AI agents

A chatbot answers you; an agent does the work

The difference isn't in how it talks, it's in whether it touches anything. An email comes in with an order. An assistant that only replies summarises it and explains what you ought to do; doing it is still up to you. An agent has tools connected: it reads the email, finds the product references in your database, checks what terms that customer has in the CRM, generates the document, puts it where it goes and notifies whoever has to sign it. At the end there's a record and a PDF, not a well-written reply. It's the confusion we most often have to clear up on the first call, and we compare the two in depth in our blog article on agents versus chatbots (in Spanish).

How much agent do you need?

There's no need to start big. The first level is an agent for one specific case, with one data source and one flow: it reads what comes in, checks one place, does one thing. It's set up, measured, and we see whether it holds up day to day. The second appears when that case has branches: several flows, memory of what has already happened with that customer or that file, and tools connected to your real systems. The third is a whole process covered by several specialised agents passing work between them: one sorts what comes in, another prepares it, another checks it. You get there because the process calls for it.

An agent without a test suite is a roulette wheel

An agent doesn't always respond the same way to the same input, and that's its weak spot. So before putting it to work, we build a set of real cases from your business with what should happen in each: what it should answer, which tool it should use and what it must never do. The whole set is run every time something changes, because adjusting one detail can break another that was working, and without tests nobody finds out until a customer does. It's work nobody sees, and it's what separates an agent you can leave to run from a demo that looks good in a meeting.

What we don't let it decide on its own

An agent carries out well-defined work well and gets things wrong when it has to use judgement. So anything with consequences (money going out, a contract, an email to an angry customer, deleting something) goes through approval: the agent gets it ready and a person presses the button. It's also worth knowing that an agent only knows what its systems let it read: if the key information is in the warehouse manager's head, it doesn't have it. And when the agent talks directly to a person, the EU AI Act requires it, from 2 August 2026, to identify itself as an automated system. We build it that way as standard.

Source: European Commission, regulatory framework for artificial intelligence (Regulation (EU) 2024/1689, Art. 50)

The proof we can show you is our own

We don't publish client case studies, so we show you our own. The form on this website doesn't email anyone: it writes straight into our CRM, creates the contact record and attaches your message, so when we call you we already have in front of us what you told us. Behind the scenes, that data syncs itself with our knowledge base and a monitor alerts us when something stops working. That part you'll have to take our word for; the outside you can check yourself right now, without asking our permission. This website publishes an llms.txt file, the index the models read. The robots.txt gives explicit permission to GPTBot, ClaudeBot, PerplexityBot and OAI-SearchBot. Every page carries entity structured data with the registered company name, tax ID, founders and service area. And there's a library of guides by sector and by town, written one by one. We track what comes out of it in Search Console, not by guesswork. We write the code of what we sell, and this comes from using it ourselves first.

Source: llms.txt on xoraagencia.com

When we'll tell you no

There are situations where an agent doesn't pay off, and we'd rather say so up front. One: the process changes every week. If today's rules aren't next month's, whatever you build spends its life in maintenance. Another: there's no data. If the information lives on paper, in someone's head or in a spreadsheet only one person understands, the agent has nothing to read from, and the first step isn't an agent: it's putting that in order. And if your case can be solved with a simple automation, without a language model involved, we'll tell you that too.