TOTVS, a major Brazilian enterprise software provider, is optimizing its data architecture to support enterprise-grade AI agents by balancing deterministic software models with probabilistic generative AI reasoning.

Key Points

• TOTVS manages systems processing approximately 25% of Brazil's GDP, requiring high-precision, enterprise-grade AI agent integration.
• The company utilizes a "data mesh" architecture where data products are treated as microservices with stable interfaces and clear ownership.
• A multi-layered data platform provides low-latency access via Postgres and DuckDB, alongside high-latency batch processing for historical data.
• Semantic Web standards, including RDF and ontologies, are used to reduce ambiguity and improve LLM response precision by up to 40%.
• To manage token costs and context limits, TOTVS implemented "MCP Fabric" and dynamic tool searching to inject only relevant tools into agent contexts.
• Security is maintained through identity propagation, ensuring agents only access data authorized for the specific user via OAuth tokens.

Why it Matters

Integrating AI agents into legacy transactional systems requires a strategic division between deterministic code and non-deterministic AI models to ensure security, precision, and cost-efficiency. By adopting standardized data governance and semantic layers, enterprises can scale AI capabilities without sacrificing the reliability expected in mission-critical business environments.