AI agents for business: what they can and can't do alone

Imagine a building renovation firm in Girona with six people. Every morning twenty emails come in: three ask for a quote, one chases an invoice, two are from suppliers and the rest is advertising. Someone reads them one by one, checks whether the customer is already in the system, looks at the calendar and replies. An AI agent can do that first pass: read each email, check whether the customer already exists, suggest a time for a site visit based on the calendar and leave the reply written in the drafts folder. The person just reviews it and presses send. That is what AI agents for business can already do. The difference from a chat tool is that an agent touches your software, which is why what you let it touch matters so much.

What an AI agent is, without the jargon

An AI agent is a program that uses a language model, the same kind of artificial intelligence behind ChatGPT, Gemini or Claude, to reach a goal by taking several steps on its own. The Spanish Data Protection Agency (AEPD) explains it in its guidance on agentic AI, published in February 2026: these are systems that, as well as answering questions, act autonomously to achieve what they have been asked to do. That's where the name agentic AI comes from.

Inside, it has three parts. OpenAI's practical guide to building agents calls them model, tools and instructions, and in a business they look like this:

• The model reads, reasons and decides what the next step is.
• The tools are what it can touch: email, the calendar, a spreadsheet, the CRM or the invoicing software.
• The instructions say what it has to do, how, and what it must never do. They're a carefully written prompt that doesn't change from day to day; in how to write a good prompt we explain how to write one.

Agent, chatbot and automation: how they differ

All three use similar technology and are sold under similar names, but they don't do the same thing. You can see it with an email from a customer asking for a quote to redo a bathroom.

A chatbot answers

A chatbot holds a conversation: it answers with the information it has. It can explain what renovation work you do and how to ask for a quote, but it doesn't open your calendar or record anything anywhere. OpenAI's guide is clear on this: a simple chatbot is not an agent. When one helps and when it gets in the way is covered in chatbot for your website.

An automation always follows the same path

An automation runs fixed steps: if an email arrives with the word "quote" in the subject line, it files it in a folder, creates a task and sends an acknowledgement. It always does the same thing, which makes it predictable and cheap. But if the customer writes "how much would it cost to swap the bath for a shower tray?" without mentioning "quote" anywhere, the rule never fires.

AI agents decide which steps to take

The agent reads the email, understands it's a quote request even though it doesn't say so, checks whether the customer already exists, looks for free slots for the visit and prepares the reply. Anthropic's guide to building effective agents, from the company that makes Claude, draws the line right there: in an automation the path is set by code, while in an agent the model decides which steps to take and which tool to use at each point. That freedom is what makes it useful, and also what lets it make mistakes.

What AI agents can do in a small business

There's no need to picture a robot running the company. What works today is specific tasks with a clear start and finish, where the agent prepares and a person decides.

Sorting email and leaving replies as drafts

Sorting what comes in (quotes, problems, invoices, suppliers), summarising long threads and preparing replies to the usual questions. It's the example from the start, and it's usually the first one to be set up because the time saved shows every morning.

Preparing quotes with your prices

With the customer's request and your price list in front of it, the agent puts together a first draft with the items that fit. Whoever knows the job reviews it, adds what only they know and sends it.

Finding and cross-checking data in your software

"Which customers have an invoice more than thirty days overdue and haven't had a reminder yet?" Today a question like that means someone opening two or three programs. An agent with read access can run the search and hand you the list.

Handling enquiries out of hours and passing them to a person

Answering an enquiry at eleven at night, asking for the missing details and leaving the case ready for someone to pick up first thing. The second part is what matters: the agent recognises what isn't its job and hands it over.

In every case, the AI drafts, summarises, sorts or replies inside the tools you already use, which is how we work in our AI integration and training service.

When an automation is all you need

The most useful advice in Anthropic's guide is also the least flashy: look for the simplest solution possible and only add complexity when it's needed, which sometimes means building no agent at all. Agents gain flexibility at the cost of being slower and more expensive, because every step is another call to the model.

If a task always has the same steps and clear rules, process automation will do it faster, cheaper and with no surprises. An agent makes sense when it has to read texts that are different every time, work out what someone is asking for or choose between several options. To see which group each task falls into, which tasks to automate with AI has a four-question filter.

There's another reason to take it slowly. In June 2025, the research firm Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. It also warned about "agent washing": ordinary assistants, chatbots and automations being rebranded as agents. By its estimate, of the thousands of vendors claiming to sell agents, only about 130 are the real thing. Before paying for an "agent", ask what it decides on its own and what it does when it gets something wrong.

The risks of handing an AI agent the keys

A chat tool that gets it wrong gives you a bad answer. An agent that gets it wrong can send, delete or change something in your software. The risks are well known, and most of them shrink when you choose carefully what it's allowed to do.

It's wrong with great confidence

A language model can make up a fact with the same assurance it uses for everything it gets right. In an agent, a mistake in the first step also carries through to the next ones. Anthropic puts it plainly: agents' autonomy means higher costs and errors that compound, and it recommends testing them thoroughly in a controlled environment with guardrails before letting them loose.

It follows hidden instructions

This is the newest risk and the least intuitive. It's called prompt injection: hiding orders in a text the agent is going to read. The AEPD gives the examples of a PDF, an email or a web page carrying instructions that a person can't see but the agent takes as legitimate orders, which can end in a data leak. If your agent reads incoming email, anyone who writes to you can try it.

It has more permissions than it needs

OWASP, a non-profit foundation known for its lists of IT security risks, calls this excessive agency: too much functionality, too many permissions or too much autonomy. Its example could happen in any office. An assistant that was only meant to summarise email uses a connection that can also send, and a rigged email tells it to search the mailbox for sensitive information and forward it to a stranger. The fix it proposes is simple: a read-only connection, and a person who reviews and presses send on every email the agent drafts.

It handles personal data

If the agent reads customer emails, records or invoices, it's processing personal data, and responsibility stays with your business. The AEPD's guidance asks for the level of autonomy to be set for each use according to its risk: acting on its own where the impact is low, with mandatory human intervention where the risk is high.

Five rules before you give it work

They're the same ones you'd apply to a new person in the office, who doesn't get access to everything on day one either:

1. One task, small and checkable. "Sort quote emails and leave drafts" can be measured. "Handle customer service" can't.
2. Only the permissions it needs. Read before write, draft before send. If it only needs to check the calendar, it shouldn't be able to change it.
3. A person approves anything that can't be undone. Sending to a customer, paying, deleting, publishing or cancelling an order. OpenAI's guide gives cancelling orders, authorising large refunds and making payments as examples of high-risk actions, and says they should go through a person until the agent has earned trust.
4. Everything is logged. What it read, what it decided and what it did, so it can be checked when something doesn't add up. The AEPD calls this traceability, and OWASP recommends logging and monitoring what the agent's connections do.
5. Whoever reviews has the time and judgement to do it. If everything gets approved without being read, the oversight is for show. The AEPD warns about automation bias, the tendency to trust whatever the machine proposes, and about the temptation to blame the person supervising when the fault lay in how the system was built.

How to tell whether an AI agent pays off

After a few weeks it's time to look at the numbers, which is the last phase of any project of ours: measure and improve. Four questions give a good picture:

1. How much time is saved each week, counted by the person who used to do the task?
2. How much of what the agent prepares goes out unchanged, and how much needs correcting?
3. What mistakes were caught at review, and what would have happened without it?
4. What does the model's usage cost each month, compared with the hours it frees up?

If half the drafts need rewriting, the problem is usually in the instructions or the data it receives, not in the idea. And if it still doesn't pay off, you go back to a standard automation and nothing is lost.

How much an AI agent costs for your business

There's no catalogue price, because an agent that sorts email doesn't cost the same as one that cross-checks data from three programs. As explained on our pricing page, automation and AI integration are quoted per project, not as a monthly fee: you tell us what you need and we give you a closed price in writing before we start.

On top of the setup, using the model has its own cost, which grows with the number of tasks and the steps each one needs. That's why it makes sense to start with one task and look at the real numbers before expanding.

If you don't know whether your case calls for an agent, an automation or nothing at all, the first step is an AI audit: we look at how you work and tell you where AI saves you time and where it doesn't, with priorities and costs.

AI agents with both feet on the ground

We're three partners working from the Palau Firal in Sant Antoni de Calonge, with businesses across Girona, Barcelona and the rest of Catalonia. Automations and connections to your software are built by Ivan, who would rather ship something small that works than a huge project that lands late. Access is checked by Izan, who makes sure data is where it should be and that every account has the permissions it needs, and not one more.

If you have a task that comes round every day and you think an AI agent could do it, message us on WhatsApp and we'll look at it in a first conversation with no obligation. You can also write to us through the contact form.