AVA
AVA - Orchestrating generative AI at the scale of a global group
Le contexte
Havas Group chose to equip all of its employees with the best AI models available, within a unified, secure and traceable framework. 23,000 people, 50 countries, dozens of agencies with their own structures and their own needs. The challenge was as much organisational as technical: how do you give everyone smooth access to the latest LLMs while keeping control of costs, permissions and data at group scale?
The engines already existed: Claude, GPT, Mistral, Gemini. Powerful, constantly improving, available. But a subscription with a single provider is like driving a car whose manufacturer keeps the keys. Havas wanted its own car. With the freedom to choose the engine according to the need, and without depending on the pricing or technical changes of a single supplier. Havas entrusted the design and development of AVA to ekino.
Les objectifs
A single conversational portal, available to every employee in the group, with Claude, GPT and Gemini accessible from the same interface. Centralised management of budgets, permissions and costs by entity. Full compliance with European legal requirements on data. Delivered in three months, by a small team.
Chiffres clés de l'étude de cas
months from design to going live
3
models orchestrated simultaneously
6
employees in 50 countries
23K
European (hosting)
100%
Le(s) challenge(s)
Claude, GPT and Gemini do not speak the same language. Each exposes its own API, with its own syntax, its own response format, its own billing model. Anthropic bills cache writes where OpenAI counts differently. Every AVA feature had to be developed taking those three logics into account in parallel. The user chooses their model according to the nature of the task. The complexity stays on ekino’s side. A security gateway filters non-professional content upstream of every model, before it reaches a provider. That orchestration work is the core of AVA. It is structural, and it does not show.
On hosting, the European legal framework was a non-negotiable requirement. All the models run on European servers through two dedicated infrastructures: Microsoft Foundry for Claude and GPT, Google Vertex for Gemini. Both providers contractually undertake not to use conversation data to train their models. Access goes exclusively through enterprise endpoints.
Notre solution
What ekino built: two applications, one architecture
AI models — Claude, GPT, Gemini — are components you integrate. What ekino built is the car around them: a portal for users and a back office to steer costs, permissions and models. Everything that turns those components into an enterprise product that is usable, traceable and controlled at group scale.
The user portal: the interface employees see. Conversations, file management, image generation. It is the portal that orchestrates the mesh between providers, produces interactive artefacts that can be displayed directly, and reports actual consumption for every exchange.
The management back office: permissions by user and by agency, budgets and quotas, cost visualisation, handling of requests for additional credit. Havas management has real-time visibility on usage and spend across the group.
The rollout went market by market: France and the UK first, then the United States and Canada, then the rest of the world, in order to scale up without any interruption of service.
We know how it works because we did it ourselves. We know how to use it because we use it ourselves. Tomorrow we are going to start having agents that our experts will fine-tune. It is not outside, it is in-house.
Les résultats
With a subscription to a single provider, you depend on its changes, its pricing, its roadmap. To take the car analogy again: if the engine changes, you have to live with it. With AVA, Havas can change engine without rebuilding the car. If tomorrow Claude pulls ahead on a specific use, or if a new model takes over, AVA integrates it without a rewrite, and without employees having to change anything in their habits.
Collaborative project mode, RAG connected to specific SharePoint folders, MCP connections, the agentic part planned for autumn 2026: all features that build on a stable base. What changes from one deployment to the next is the organisational map to be modelled. The car itself stays.
Réalisations
Un projet ?
Parlez-nous de vous, de vos projets, de vos ambitions, de vos enjeux, et construisons ensemble des solutions concrètes qui contribuent à votre transformation.