“Should we redirect the AI budget towards generative AI?” The question reaches the executive committee in this form, or in another: “this sales forecasting project, is it generative AI?”. It assumes two competing technologies. The difference between AI and generative AI is of a different nature. AI refers to the whole field: all systems that derive a prediction, a classification, a recommendation or a decision from data or rules. Generative AI is one of its branches, whose output is new content: text, image, sound or code.
What committees contrast in practice are two families of models. The first, often called predictive AI, estimates a value or places a case in a category. The second writes, summarises or answers. They do not replace each other. The choice starts from what the system must produce: a figure or a category, which the facts will confirm or contradict, or content, which a person will have to read and judge.
AI, machine learning, generative AI: nested sets
Artificial intelligence is an old field. It covers systems based on rules written by experts, optimisation methods and machine learning, in which a model learns from examples rather than rules. Deep learning, based on neural networks, is a part of machine learning.
Generative AI, as it is being deployed today, belongs in turn to deep learning. It groups together models trained to produce content: large language models for text and code, diffusion models for images. All generative AI is therefore AI. The reverse is false: a rules engine or a scoring model generates no content.
Vocabulary fuels the confusion. “Predictive AI”, “traditional AI”, “classic AI” or “discriminative AI” refer, by contrast, to what is not generative. None of these terms is standardised: in a scoping note, have the meaning of each spelled out. The boundary depends as much on use as on technique. A large language model (LLM) can classify an email or extract a date from a contract, which amounts to entrusting a classification or extraction task to a generative model.
The difference between AI and generative AI: a point-by-point comparison
The table compares the two families as they appear in a project: on one side specialised predictive models, on the other pre-trained generative models.
| Criterion | Predictive AI | Generative AI |
|---|---|---|
| Output | A value, a probability, a category, a ranking | Content: text, image, sound, code |
| Raw material | Structured, historised data, often labelled, specific to the organisation | Very large general corpora for initial training, then your documents supplied at the time of the request |
| Origin of the model | Trained by your teams, or configured in software, for a given task | Pre-trained, most often by a third party, then adapted through instructions, documents or fine-tuning |
| Scope | One task per model | Many tasks with the same model |
| Stability of output | For the same version, the same input gives the same output | The output can vary from one call to another |
| Verification | Comparison of the prediction with what actually happened | Critical reading: accuracy, faithfulness to sources, compliance with the expected form |
| Typical error | A wrong prediction, countable after the fact | Plausible but false content, which nothing flags |
The last two rows weigh most in a decision. A predictive model makes mistakes in a way that can be counted: on a test set, the error rate is known before going live. A generative model makes mistakes by writing a true sentence and a false one with the same assurance. Its evaluation requires a set of reference cases and a degree of human judgement. The two families are not operated in the same way either: the comparison appears on our MLOps and LLMOps page.
The situations in which a predictive model remains the right choice
The expected output is a number or a category, and reality eventually decides. Demand forecasting, risk scoring, fraud detection, predictive maintenance, churn probability: in all these cases, the prediction is sooner or later compared with an observed fact. This comparison makes it possible to measure the model, then to correct it.
The raw material consists of structured data: transactions, sensor readings, time series. A language model does not calculate from your histories. It produces plausible text, figures included.
The decision must be reproducible and explained. When a result affects a person (a price, a processing priority, the outcome of a request), you must be able to show that two identical files receive the same answer and say which variables carried weight. A predictive model whose version is frozen lends itself to this better than a model whose output varies.
The volume is high and the response time short. A small, specialised model handles a continuous flow of cases. Operating it mainly requires monitoring its drift.
Where generative AI prevails
The material is language or unstructured content: contracts, procedures, emails, minutes, code. Writing a first draft, summarising a file, rephrasing, translating, answering a question from a body of documents are tasks to which predictive models offer no general answer. Querying an internal corpus most often relies on a RAG architecture.
There is no labelled data set. A predictive model requires gathering a history and qualifying it before the first result. A pre-trained model starts with an instruction and a few examples, which shortens the path to a first trial, without removing the need to evaluate it.
The tasks are numerous and changing. The same model handles requests of different kinds, where each would have required its own model.
A person reads the output and can correct it. This is the condition that makes plausible error bearable: a draft that is reviewed, a summary checked against its sources, an answer that cites the passages used. The conditions for reliable go-live (test set, sourced answers, planned refusals, operation) are detailed on our page devoted to generative AI in the enterprise.
When the two families work together
Many processes call for both, at different stages.
Upstream of a predictive model, a language model transforms free text into usable data. It extracts from a claim declaration the cause, the date and the circumstances, which become variables for a scoring model.
Downstream, it puts into words a result calculated elsewhere. The predictive model produces the forecast, the language model writes the commentary intended for the business owner, with the instruction not to change any figure.
The reverse arrangement is the trap to avoid: asking an assistant to produce the forecast or the score itself, because it is already in place and it always answers something.
What the distinction changes in a committee decision
The budget is not split between two competing envelopes. It is decided use case by use case, and the model family follows from the need. The cost items differ: a predictive project invests first in data preparation and model building, a generative project in the corpus, integration and evaluation, then, most often, in usage billed by volume.
The skills are not the same either. The first projects draw on statistics and data science. The second fall more under software engineering, document preparation and the design of evaluation sets.
Finally, the regulatory framework does not follow this boundary. Regulation (EU) 2024/1689 grades obligations according to the risk level of the use and provides a specific regime for general-purpose AI models, as our entry on the AI Act recalls. A predictive score used in a sensitive area can call for more requirements than a writing assistant.
Prioritising use cases and their business case falls under AI strategy. For each line of the portfolio, the question put to the committee is then reframed as follows: what must the system produce, from what data, and who will check its output?