Why artificial intelligence in business requires much more than a good model. The real obstacle to AI adoption in companies is not the quality of the models. It is the quality of the information infrastructure on which those models must operate.
When a company decides to adopt artificial intelligence, the first question almost always concerns the model: which one to choose, how to train it, which cloud to run it on. This is the wrong question — or rather, a premature one. An excellent model on poor data produces poor results. The metaphor comes from old computing: garbage in, garbage out. In the AI era, the garbage is not just noise in the data: it is fragmentation, semantic ambiguity, unstructured and ungoverned knowledge.
The first challenge concerns data that lives in ERPs, CRMs, MES systems, and operational platforms. Here there are three critical issues. The first is integration: the real problem for companies is not having too little data, but having too much, scattered across non-communicating systems, with Excel spreadsheets serving as fragile bridges between them all. Giving AI access to this fragmented ecosystem requires strategic architectural choices, not just technical ones.
The second is quality: duplicate, incomplete, or outdated data produces wrong answers delivered with confidence — one of the most insidious risks of enterprise AI. The third is access governance: not all AI systems should be able to query all databases. Defining who has the right to access which data, in which context, is both a regulatory (GDPR, AI Act) and an organizational requirement.
Even with integrated, high-quality data, a subtler problem remains: semantic context. “Gross margin” for one division has a different meaning from “gross margin” for another. “Customer” in the ERP may have a different definition from “customer” in the CRM. Without a shared semantic layer — a company glossary that establishes what each term means, in every context, for every division — the model will correctly answer the wrong question.
Building this layer is not glamorous and does not appear in press releases. But it is the difference between an AI system that responds and one that truly understands.
Alongside structured data exists the documentary heritage: contracts, emails, presentations, proposals, org charts, source code, meeting minutes. This is the living memory of the company. Modern language models are powerful in processing this content, but only if that knowledge is accessible, organized, and up to date.
In reality, this knowledge is scattered across shared folders, email inboxes, obsolete intranets, and unmaintained repositories. In most companies, there is no coordinated knowledge management system. Each function manages its own knowledge with independent logic. The result is that AI accesses a chaotic documentary base and produces chaotic responses; or worse, it does not access it at all and the heritage remains entirely unused.
Leakage is not a future risk. It is a present risk, occurring every day in which an employee copies a confidential document into a consumer chat window.
While companies build — or postpone — their own AI infrastructures, people find alternative solutions. The phenomenon is called Shadow AI, and it is the direct heir to the Shadow IT of the 2000s: employees using ChatGPT, Claude, or other consumer systems to work, unknowingly sharing source code, contracts, customer data, and corporate strategies with ungoverned systems.
The answer is not prohibition — it rarely works and only creates an even less visible underground AI. The solution is to build capable, secure, and easy-to-use internal alternatives, accompanied by clear Acceptable Use policies. People use consumer tools because they work. The enterprise solution must compete on that dimension, not just on security.
There is one final risk that emerges when AI starts working properly: the productivity paradox. AI increases individual output: more code, more presentations, more analyses. But more output does not mean more value.
The first effect is volume inflation: the cost does not disappear, it shifts from the producer to the consumer. Whoever receives the code still has to review it. Whoever reads the presentation still has to evaluate it. If the pace of consumption does not adjust, informational overload and decision-making slowdowns are created.
The second is boundary blurring: AI lowers domain barriers. A junior employee produces software architectures, a salesperson generates financial analyses. The semi-finished products look complete but are not, and those who need to validate them often lack the time — or the context — to notice. The third effect is the diffusion of accountability: when everyone produces everything, the chain of responsibility becomes opaque. Governing AI also means governing roles: who can do what, with what supervision, with what responsibility for the output.
Artificial intelligence amplifies what it finds. It will find fragmented data and fragment it further. It will find ungoverned knowledge and make it even harder to manage. But it will also find — where it exists — an organization that knows what it knows, and in that case it will generate value in proportion to that maturity.
AI is not a technological problem. It is a test of organizational maturity. The companies that will pass it are not necessarily those with the best models, but those that will have done the most tedious and most necessary homework: integrating data, governing knowledge, defining roles.
The time to do this homework is now — before choosing the model, before launching the pilot, before the problem becomes large enough to no longer be ignored.
Article by Alessandro Geraldi, Group CEO Impresoft, published on HuffPost. Original source (Italian): HuffPost Italia