7 min read
A company has spent months evaluating which AI tool to implement. The operations director wants to improve weekly demand forecasting by SKU. The marketing team is convinced the solution is a system based on language models. Nobody is wrong about wanting to improve the process; both are wrong about which type of system improves it.
The problem is that the term "AI" stopped referring to a technical category and started referring to a single family of systems — more than a technical judgement call by the team. That naming shift has practical consequences: companies dismiss tools that would solve their problem because "they're not AI", or buy tools that are not the right fit because they are. The confusion generates two symmetric errors. The first: dismissing a predictive machine learning solution because "it's not real AI" (meaning it is not generative). The second: buying a language system to solve a problem where a classical predictive model would work better, faster and at lower cost.
The renaming that nobody announced
In 2020, the set of things the industry called "applied AI" was broad: classification models for detecting customer churn, time series for demand forecasting, recommendation systems, image recognition in manufacturing. All of that was AI. There was reasonable consensus on the term.
ChatGPT launched in November 2022. In the months that followed, the term "AI" re-anchored in the market conversation to a single type of system — large language models and their derivatives. The generative became the default reference. What was previously the core of "applied AI" started to be called something else: "advanced analytics", "predictive models", or simply "statistics".
Classical machine learning did not disappear. It kept training, deploying and producing predictions every week in thousands of companies — but it lost the name it had in the market conversation. The practical result is awkward. Many companies that today are evaluating whether to adopt AI already have AI running in production. They just call it something else, and that makes it hard to see what they have and what they are missing.
Two technical families, one name in the market
The distinction that matters for decision-making is about what type of signal each system processes well — more than platform or vendor.
Classical predictive models operate on structured variables: numbers, dates, categories, proportions in tables. They train on labelled examples and learn to generalise a pattern. Given a customer's purchase history and characteristics, the probability they will buy again in the next 30 days. Given weekly sales history by SKU and seasonality variables, the forecast for next week. The output is specific, bounded and reproducible. A gradient boosting model on time series does not improvise or reason about context: it solves a well-defined function on well-defined data.
LLMs operate on text. Training is on massive text corpora: books, articles, code, conversations. Generation is autoregressive: the model estimates the next most probable piece of text given what came before, over and over. The strength is linguistic flexibility: understanding natural language instructions, extracting information from documents with variable formats, synthesising, reformulating. The relative weakness is opacity, computational cost and a tendency to produce plausible but imprecise outputs when the problem requires numerical accuracy.
Both are probabilistic. Both learn from data. Both can be wrong. The difference lies in the type of signal each processes with advantage and in the properties of the output it produces.
Three problems, three different answers
Weekly demand forecasting by SKU. The problem is tabular: sales history, calendar variables, price, seasonality. The expected output is a number with confidence intervals for each product. The model must be reproducible, auditable and able to update its estimates when the consumption pattern changes. A gradient boosting model trained on that history produces exactly that. An LLM applied to the same problem operates outside its zone of advantage, adding computational cost and reducing output control without improving precision. This is a classical machine learning problem.
Data extraction from invoices with different formats. The problem is textual and unstructured. Each supplier has its own format, field layout and conventions for dates and amounts. A rule-based or classical classifier system requires one rule per format and breaks with every new supplier. An LLM that receives the document as text and returns the fields in JSON structure generalises to new formats without additional rules, because it understands semantic context regardless of the position on the page. This is a problem where the linguistic flexibility of LLMs has a real advantage.
Next best product recommendation. The problem is hybrid. The propensity part — the probability that a specific customer is interested in a given product given their history — is tabular and is better solved with a classical predictive model trained on purchase history. The message personalisation part — adapting the communication text to the customer's context — is linguistic. The correct architecture combines both. The predictive model generates candidates ordered by purchase probability; the LLM produces the personalised message. Trying to solve everything with a single type of system introduces unnecessary compromises on one side or the other.
The criterion for choosing lies in the problem. It depends on what type of signal the system processes and what properties the output must have.
LLMs are built on top of machine learning
Language models are an extension of classical machine learning, built on top of it. Fine-tuning an LLM uses gradient descent on a supervised loss function — the same mechanics as a regression or traditional classification model. RLHF (reinforcement learning from human feedback), the method that aligned early chat models' behaviour with human preferences, is a variant of reinforcement learning with a learned reward function — a technique that has been in development since the 1990s. The embeddings that allow retrieval systems to find relevant fragments in a document base are evaluated with precision, recall and ranking metrics — the same metrics of classical ML applied to an information retrieval problem.
The infrastructure on which the generative AI of 2026 operates is, to a large extent, machine learning that has been developing for decades. The leap between 2020 and 2026 was in the scale of training and in the versatility of the natural language interface. The underlying mathematical principles are the same. Generative AI is built on top of classical ML, with a new layer on top. The investment decision, then, is about layers of the same stack, each with its zone of advantage and with the type of problem it solves best.
The question that defines the path
Before evaluating what system to buy, there is a question that clarifies the map. What type of signal does this system need to process?
If the signal is structured, quantitative and historical, the classical predictive model has the advantage. It produces specific outputs, it can be audited, it updates with new data and has a lower operational cost. If the signal is unstructured text, documents with variable formats or natural language from users, an LLM has a real advantage. If the problem combines both types of signal, the correct architecture combines both types of model.
A second question complements the first. What properties must the output have? A prediction that feeds a CRM system must be deterministic, traceable and reproducible. A personalised recommendation text can have more acceptable variability. The properties of the output condition what type of model can produce it reliably.
The right diagnosis starts before choosing tools. You need to decide what type of decision you are trying to improve and what kind of signal you are working with to improve it. That conversation — about the problem before the technology — is what makes a project reach production instead of staying as a pilot.