) and remembers nothing from before. The model-based reflex agent keeps an internal state, what has happened so far, and uses it to decide when what it perceives isn’t enough. The goal-based agent no longer just reacts: it asks what will happen if it does one thing or another and picks the one that brings it closer to a goal. The utility-based agent compares several possible outcomes and keeps the one that suits it best. And the learning agent can be any of the above with one more piece, a mechanism that evaluates how it has done and adjusts its behavior.

This taxonomy helps you understand the mechanism, but it was designed for research, not for buying software. No vendor is going to sell you “a utility-based agent”.

The three types the market sells and what task each one is for

What an SME finds on the market boils down to three things. An assistant, such as ChatGPT, Claude or Gemini, answers when someone asks it something and doesn’t do anything on its own. A workflow agent is connected to your tools and carries out steps: it reads an email, looks up the CRM, drafts the reply and leaves it in drafts. A multi-agent system splits the work between several agents that hand tasks to each other: one researches, another writes, another reviews.

In customer service, the usual choice is a supervised workflow agent that sorts incoming queries, answers the recurring ones with information from your website or your orders and passes the ones it doesn’t understand to a person. If what you need is for your team to quickly find the returns policy, an assistant loaded with your documents is enough.

In sales, a workflow agent makes sense for qualifying leads that come in through a form, looking up the company, scoring it and leaving the record ready for the sales rep. The conversation with the customer still belongs to the rep.

a workflow agent that reads and prepares the company's emails

Back-office admin is, in our view, the area where AI is most often unnecessary. Posting invoices to the accounts or flagging an overdue payment are fixed rules, and an automation without a language model handles them just as well. AI comes in when you need to read documents that don’t follow a set format, such as invoices from different suppliers. For data analysis, on the other hand, the assistant pays off: you give it a sales spreadsheet and ask what has changed this quarter, as long as someone checks the figures before making a decision based on them.

How to choose: autonomy, integration and oversight

  • Autonomy. Does the task need to move forward on its own, without anyone kicking it off? If someone requests it every time, an assistant is enough. If it has to be triggered by a new email or an order, you need a workflow agent.
  • Integration. Does it need to read from or write to your systems, such as the CRM, the online store or the ERP? Every connection means setup and maintenance work, and what isn’t connected can’t be automated.
  • Oversight. What happens if it gets something wrong? A badly written draft that someone reviews costs nothing; a refund issued by mistake does. The more expensive the mistake, the closer a person needs to be before the action goes out.
a person reviews what the agent has prepared before sending it

Tools like n8n or Make let you build a workflow agent by connecting your apps and calling a model such as Claude or ChatGPT at the step that needs to read or write.

When you don’t need to move up a type

A multi-agent system sounds better in a sales proposal, but every agent you add is another point where something can fail, another model usage cost and another piece someone has to understand when it breaks.

Our rule is to always start with the simplest type that does the job, and only move up when it falls short. If the task always follows the same steps and the data always arrives in the same format, Zapier, Make or n8n without any language model will do it more cheaply and more predictably, and they don’t make anything up. If you need to understand free-form text, you add AI at that step and only that one.

A multi-agent system is justified when the task has genuinely different stages, with different criteria, and a single instruction doesn’t produce good results. In an SME that happens less often than you’d think, and we’d rather tell you so before quoting for it.

If you have a task in mind and aren’t sure which type it calls for, tell us about it and we’ll tell you what level will do.

About the author

Blai Lloveras

Articles by Blai

Let's talk

Do you have a project in mind?

Tell us about it and let's grow your business together.

Analyze my sales opportunities