The scene that keeps repeating
A company tries an AI assistant on its own data. The first demos are spectacular: you ask questions in plain language and you get answers. Three weeks later, someone from the commercial side asks "who are my five most profitable customers?" and the AI returns a list that doesn't match the reality that person knows. The rushed conclusion is always the same: "AI doesn't work".
The AI works perfectly. It is answering, with total confidence, based on data that contradicts itself. The failure isn't where everyone is looking.
Models don't invent from nothing: they amplify what you feed them
When an AI agent receives fragmented and contradictory context from your CRM, your ERP, and three shared folders, it makes firm decisions on bad information. Hallucinations, biases, and inconsistent recommendations almost always come from noisy, incomplete, or poorly governed data, not from a flaw in the model.
Put another way: a powerful AI on bad data is just an expensive and very fast way to be wrong. Switching tools when the problem is in the data is like switching cameras because the photos come out blurry while your hands are still shaking.
The numbers that should stop anyone before investing
And this isn't only my opinion. A 2025 MIT NANDA report found that 95% of organisations saw no measurable return from generative AI, mostly because the tools don't fit or learn from the way the company works (coverage of the report). Gartner predicts that through 2026, companies will abandon 60% of the AI projects that aren't supported by AI-ready data (Gartner, February 2025).
Both point to the same place, and it sits underneath the model.
Nobody fails at AI by choosing the wrong model. They fail by handing it to data they don't even understand themselves.
What "AI-ready data" means in an SME
I'm not talking about a data lake of millions of records or a two-year project. I'm talking about something more boring and far more useful: that the relevant data lives in one common place and isn't scattered, that each important metric has an agreed definition, that you know where each piece of data comes from and with what quality, and that someone is accountable for it. That is governance, and it's exactly what's missing when an AI starts giving strange answers.
The order mistake: buying the AI before tidying the house
What I see over and over are companies that invest in the AI tool before they've sorted out the foundation. It's the same reversed sequence that shows up in almost any poorly framed data project: you start with the technology instead of the problem. And you end up adapting the business to the tool instead of the other way around. The tool, however good it is, can't fix what you haven't yet defined.
How to tell if your problem is the model or the data
There's a simple test. Take three important business questions and ask a trusted person to answer them by hand, going to the sources. If that person takes hours, runs into numbers that don't add up, or has to decide "which of the two is the right one", your AI is going to have exactly the same problem, only faster and with more poise.
If a capable person can't answer with your data without fighting it, no AI is going to do it by magic. The bottleneck isn't the model.
Where to start
Before switching AI, the order that works is this. You pick two or three questions the AI should answer well. You follow the trail of the data it needs and fix the definitions and the source. You expose to the AI only curated data, not the whole warehouse (the systems that reach usable reliability tend to lean on a few clean views, not the raw tables). And only then do you evaluate the model. In that order, almost any decent model on the market works.
The bottom line
Next time an AI gives you a bad answer, don't switch models on reflex. Ask yourself first whether you could answer that question with your data without fighting it. If the answer is no, you already know where the real work is. And it's exactly the work that makes AI, whichever you have, finally start to deliver.