AI agents for business: what they are, what types there are and what they do

“AI agent” is what almost anything gets called today: a chatbot with its own name, a four-step workflow and a system that does real work. When a word can mean anything, it stops being useful for making decisions.

This is what an agent is when you look at it from inside a business, not from the developer’s side: what it does, how it differs from what you already have, what types exist and when it isn’t worth building one.

What is an AI agent

An AI agent is a system that receives a goal, decides for itself which steps to take to achieve it and carries them out using the tools and information you’ve given it access to. The difference from a traditional automation lies in the decision: the automation always follows the same sequence, and the agent chooses the sequence according to what it finds.

Three elements define it, and it’s worth having them to hand the next time someone shows you a demo:

• A goal, not a one-off instruction. “Prepare the file for this tax request” rather than “summarise this PDF”.
• Tools it can call on. Your email, your ERP, a network folder, a website, a database. Without access to anything, an AI model can only write text.
• A bounded scope for decisions. What it can do on its own, what it has to check and where it stops.

If any of the three is missing, what you’re looking at is something else. It may still be useful, but it’s worth knowing what you’re going to pay for.

An agent isn’t ChatGPT under another name

ChatGPT and similar tools work in turns: someone writes, the model replies, and that’s it. The work is still done by the person, who now just writes faster. It’s useful, and it’s where most of the businesses that say “we already use AI” are today.

An agent changes who does the work. It starts up without anyone opening it, goes into the systems where the information lives and leaves the work done, or ready for someone to approve. That’s the real boundary, and it isn’t set by the technology: it’s set by whether the result depends on a person sitting in front of it.

That’s why a business can spend two years with AI licences for the whole workforce and not have changed a single process.

Automation or agent: how they differ

The most expensive confusion isn’t between ChatGPT and an agent. It’s between an automation and an agent, because both do work without anyone watching and they cost very different amounts.

The rule of thumb fits on one line: if you can draw the whole process in a diagram without writing the words “it depends”, you don’t need an agent. You need an automation, which is cheaper, quicker to build and easier to fix when it fails.

The reverse also holds. If every time you explain the process three exceptions and a “Marta handles that one, she knows how” crop up, automating it the old-fashioned way produces a system that skips half the cases.

What types of AI agents are there

If you search for this question, the first thing you find is an academic classification: reactive, deliberative, goal-based, utility-based and learning agents. It’s correct, and it doesn’t help you decide anything. It describes how the system reasons on the inside, when what you need to know is what work it does on the outside.

This is the classification we actually use when designing, ordered by what the agent does:

• Capture agents. They read what comes in (emails, invoices, forms, case files in PDF) and extract organised, usable information from it. It’s the most common type in an SME because the volume is high and the work is pure paper-shuffling.
• Execution agents. They do a complete task from start to finish within a set of rules: they gather, cross-check, draft and leave it ready for someone to review and sign. This is where the big savings are, and where it matters most where you set the limit.
• Query agents. They answer questions about the business’s documentation: contracts, procedures, project history, internal rules. They’re useful when the knowledge exists but is spread across folders nobody remembers.
• Monitoring agents. They watch something that changes and flag it when needed: a tender portal, an inbox, the deadlines on a case file, a supplier’s prices. They work while nobody is watching them, which is exactly what a small team can’t afford to do.
• Coordination agents. They don’t do work: they hand out work to other agents and check it has been done. They only make sense when several are already running, and building them earlier is like drawing up an organisation chart for a one-person workforce.

There’s a second axis that matters more than the list above: how long a leash the agent is on. The same system can propose and wait for approval, act and report afterwards, or act without saying anything. Our approach is always to start with a short leash and let it out in stages, as each part proves itself. Not as caution for show: the first months are what show you where the criterion you wrote down fails.

Examples of AI agents in a business

A tax case file at an advisory firm. At JGM Consultores, requests from the Spanish Tax Agency took around three working days each and arrived in bursts. The system doesn’t reply or sign: it gathers the documents, cross-checks the data, spots what’s missing and leaves the file assembled for the professional to decide. The measured result: from 24 hours to 2 hours per case file, and from €480 to €10 in operating cost.

What made that number possible wasn’t the AI model. It was having first separated out which part of the work was professional judgement, which stays with the person, and which part was moving information from one place to another. Without that cut, the project would have tried to automate the whole case file and would have failed precisely in the part that matters.

Reviewing what the automations produce. At Grupo Platea, human review of the automated work took between four and five working days. Today it’s around an hour a day. It’s an unglamorous and very representative example: the work that appears when you automate something can also be reduced, and almost nobody mentions it in the initial proposal.

Tender tracking. A system that tracks calls for tenders, discards those that don’t fit the company’s profile and prepares the repetitive paperwork for those that do. It went live at a client this month, after two days of implementation. There’s no savings figure yet and we’re not going to make one up: it will be published when it exists.

Notice that none of the three is an agent that “runs the business”. All three handle a specific process, with a written limit and a person who decides at the end. That’s what this really looks like when it works.

When an AI agent isn’t the answer

There are four situations in which building an agent is spending, not investing. Saying so beforehand works out cheaper than finding out afterwards.

When the process always runs the same way. There, the ability to make decisions is surplus to requirements, and you pay for it anyway. A well-built automation is more reliable and costs a fraction.

When it happens three times a year. Without volume no return is possible, however irritating the task. Document it properly and keep doing it by hand.

When the criterion isn’t written down anywhere. If the decision lives in the head of someone who’s been with the company for fifteen years, the first deliverable isn’t an agent: it’s putting that criterion in writing. That work is valuable in itself, because the business stops depending on that person being available.

When nobody is going to supervise it. An agent with no one responsible for it isn’t a saving, it’s a risk working office hours. And if the process is broken to begin with, automating it only multiplies the mess: that’s why it’s worth looking first at which business processes to automate and which aren’t ready.

If you don’t know which of these four your case falls into, that’s exactly what we look at in the free AI assessment: whether there’s a real lever, what kind of solution fits and what would be needed before building anything.

What to ask before hiring an AI agent

Five questions. The answers separate those who have implemented this from those who have seen a demo.

1. What does it decide on its own and what does it check? If there’s no written answer to this, there’s no design. There’s an intention.
2. What happens when it gets something wrong? Good systems aren’t the ones that never fail: they’re the ones that fail visibly and recoverably. Ask for the exception plan, not the promise of accuracy.
3. Where does my information live and who can see it? What data leaves your business, where to, with which provider and on what terms. If the answer is vague, the question was the right one.
4. What am I left with if we stop working together? The code, the written criterion and the accounts should be yours. The monthly service is paid because it keeps the system up to date and brings improvements, not because without the provider the whole thing switches off.
5. What number are we going to use to measure whether it worked? A starting figure and a finishing figure, for a specific task. Without a starting figure, a year from now there will be no way to defend the investment to anyone.

A business that can’t answer the fifth doesn’t need a provider yet. It needs to measure a process.

Frequently asked questions

What is an AI agent in simple terms?

A program you give a goal rather than a list of steps, and which decides for itself how to achieve it with the tools it has access to. An example: instead of asking it to draft an email, you ask it to prepare a customer’s complete file, and it looks for the information, organises it and leaves the draft ready for someone to review.

What is the difference between an AI agent and a chatbot?

The chatbot converses and the agent acts. A chatbot replies to whoever writes to it and stops there; an agent goes into your systems, does the task and leaves a result. Many products sold as agents are chatbots with access to a search engine: the test is to ask which complete task it does without anyone writing anything.

What are the best AI agents for a business?

There’s no useful ranking, and whoever gives you one is probably selling the platform underneath. Models and tools change every few months and all the serious ones work well; what decides the result is how well the process is defined, whether the criterion is written down and whether someone supervises. The best agent for your business is the one that solves the process that costs you most today, and that one isn’t on any list.

How much does it cost to build an AI agent for an SME?

It depends on three things, and anyone who gives you a price without asking about them is improvising: how many systems need connecting, how many decision rules have to be written from scratch and how much supervision the process needs. What’s worth being clear about is that the cost doesn’t end with the implementation: an agent that touches live processes needs a monthly service to keep it up to date (servers, updated models and the AI costs, which aren’t zero), just like any other system in the business.

Do I need a technical team to have an AI agent?

To use it, no. To maintain it, someone has to take responsibility, and they don’t need to be technical: they need to know the process and have the authority to say how it’s done from now on. Implementations that fail almost never do so for lack of technical knowledge, but because nobody inside made the decision to change the way of working. That’s where training the team matters more than the tool.

Is it safe for an agent to access my business’s data?

It depends on how it’s built, and it’s a question you should demand an answer to in writing before signing: what data leaves, to which provider, whether it’s used to train models and how long it’s kept. There are architecture decisions that change the answer a lot, from using models with a data processing agreement to processing the most sensitive material locally. What won’t do is the generic answer “it’s secure”.

In short

An AI agent isn’t a product category you buy: it’s a way of solving a process that has exceptions and judgement built into it. When the process doesn’t have them, there are cheaper and more reliable solutions, and saying so is part of the job.

What separates an agent that’s still working two years from now from one that switches off after three months isn’t the model underneath. It’s having decided, before building, what it does on its own, what it checks, who answers when it gets something wrong and what number it’s measured by.

If you have a process in mind and want to know whether it needs an agent, an automation or neither, tell us about it. We’ll tell you what we see, and if we don’t see a real fit, we’ll tell you that too.