TOTVS Unveils Design Methodology for Enterprise Data Architecture Tailored to AI Agents

Fabiane Nardon from TOTVS, a major Brazilian ERP vendor, has unveiled a design methodology for enterprise data architecture tailored to AI agents. AI agents are autonomous AI systems that execute tasks based on human instructions while referencing large volumes of data. Nardon details how companies can reorganize their data layers to enable these agents to operate efficiently.

One of the challenges enterprises face as AI agents proliferate is "token cost." Large Language Models (LLMs) process text in units called "tokens," and as the volume of referenced data increases, so do costs and processing time. In particular, when data from transactional systems—such as procurement or inventory management—is passed directly to AI, it often results in sending unnecessary information in large quantities. For enterprises with an ERP like TOTVS at their core, this challenge represents an unavoidable practical issue.

One methodology Nardon introduced is "data mesh," an architectural design approach where data is not centrally managed in one location but rather distributed, with each business domain taking responsibility for its own data. Additionally, he highlighted the importance of adopting low-latency (fast-response) database configurations to enable AI to retrieve data quickly. Furthermore, he showcased strategies for leveraging a conceptual model called "semantic ontology," which systematically organizes the semantic relationships within data, enabling AI to understand context more accurately.

He also addressed the technique of "MCP (Model Context Protocol) tool selection," which dynamically selects which tools AI should use and when based on the situation. This mechanism efficiently utilizes the information scope (context window) available to AI by selecting only necessary tools on demand, thereby reducing unnecessary token consumption. By combining these multiple approaches, the goal is to achieve a design that reduces token consumption while maintaining accuracy in agent operation.

At the same time, Nardon highlighted the challenge of balancing three critical dimensions: "accuracy," "security," and "cost." Since LLMs operate probabilistically and produce non-deterministic outputs, whereas business systems rely on deterministic processing, the fundamental challenge in enterprise AI design is determining how to reconcile these "non-deterministic LLMs" with "deterministic business logic."

What this effort demonstrates is the reality that AI agent implementation depends not only on "model performance" but also deeply on "data design." No matter how powerful a model is, if the data passed to it is not well-organized, accuracy will not improve and costs will escalate. In the enterprise domain, there is a growing recognition that reviewing the data foundation itself prior to AI adoption is essential. Going forward, insights into such data design principles may become a critical competitive factor that determines the success or failure of AI adoption initiatives.