Zaion outlines its criteria for selecting AI models

Zaion selects its AI models based on their price, quality, latency, sovereignty potential, and server and carbon footprint. The company, which combines AI engineers and customer relationship specialists, automates exchanges and equips advisors with real-time tools.
Zaion Tests and Replaces Its Models
The choice of a model is not final at Zaion, explains Imed Laaridh, PhD in computer science and VP of AI & Data at the company. He describes a portfolio with expiration dates: a solution is replaced when another surpasses it according to the selected criteria. The company claims to own its GPU servers to conduct evaluations and research and development work.
The process begins with real internal data, particularly for open-source models. This is followed by so-called "shadow testing": production flows are processed in parallel, without any actual impact on the service, the manager specifies.
According to him, these tests are complemented by automatic quality checks and human oversight. In partially automated improvement cycles, the AI suggests modifications, but a person ultimately validates them. Imed Laaridh presents this division as a way to maintain traceable and explainable decisions.
Models Tailored for Customer Relations
For Imed Laaridh, generic scores are not sufficient to measure a model's usefulness: power must be assessed in light of the expected service. The automatic resolution rate and response time carry different weights depending on the sectors. An acceptable delay for asynchronous processing can become detrimental in a live conversation.
The manager puts forward a hypothetical scenario: a customer service agent uses a cutting-edge model whose in-depth reasoning takes several seconds, or even minutes. Even if correct, the response might arrive after the customer has hung up. He associates this wait with increased drop-offs, repeated inquiries, overwhelmed advisors, and degraded profitability.
To classify requests or transcribe speech, small specialized models can, according to him, perform better than a large general-purpose language model. He cites natural language understanding for classification and automatic speech recognition for transcription, with an adaptation to Zaion's data. He attributes potentially better accuracy, faster responses, and lower costs to them. At the same time, he believes that ChatGPT has accustomed users to more natural and complex dialogues.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
However, customers access a platform developed by Zaion, not directly the models, the executive indicates. The company presents this interface as a safeguard that standardizes exchanges, preserves their transparency, and takes into account the skills of the teams. It claims that the interface remains the same during a replacement and that users do not need to master the internal mechanisms or adapt the models themselves. Preliminary checks would, according to it, ensure that neither their experience nor the practices of business teams are altered.
Costs and Sovereignty in Selection
The Zaion Lab grid does not separate technical results from operational constraints: price and sovereignty coexist with quality, speed, and material and carbon footprint. Imed Laaridh notes an increasing importance of inference costs and carbon. He also claims that some reasoning operations are charged even though they remain invisible to the user.
The manager describes divergent pricing trends: less efficient models would cost less, while the price per task for cutting-edge models would often increase, especially with reasoning. He mentions GPT-6 and Claude Fable in this context. According to him, this trend encourages companies to consider smaller models, often open-source, that they can adapt and host locally.
Data control is another criterion, particularly in banking and insurance, he insists. He sees providers subject to the American Cloud Act or extraterritorial legislation as a risk of geopolitical dependency. He specifically states that American export restrictions temporarily suspended access to certain models in June 2026.
Imed Laaridh advocates for a sovereign infrastructure to maintain control over data and models in enterprise services. Zaion claims to control its own models on essential components. For large language models from third parties, the company states it imposes hosting within the European Union and masks the data.
Laaridh Anticipates Requirements on Decisions
Looking ahead five to ten years, Imed Laaridh envisions a more thorough justification of carbon footprints, possibly accompanied by regulatory frameworks encouraging sobriety. He primarily anticipates increased obligations to trace and explain algorithmic decisions. The granting of a mortgage is the example he cites as a decision that should be justifiable. These requirements would, in his view, favor understandable models governed locally.
The manager also envisions rules facilitating the replacement of one AI provider with another. He draws a parallel with the common USB-C charger mandated by the European Union, which prevents iPhone users from relying on a specific cable. In AI, regulators could, according to him, impose easier export of data and accumulated intelligence to reduce dependencies and allow for provider changes.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.