Frugal AI: When Simplicity and Quality Go Hand in Hand

• AI
• CSR
• Deep learning
• Machine learning

Share on Twitter : Frugal AI: When Simplicity and Quality Go Hand in Hand (New window) Share on Linkedin : Frugal AI: When Simplicity and Quality Go Hand in Hand (New window)

Frugal AI: continuity of best practices rather than revolution

In the world of data science, a quiet revolution is underway, although it is rooted in existing “algorithmic heritage.” Contrary to the common belief that the most powerful systems are necessarily the most energy-intensive, a comparative study of several technical solutions highlights empirical results that tend to demonstrate the opposite. By analyzing a corpus of tweets from customer comments in the air travel sector, researchers at Orange compared different approaches to predicting the category of customer dissatisfaction (lost luggage, delayed flights, in-flight service quality, etc.). The results are clear: the most frugal approaches can outperform the most energy-intensive systems, including the latest generative AI models.

La frugalité n’est pas un arbitrage au détriment de la qualité. Au cas par cas, on gagne souvent sur les deux tableaux !

Comparative approaches

The study evaluated three categories of approaches, listed below from the most energy-intensive to the most energy-efficient: The tests were conducted in two main scenarios.
1. “LLMs” (Large Language Models), in other words, generative AI (such as ChatGPT). Four of these solutions were compared: GPT-4.1-nano, GPT-4.1-mini, GPT-4.1, and Claude Sonnet 4. First, without providing examples (the “prompt” technique is called “zero-shot”): we provide only the tweet to the AI and ask it to classify the tweet, while requiring it to choose from a defined list of categories (lost luggage, delayed flight, in-flight service quality, etc.). Next, with a few examples provided (known as “few-shot”): we start by providing the AI with a number of examples in the “prompt,” specifying that a given tweet corresponds to a certain category, another tweet corresponds to a different category, and so on. Then, as before, we provide the AI with the tweet and the list of categories and ask it to identify the category of that tweet, based on the examples provided.

2. “Embeddings” + machine learning: This approach uses statistical algorithms considered standard in the field of prediction and probability: “Light GBM,” “Random Forest,” and “Logistic Regression.” There are many others .These algorithms are based on machine learning. They all require “training.” In other words, these systems must first be provided with a large number of examples—consisting of tweets and their categorizations—so that they can “learn” to recognize tweets and subsequently be able to “recognize” tweets they have never seen before. These algorithms process only numerical data, not text. It is therefore necessary to transform the various tweets into vectors that constitute a mathematical representation of these tweets; each tweet will thus be “converted” into vector form (for math buffs: in 512 dimensions, or more): this is what is called “embedding” (or encapsulation). This transformation into vectors, which first appeared in 2013 with “word2vec” [https://arxiv.org/abs/1301.3781], is nevertheless quite computationally intensive, as it uses “deep learning” techniques (learning through “neural networks”) requiring several (hundreds of) thousands of calculations. It should be noted that these vectors are in no way directly interpretable by a human being. It is then these vectors (representing the tweets) that can be processed by statistical algorithms to predict their category (lost luggage, delayed flight, in-flight service quality, etc.).

3. The “naive baseline”: It uses the Khiops software (an open-source solution “by Orange” [https://khiops.org/]), an open-source solution developed by Orange for data analysis and machine learning that stands out for its simplicity and effectiveness. Our study compared versions 10 and 11 of Khiops, using the words in the tweets as-is (without embedding calculations) and sequences of consecutive words, which provide additional information. Example: the presence of the term “bad” in a text does not necessarily imply a negative meaning; on the contrary… if it appears specifically in the sequence “not bad,” it is therefore necessary to analyze these combinations rather than individual words taken in isolation.

Speak for themselves

The findings of this study are particularly illuminating:

• Predictive performance: The second family of approaches (embeddings + statistical algorithms) is by far the most effective in terms of predictive quality (“recall,” “precision,” and “F1-score ” [https://fr.wikipedia.org/wiki/Précision_et_rappel]), even when compared to the largest LLM model.
• Robustness: This family of solutions remains effective even for “catch-all” classes (those without a homogeneous theme: “dissatisfaction during the flight,” for example, can have a thousand different reasons), whereas LLM performance collapses in such cases (but this is logical and inevitable: no algorithm can assign a text to a category that is inherently vague from the outset, unless it has a very large number of examples).
• Energy Efficiency: Over 99% of the energy cost of using the embeddings + statistical model solution is due to the embeddings themselves (which makes sense, since calculating a conceptual vector is still deep learning!). LLMs consume between 1.6 and 18 times more energy in use than this family of solutions.
• The Naive Solution: The Khiops 10 solution with undifferentiated bag-of-words is slightly inferior in predictive quality to LLMs (admittedly, 72% vs. 82% on the primary target class… and 66% vs. 68% on the average of the 5 main classes), while consuming between 150 and 1,500 times less energy! Note that this solution also consumes nearly 100 times less energy than the embeddings + statistical model solution, albeit with a slight degradation in results.
• Response time: Khiops makes a prediction in 1 to 2 milliseconds (ms) on a business laptop. On the same hardware, the embeddings-based solution takes between 30 and 40 ms (almost entirely due to the cost of calculating the vector). The request to the LLM (for which network latency can reasonably be ignored) takes nearly 400 ms.
• Side note: Moving slightly away from the initial objective (identifying tweets in the “lost luggage” category), we finally note that across all five main categories, version 11 of Khiops—which here processes only simple, non-vectorized word groups—produces (by a very narrow margin) better predictions on average than all the LLMs…

Comparative summary table

Toward a more responsible AI

These findings confirm the need to choose “the right model for the right use” and to think beyond the “all-generative-AI” mindset. They pave the way for a more responsible AI, where performance and energy efficiency are not viewed as an inevitable trade-off but can, on the contrary, coexist.

It is not a matter of imagining that energy-efficient AI solves all problems. There would be no generative AI at all if that were the case. LLMs prove indispensable for any “intrinsically generative” content (computer code generation, “reasoning” on a text, and more generally any task that cannot be analyzed by traditional statistical algorithms (including by deep learning prior to the generation of LLMs). The goal is rather to implement best practices for choosing the most frugal AI possible wherever they can be applied. To do this, annotated datasets are required for training, and some time must be dedicated to the process. This requires a degree of initiative, but the gains in cost control, environmental impact, and explainability can be substantial in the end.

Moving toward responsible AI also means exercising discernment regarding its uses and, above all, the levels of performance that are truly necessary.

This study reveals that Khiops delivers results comparable to large language models (LLMs), while reducing energy consumption by 150 to 1,500 times. Furthermore, when compared to techniques combining “embeddings” and statistical models, the Khiops tool—even if slightly less accurate—consumes about 100 times less energy. Given these impressive figures, a question arises: do we really need the slight performance difference offered by more energy-intensive solutions? Does this slight difference in quality justify consuming 100 times more resources?

Orange researchers are already planning to expand these studies in two complementary directions:

• the new context brought about by agent-based AI;
• the possibility of detecting, over time, “run” situations where frugal AI could theoretically be relevant (a matter of rationalization and cost control).

We hope to be able to discuss this further with you sometime in 2026!

This text has been translated by an artificial intelligence.