There is debate over which AI model is the most powerful. Meanwhile, the real bottleneck for AI projects in SMEs remains taboo: data quality. Concrete feedback, supporting figures.
When I wanted to equip my own site with an assistant capable of responding to visitors based on my content (a RAG system, this architecture which makes an AI respond based on a given documentary database), I started with the “intelligent” part: the model, the vector search engine, the settings. The part that makes you dream.
The first indexing of my site produced 593 vectors. In everyday language: the content has been cut into 593 small pieces, transformed into digital representations that the AI can find and cite. On paper, everything worked. In fact, the responses were poor. The assistant cited fragments of navigation menu, returned snippets of layout code, invented structure. The model was not in question. He did his job exactly: he faithfully indexed what was given to him. The problem is what we gave him.
The fuel was adulterated
My pages had been built with a visual builder. Excellent for designing a site; disastrous as a source for an AI. The exported content was riddled with technical noise: inline CSS styles, column and section tags, menu and footer text repeated on every page. The human sees none of this: his browser displays a clean page. The machine ingests everything, including the piping.
So I did the thankless job: cleaning. Filtering using regular expressions to remove CSS and layout markup, isolate text that actually had meaning, and remove structural repetitions. Reindexing: 385 vectors. A third less than the first time. The same site, but reduced to its useful part. And the quality of the responses jumped. The lesson is in one sentence: better 385 clean vectors than 593 noisy vectors. In AI, the quantity of data is not a virtue; cleanliness, yes.
I add a more discreet trap, for the curious: my index expected vectors of a certain size (1,536 dimensions) while the model produced another (1,024). A silent inconsistency like this doesn’t crash the system: it just makes it respond sideways, without an error message. This is the treachery of these projects: they apparently “work” while producing falsehoods.
It’s not my problem, it’s THE problem of SMEs
I could put this experience down to my clumsiness. That would be misdiagnosing me. Because the data of small and medium-sized businesses is not absent: it is dirty. Exported from tools that were never designed to feed an AI: site builders, crookedly scanned PDFs, spreadsheets with merged cells, client files where the same name is spelled three ways, histories scattered across five software programs that don’t talk to each other.
This observation is not an isolated intuition. According to Adobe’s AI and digital trends 2026 study, conducted with Oxford Economics among 3,000 managers, 78% of French companies cite data integration and quality as the main obstacles to scaling up AI. And less than half believe that the quality of their data today is sufficient to take advantage of it. My 385 vectors are this statistic in miniature: the usable third of a set that I thought was ready.
Reverse the reflex
The dominant reflex “let’s choose the most powerful AI” takes the problem in reverse. Before the model, there is invisible, thankless and decisive data preparation work. “Garbage in, garbage out” is not a conference slogan: it is the daily life of those who deploy these systems for real.
The good news for an SME is that this work requires neither the biggest model nor the biggest budget. It requires rigor: knowing where the data comes from, what it is worth, and cleaning it before entrusting it to a machine. A modest project considering what it unlocks.
The question to ask is therefore not “which AI to choose?”, but “can my data be read by a machine?”. Most leaders discover, as I did that day, that the answer is: not yet. And this is precisely where the real project begins.
