Generative AI or classic machine learning: choosing the right tool
- 1The right criterion is not how modern the technology is but the nature of the task: predicting from structured data, or understanding and producing language.
- 2Classic machine learning is usually more stable, reproducible and auditable for numerical decisions; generative AI excels on unstructured text and documents.
- 3At SME scale, the usage cost of a generative model is rarely the issue; the real cost sits in human supervision, data quality and compliance.
- 4Hybrid designs, where generative AI structures information and a classic model or rules decide, are often the most robust.
For several months the same question has come up in executive committees: should we "put generative AI" on this process? The question is poorly framed. It starts from a tool, when an investment decision should start from a task. Generative AI and classic machine learning do not solve the same problem, do not fail in the same way and do not carry the same compliance obligations. Choosing one for the other is like buying a truck to deliver an envelope, or the reverse.
Two families, two logics
Classic machine learning: predicting from structured data
Classic machine learning covers regression, decision trees, classification and time-series models. It learns from tabular data (invoices, order histories, sensor readings, HR records) to produce a value or a category: a probability of late payment, a demand forecast, a machine failure risk score.
Its strengths are well established. Results are reproducible: the same data gives the same output. Error can be measured against history, with indicators a finance team understands. The model is usually light and inexpensive to run, and the factors driving it can be explained. Its limit is that it needs clean data, a well-defined scope and preparation work that is almost always underestimated.
Generative AI: understanding and producing language
Generative models, including large language models, work on unstructured content: contracts, emails, reports, manuals, tenders. They summarise, classify, extract, rephrase, translate and draft. Their main contribution is handling, without heavy configuration, texts that could not previously be processed automatically.
The trade-off is just as structural. The same request can produce different answers. The model can state something inaccurate with full confidence. Error is hard to quantify in advance, and explainability is limited. This is not a defect to be fixed but a property to be managed through process design.
Four questions to decide
With clients we use a deliberately simple grid.
1. What is the nature of the input? Tables and numerical series point to classic machine learning. Texts, documents and correspondence point to generative AI.
2. What is the nature of the output? A number, a probability, a class from a closed list: classic. A text, a summary, a draft: generative.
3. What does an error cost, and can it be explained? If the output triggers a decision with significant effect on an individual (credit, hiring, individual pricing), reproducibility and traceability come first. If the output is a draft reviewed by a person, tolerance is higher.
4. Would a simple rule do? Some use cases presented as AI can be solved with a pivot table, a business rule or better organisation. The best model is sometimes no model.
A worked example
Take a distribution company of 400 people with two candidate projects. The figures below are working assumptions, not measurements.
First project: anticipating late customer payments. The company holds three years of history, about 30,000 invoices in its ERP, with amounts, terms, customer sector and reminder history. Structured input, output as a probability, prioritised collection actions. This is classic territory: a classification model is trained and tested on history, the finance team sees which variables carry weight, and performance can be tracked monthly. Asking a generative model to "guess" late payments from tables pasted into a prompt would give an unstable result, hard to audit, with no gain in accuracy.
Second project: the customer service team receives on average 15,000 emails a month, to be sorted and prepared for reply. Unstructured input, text output: this is generative territory. Assume a usage cost in the order of CHF 0.01 per email processed. That is about CHF 150 a month. The technical cost is negligible. The real cost is elsewhere: if 5% of cases need detailed review at three minutes each, that is 750 emails and roughly 37 hours a month. It is this supervision, not the model invoice, that must be budgeted and organised.
The lesson: at SME scale, the price of the model is rarely the criterion. Data quality, supervision and compliance are.
The compliance angle: the technology matters less than the use
Neither Swiss law nor European regulation reasons by family of algorithms. They reason by use and by risk.
In Switzerland, the revised data protection act, in force since 1 September 2023, applies whenever processing touches personal data, whatever the model. Two points stand out. First, informing individuals when a decision is made solely by automated means, which is more likely with a predictive score than with a drafting assistant. Second, disclosure of data abroad: using a generative service hosted outside Switzerland requires knowing which data flows there, under what safeguards and with what retention. A case such as client data privacy compliance shows that this mapping is the real prerequisite.
On the European side, the EU AI Act entered into force on 1 August 2024 with phased application: prohibitions since February 2025, obligations for general-purpose models since August 2025, most other obligations from August 2026. A classic scoring model used to assess the creditworthiness of individuals falls under high risk, while a generative drafting assistant is mainly subject to transparency obligations. In other words, the "old" predictive model can be more regulated than the fashionable one. A Swiss company that sells into the Union, or whose outputs are used there, should check its exposure.
Combinations that work
The choice is not always binary. The most robust designs we see combine approaches: generative AI reads heterogeneous documents and extracts structured fields, then a classic model or explicit rules make the decision, with a human wherever the stakes justify it. Automated regulatory monitoring is one example: the generative model reads and summarises texts, but deciding what actually concerns the company stays with a person.
What leaders can do
Start from the task. For each idea, describe the input, the output, the cost of an error and the volume before naming a technology.
Measure before deploying. Define a set of test cases and an acceptable error threshold, including for a generative tool, so that the comparison is with current practice and not with perfection.
Budget for supervision. Count review hours, maintenance and data preparation, more than the licence.
Frame the uses. An AI governance charter sets out which data may leave the company, who validates what and how decisions are traced. Without it, tool choice happens case by case inside teams.
Plan for adoption from the start. A well-chosen tool that is poorly embedded in daily practice goes unused, which is a matter of organisation and adoption as much as of technology.
These trade-offs are part of our approach to data, AI and compliance, where AI is a lever only when it serves a precise objective. If you are weighing several options or want to check that a project is sized correctly, our diagnostic helps sort your use cases and define the first steps.
A personalised first reading in three minutes.