The debate between local and cloud AI is becoming increasingly relevant in the retail sector, where agentive AI is introducing a hybrid approach to balance cost, security, performance, and data control. Retailers are now considering the cost of AI queries as a critical factor in managing their operating margins. Traditionally, AI's cost was measured by individual tasks, such as generating a product description or answering customer questions. However, with AI agents now capable of performing complex, autonomous tasks—like comparing prices, processing transactions, and managing workflows—the number of AI calls has surged into the tens or even hundreds of thousands per day. This shift transforms AI from a budgetary concern into an architectural challenge. For example, a pricing agent that continuously monitors and adjusts thousands of product references might make tens of thousands of daily checks. Similarly, a purchasing agent that compares offers and finalizes orders without human input can multiply calls at every stage of the customer journey. These operations are not occasional but continuous, making the cost per action more significant than the cost per token. While billing by usage is ideal for unpredictable volumes, it becomes less efficient when tasks are repeated thousands of times daily. In such cases, the cumulative cost of a token-based model may exceed the cost of maintaining an internal infrastructure. Hosting AI models on local servers offers three key advantages for high-volume tasks and sensitive data: cost control independent of external pricing, data sovereignty by keeping information within the company, and enhanced security, especially when handling customer or competitive data. However, this approach is not without its challenges. It requires an initial investment in hardware, maintenance, and internal expertise, and it may limit the model's reasoning capacity compared to advanced cloud models. Thus, the decision is not simply about choosing between local and cloud AI, but determining at what volume and data sensitivity local hosting becomes more cost-effective. The optimal solution lies in a hybrid architecture, where local AI handles high-volume, low-complexity tasks such as catalog checks and duplicate detection, while cloud AI manages complex or rare operations like generating detailed content or interpreting ambiguous queries. The success of agent-based commerce in retail will depend not just on the choice of AI models but also on the quality of product and media data. Well-structured data with clear taxonomies and business rules allows more tasks to be handled by economical local infrastructure, ensuring reliability for autonomous agents. Retailers that build and maintain such hybrid systems will likely gain the most value, adapting their architecture as their usage volumes and security needs evolve. For more insights on this topic, join us at Tech for Retail 2026 on November 30 and December 1 at Paris Expo, Porte de Versailles.