An adviser proudly showed us the chatbot he had signed up for the year before. It was on his website, it had a name and it answered. We asked how many enquiries it resolved a month. He stopped to think. He did not know. Nobody had looked at it since the day it was installed. We opened the dashboard: the last conversation that had ended well was from February. From then on, one after another, questions the bot did not understand and customers ending up writing "I want to talk to a person".
The tool was not bad. The way it was put into the business was. And this is what almost nobody says: mistakes when implementing artificial intelligence in an SME are almost never technical. They are decisions made before touching any tool, or decisions not made. And they repeat so often that you can make a list of them.
This is that list. Seven mistakes we have seen up close, each with its antidote. There is no magic figure in here. There is judgement, which is what really separates a project that holds up from one that has been switched off since February. If you have not started yet and want to see the whole process, first read how to implement AI in your company in five phases.
Before the mistakes: the good news
It is worth starting with something that plays in your favour. Few people are doing this yet. According to the INE survey on ICT use in companies, published in October 2025, 13.4% of Spanish companies with fewer than ten employees used artificial intelligence technologies in the first quarter of 2025. And among those with ten employees or more, according to the ONTSI indicators on AI use published in April 2025, the figure was 11.4% in 2024.
However you look at it, it is still a minority. And that is the interesting part: most of those who are going to do it have not done it yet. You have room to do it well while the competition has not yet tried. And "doing it well" starts with not tripping over the same seven stones as always.
Mistake 1: automating a process that is broken
This is the most expensive and the most common. You have a process that works so-so (the order flow, signing up a customer, the way a quote becomes an invoice) and you decide to put AI on top of it to make it faster.
The problem is easy to state: automation does not fix a bad process, it speeds it up. If today the data is entered twice and sometimes does not match, tomorrow it will not match just the same, only faster and with fewer people looking. You automate the error along with the success.
The antidote is boring and it works: fix the process first, on paper, without technology. Tidy it up, take out the steps that are not needed, make clear what is done when the data comes in odd. When the process is clean and you can explain it out loud from start to finish, then automate it. If you cannot explain it, it is not that you lack a tool: it is that you do not yet know what you want to automate.
Mistake 2: starting with the biggest thing instead of the most measurable
The temptation is to go for the project that impresses most: the assistant that does everything, the complete overhaul of the system, "while we're at it, let's do the whole thing". It is the quickest road to a project that never ends and that nobody can say whether it has helped.
The antidote: choose what is most repeated and measurable, not what is biggest. A process that happens many times a day, that you can count and that has a result you can see. Appointment reminders. Answers to the same five questions. Moving a web form into your system. Something small, closed and with a before and after you can compare. If the first project works and it shows, you have permission for the second. If you start with the biggest and it fails, you are left with no project and no appetite. We cover it in more detail here: how to implement AI in your business without knowing how to code.
Mistake 3: not measuring the "before"
This is the one everybody skips, and it cannot be fixed once skipped. You set up the automation, it starts to work, and three months later someone asks the obvious question: is this paying off? And you cannot answer, because you did not write down how things were before you started.
If you do not measure the before, you cannot prove the improvement. Not to others and not to yourself. You are left with the feeling that it is going better, which is exactly what you had before spending the money.
The antidote is a notebook and a week. Before touching anything, count the process you are going to automate: how many times it happens, how long each time lasts, how many enquiries go unanswered, how many things have to be redone. In writing, not from memory, because memory inflates what annoys and hides what has become normal. That one-week count is the only thing that separates knowing it worked from believing it worked. And it only exists once: once automated, the "before" disappears. On how to do that count and what to put in it we have written a whole guide: how to calculate the ROI of AI automation.
Mistake 4: removing the human touch exactly where the customer values it
AI is very good at taking the repetitive out of the way. The problem starts when, in the enthusiasm, you also automate what the customer came looking for from a person. The complaint from someone angry. The delicate enquiry. The moment when the customer wants someone to tell them their problem matters.
A bot replying "I understand your frustration" to someone who has just had a real problem does not save work: it creates a bigger job. The antidote is to draw a line before starting: what is automated and what is not. The repetitive and predictable, for the machine. What needs judgement, tact or a genuine apology, for a person. And always an exit to hand: the customer must be able to move to a human without fighting the system, in the first message if need be.
There is also a reason that is no longer just common sense. The European Commission's EU Artificial Intelligence Act (Reglamento Europeo de Inteligencia Artificial), in force since August 2024, brings in transparency obligations for chatbots: from August 2026, when someone talks to an AI system, they must know they are talking to a machine. Hiding it, besides going down badly, stops being an option. Say it clearly. A customer who knows a bot is serving them and who can ask for a person whenever they want gets far less angry than one who was led to believe they were talking to someone.
Mistake 5: buying an off-the-shelf tool that doesn't fit how you work
There is a pattern that repeats: someone buys the tool they were recommended, the one "everybody uses", and then spends half a year bending their business to fit inside it. They change their processes to suit the program, instead of the other way round. They end up working for the tool.
The antidote is to turn the question round. It is not "which tool do I buy?", it is "how do I work and what do I need it to do?". The technology adapts to your way of working, not you to its. Sometimes the answer is a well-chosen off-the-shelf tool, and fine. Other times your process has something of its own (the way you treat your customers, an integration with the system you already use, a flow that does not look like your neighbour's) and then what you need is something built to your measure, not a forced template. Telling one case from the other is half the decision. You can see what we work on to get an idea of where that line lies.
Mistake 6: forgetting the GDPR with your customers' data
An AI system that serves your customers processes personal data: names, phone numbers, sometimes more sensitive things. Putting that data into a tool without stopping to think where it comes from, where it ends up and who sees it is a mistake you do not see until you see it, and then it is too late.
The antidote is not to become a lawyer, it is to treat the matter as part of the project and not as a nuisance at the end. The Spanish Data Protection Agency (AEPD) published a guide on GDPR compliance for processing that involves artificial intelligence; it is the place to start. Three things worth being clear about before you begin:
- Basis and purpose. What you use each piece of data for and on what justification. Data you asked for one thing is not valid for anything else.
- Transparency. The customer must be able to know that their data is processed by an automated system and what is done with it.
- Automated decisions. Article 22 of the GDPR gives the right not to be subject to a decision based solely on automated processing when it produces legal effects or significantly affects the person. In plain terms: if something serious for the customer depends on an automatic answer (granting or refusing something), there must be a person who can review it.
Data is processed in line with the GDPR because it is the law, yes, but also because it is what separates a business you can trust from one you cannot.
Mistake 7: "install it and forget it"
Back to the adviser at the start. His mistake was not choosing the chatbot badly. It was thinking an AI system is like a household appliance: plug it in and that is it. It is not. A system that serves customers changes along with them: new questions appear, a price changes, a product comes in, and what it answered well in January it answers so-so in June if nobody looks at it.
The antidote is to accept from day one that this needs maintenance and to decide who does it. You do not need a department. You need someone to look, every so often, at what the system is answering, where it gets stuck and which new questions it cannot answer. Half an hour every two weeks at the start stops it degrading on its own until it is useless. A system nobody reviews is not cheaper: it is cheaper until the day it gets something wrong and nobody sees it.
The seven, in one sentence each
If you keep just one thing from all this, let it be the order:
- Fix the process before automating it.
- Start with what is most repeated and measurable, not what is biggest.
- Count the "before" before touching anything.
- Automate the repetitive; leave the human where the customer values it.
- Let the tool adapt to how you work, not the other way round.
- Treat the GDPR as part of the project, not as the paperwork at the end.
- Decide who maintains it before switching it on.
None of the seven is technical. All of them are decisions. That is why someone who cannot code can take them well, and why someone in a hurry takes them badly.
Frequently asked questions
Where do I start if I don't want to get it wrong?
With the most repeated and easiest-to-measure process, not the most ambitious. Count it for a week, automate only that, and count it again three months later. A small first project that works and shows is worth more than a big one that never finishes.
Can I implement AI without knowing about technology?
Yes. The obstacle in these seven mistakes is not technical, it is one of judgement: knowing what to automate, in what order and what to measure. A supplier can handle the technical side; you supply the judgement, because you know your business. On the practical side we have written this guide without jargon.
Do I have to tell my customers they are talking to a bot?
Yes, and increasingly by law too. The EU AI Act brings in transparency obligations for chatbots from August 2026. Rules aside, telling people is what takes the most heat out of the anger: a customer who knows a system is serving them and can move to a person whenever they like takes it much better.
What about my customers' data?
It is processed in line with the GDPR, and that has to be planned before starting, not after. The Spanish Data Protection Agency has a guide on AI and the GDPR worth looking at. The essentials: be clear what you use each piece of data for, be transparent with the customer and make sure a person can review any automated decision that seriously affects them.
If you'd rather not trip alone
These seven mistakes are avoided with judgement, and judgement can be applied on your own. This article is written for that: you can use it as a checklist before hiring anyone for anything.
If you would rather we looked at it together, we sit down, see which process you want to automate, whether it is ready for it and how we would approach it without falling over any of these seven stones. Tell us what you need and we will say whether we can do it, how we would approach it and what it would cost. If it is not for us, we will tell you that too.
Xora IntelligenceAI consultancy and custom software in Los Barrios (Cádiz). We write about what we build for businesses in the province. If you have a question about your case, write to us.




