Generative AI or traditional AI, which one does your problem need
Two families of technology that solve different problems. How to tell which one fits each case, and why choosing wrong makes projects expensive.
The conversation about artificial intelligence has focused so heavily on language models that many organizations try to solve problems with them that belong to a different family of techniques. The result is expensive: imprecise solutions to problems that had a better and cheaper answer.
Two families, two kinds of problem
Predictive AI works with numbers and categories derived from historical data. It answers questions of the form how much, when or how likely. Sales forecasting, credit risk, churn probability, fraud detection, predictive maintenance.
Characteristics: it needs history, produces quantifiable results, its accuracy can be measured precisely, and it's auditable.
Generative AI works with language and content. It answers questions of the form what does this say, how do I explain it or draft this for me. Summarizing documents, classifying enquiries, extracting data from text, drafting, answering questions about documentation.
Characteristics: it needs no historical series of its own, handles unstructured information well, and its evaluation is more qualitative.
How to tell which one fits
One question is enough: is the answer a number or a text?
If you need a number, a probability or a category derived from tabular data, that's predictive AI. If you need to understand, summarize or produce language, that's generative AI.
Examples of the most common confusion:
- We want to predict which customers will leave. Predictive. Using a language model here produces worse and costlier results.
- We want to understand why customers complain. Generative: free text has to be processed and themes extracted.
- We want to know how much stock to order. Predictive.
- We want to draft the right message for each customer's situation. Generative.
Why choosing wrong is expensive
Using a language model for a numerical task creates three problems: the result is less accurate, the cost per query is far higher than a statistical model, and explainability is worse exactly where it matters most.
In the other direction, trying to handle free text with rules or traditional models produces brittle systems that need constant maintenance and break with every format variation.
The combination that works
The most useful production systems usually use both families, each on its part of the problem.
An example in collections: a predictive model estimates each account's probability of payment from history, ageing and behaviour. A language model takes that output and prepares the appropriate message for each case, in the tone that fits the segment. The team approves and sends.
Another in industrial quality: a model detects anomalies in sensor data; a language model writes the incident report with the context the next shift needs.
Neither component would have solved the whole problem alone.
What each one costs in effort
A predictive project requires historical data, preparation work and statistical validation. It's slower to start and more precise when it works. The effort sits in the data.
A generative project requires access to documentation, scope definition and output quality evaluation. It starts faster and needs more fine-tuning of behaviour. The effort sits in definition and evaluation.
That difference matters for planning: a generative case can be showing results in weeks; a predictive one rarely before two or three months.
The question that comes before everything
Before choosing a technology family, answer something more basic: which decision will change because of this output?
If nobody will decide differently based on what the system produces, the project needs neither technology. That question eliminates more projects than any technical assessment, and it's worth asking early.
Frequently asked questions
When should we use generative AI and when predictive models?
Generative AI is for working with language: reading, writing, summarizing, classifying text. Predictive models estimate numbers or probabilities from historical data, such as demand, risk or customer churn.
Can a language model forecast sales?
It isn't the right fit. Forecasting calls for statistical or machine learning methods trained on the historical series; they're more accurate, cheaper and auditable.
Can both approaches be combined?
Yes, and it's common. A predictive model computes the risk or the demand, and a language model explains the result in readable text or prepares the corresponding communication.