When AI makes a mistake and nobody notices

When a system fails obviously, you see it. The dashboard breaks, the numbers don't add up, someone calls. But there is another type of error that is much more difficult to detect: the one that occurs silently, with the complete appearance of normality. **AI, when it makes a mistake, makes it very confidently. The system continues to function, continues to generate outputs, continues to give answers. And you continue to make decisions based on those answers.** This is the problem no one wants to discuss when talking about artificial intelligence applied to business. For years, we've seen the conversation about AI almost always revolve around models: which one is more powerful, which one reasons better, which one scales more. It's a legitimate conversation, but incomplete. Because the model decides nothing on its own. What decides is the combination of the model with the data that feeds it. And much less is said about that second part. In the world of retail and e-commerce, this is particularly clear. Companies have been **accumulating data** for years with the same logic with which some people accumulate objects at home: just in case. **I call it the data Diogenes syndrome**. You have a lot of data, but you don't have the capacity to consume it nor do you know what each thing is for. And then comes the promise of AI, which is supposed to sort out that chaos. But AI doesn't sort out chaos: it amplifies it. A powerful model on low-quality data doesn't produce better decisions; it produces erroneous decisions with more speed and more conviction. **The question we should ask ourselves before any investment in AI is not "what model do we use?" but "on what data are we going to use it?"** And the realistic answer, in most of the companies we work with, is that that data is not in good condition. Not because no one has cared, but because **maintaining reliable data at scale is much more difficult than it seems**. Extracting a hundred products from Amazon at a specific moment is trivial. Extracting a million products, three times a day, for a year, with a minimal error rate, is a completely different game. And that difference is systematically underestimated. There is another conversation that is also avoided: the one about maintenance. There is a widespread idea that maintenance is what you do when something breaks. A patch, a fix, support. I see it differently. **Maintenance is part of the product, it is not support**. Data systems live in an environment that is constantly changing: websites are modified, formats change, platforms update their structures. A system that is not designed to anticipate these changes is not reliable, it is a system that works until it stops working. And when it fails, the customer notices before you do. Designing with that mindset changes how you build. You don't look for the perfect solution for today; you look for one that can be correctly rebuilt when the environment changes. **Being adaptable is not having an answer for everything today. It is the ability to build the correct answer tomorrow.** I was asked a question in a recent conversation that I found very well-posed: **if all AI models disappeared tomorrow, what capabilities would still be essential?** The answer is the usual one: knowing what to measure, building systems to measure it reliably, and making decisions based on that measurement. What changes with AI is the speed at which bad data can spread throughout your decision chain. The challenge is not fundamentally technological. It's about judgment. **Knowing what data you need, with what frequency, with what level of quality, and for what specific decisions**. Data that doesn't lead to action is a waste of time. And that clarity, at a time when everyone is rushing to implement agents and automations, is scarcer and more valuable than any model. I am writing this after a conversation on the **Ecommerce News podcast** where we tried to talk about all of this with the honesty the topic deserves. If you are interested in delving deeper, it is available to listen to. But the reflection I take away is simpler: before asking yourself what your AI does, ask yourself what data it is working with. **LISTEN TO THE PODCAST HERE**