A weekend and a subscription to a couple of AI tools, and anyone on your team can build an internal system. A documentation search tool, a bot that answers customer questions, a dashboard that connects data that used to live apart. It works on the first try and the feeling is one of omnipotence: you just built, on your own, something that used to take a whole team. That's where the problem starts, because being able to do it isn't the same as knowing what you've just taken on.
Where your data ends up, and whose it is now
In 2023, Samsung engineers pasted proprietary code and internal meeting notes into ChatGPT to move faster. That content left the company and became part of a third-party service. There was no attack and no technical bug. Just someone using a tool without asking where what they typed in would end up. Samsung is a multinational with a legal department, and it still happened; at a small company, nobody asks that question before connecting the first prompt.
Every time your system sends information to an AI model, that information leaves your servers and goes to another company's, usually in another country. If you handle customer data from Andorra, that has a specific legal name: an international data transfer. Andorra's data protection regulation requires a prior assessment before making one, and if the risk can't be mitigated, you have to consult the Andorran Data Protection Agency (APDA) beforehand, not after. The APDA itself has publicly acknowledged that much of this data ends up on servers in the United States, and that a bilateral legal framework to regulate it with real guarantees is still missing. If the authority itself says so, it's not a minor detail you can skip because "it's just a prompt."
Who's accountable when the law already applies, even if you built the system yourself
The EU's AI Act already requires, since February 2025, that any company using AI ensure the people operating it know enough to do so with judgment. And it distinguishes between whoever sells an AI system and whoever simply uses it inside their own company: this second role also carries its own obligations, like real human oversight, logging what the system does, and informing employees before rolling it out if it affects their work. Building it yourself with AI doesn't exempt you from these obligations, it just means nobody has probably checked them yet.
Andorra doesn't yet have its own AI law. What it does have is data protection law, and that law makes no distinction between "a professional vendor built this" and "we put this together ourselves on a Friday afternoon." If the system handles data about customers, patients or employees, the law you already followed before still applies, with the added fact that the data now passes through a new intermediary that didn't exist a year ago.
Will you know when it gives you a wrong answer?
This is the question that gets asked the least and costs the most. There's no single reliable figure for how often AI fails at a real business task (if someone gives you a clean percentage without saying where it comes from, be skeptical): it depends heavily on the task, the data it works with, and how it was measured. What is true, and verifiable with any model you use today, is that none of them reach zero errors, and that it answers with the same confidence whether it's right or wrong. The only reliable way to know your system's error margin is to measure it against your own documents, not trust the tool's brochure.
And the reason it fails usually isn't that the AI "makes a mistake" out of nowhere. It's usually that the system doesn't know a piece of information has gone stale. An internal page said a feature was supported, a support thread says it has a limitation, a colleague says it depends on how it's configured; if nobody has flagged which of these three is current today, the system treats all of them with the same confidence. A document corrected six months ago, if nobody updates the source, keeps showing up as if it were still valid. That's not a one-off AI mistake. It's a lack of maintenance on what feeds it, and maintaining that is ongoing work, not a button you press once.
If you or your team don't have that judgment, nobody will catch the error until someone makes a decision based on it: a miscalculated quote, a wrong customer record, an answer given to a customer that shouldn't have been given. And when it happens, fixing it once isn't enough. You need someone who understands why it happened, updates the stale source, and records who validated it, so it doesn't happen again the following week.
What AI gives you in a weekend, someone has to watch, maintain and fix for years.
Build or buy: the question isn't new, what's changed is the temptation
Build versus buy is a decision companies have made for decades. What's changed isn't the decision, it's that "building it" now looks free and instant, so it gets made without going through the same filter as before. Nobody used to put together an ERP on a Friday afternoon, because it was obviously unworkable. A system built with AI in a few hours doesn't look like the same kind of commitment, but it moves the same data, falls under the same law, and can get things just as wrong.
The right question isn't "can I do this myself with AI?". It's who's accountable for the data once it leaves the country, who's checked that you comply with the law that already applied to you, and who will catch and fix the error when it shows up.
A system nobody designed to last, lasts right up until someone actually needs it to work. That's when it shows who checked the data, the law and the error, and who only checked that it ran.
When doing it yourself actually makes sense
Not everything should be bought. If what you need is small, you or a small team are the only ones using it, it doesn't process sensitive data, and you can afford it breaking without anything serious happening, just build it with AI. That's exactly what these tools are great for: testing an idea, automating a personal task, saving time on something that doesn't depend on you tomorrow if it fails.
The threshold is easy to remember: if what you're building touches customer data, makes decisions for other people, or other people rely on it to do their job, it's no longer an experiment. It's infrastructure, even if you put it together in one evening. And it's a specific kind of infrastructure: connecting a couple of tools isn't the same problem as keeping what feeds them alive. Storing information so it can be found is one thing. Making sure what gets found is still true is something else entirely.
What we bring to it
We use AI constantly, to move faster on repetitive work. But when a client asks us for a system their business will depend on, the first questions we ask aren't technical: where will the data live, what regulation applies to them, and how do we detect and fix an error the day it shows up. If none of those answers exist yet, it's not time to build it. It's time to talk about it first.
Frequently asked questions
Does building an internal system with AI exempt me from data protection law?
No. If the system handles data about customers, patients or employees, the law you already followed before building it still applies. In Andorra, sending that data to an AI model is also usually an international data transfer, which requires a prior assessment under data protection regulation before you do it.
How do I know if the system is giving me a wrong answer?
An AI system answers with the same confidence whether it's right or wrong, and it often fails because an internal source has gone stale or contradicts another one without anyone flagging it. The only reliable way to know your error margin is to measure it against your own documents, not trust what the tool's marketing promises.
When does it make sense to build a system with AI instead of buying one?
When it's small, you or a small team are the only users, it doesn't handle sensitive data, and you can afford it breaking without serious consequences. If the system touches customer data, makes decisions for other people, or other people rely on it to do their job, it's no longer a personal experiment, it's infrastructure someone has to maintain.
