AI
How to Build Data Architectures That Work for AI Agents
Enterprise organizations must rethink their data platforms to serve AI agents, combining transactional systems with semantic models and hybrid data access patterns.
InfoQ
Enterprise software development is facing a fundamental shift: the arrival of AI-powered agents that reason probabilistically rather than deterministically. For organizations like TOTVS—a leading Brazilian enterprise systems provider—this shift poses a critical challenge: data that was optimized for applications and dashboards is not optimized for agents.
The New Data Problem#
Traditional enterprise systems face a mismatch when powering AI agents. Applications needed data organized for consistent, predictable queries with near-perfect accuracy guarantees (99.99% uptime). Agents, by contrast, fire hundreds of unpredictable queries in minutes and operate on probabilistic reasoning rather than absolute certainty. Meanwhile, data warehouses were built for human analysts, not for token-hungry AI systems.
The core issue is that enterprise-grade AI agents require data prepared specifically for their needs—not repurposed datasets from legacy systems.
Deterministic and Non-Deterministic Computing#
Successful AI-powered applications don't choose between traditional software and AI; they combine both. The decision hinges on a three-variable formula:
- Precision – accuracy of results
- Security – risk exposure
- Cost – resource consumption
Each variable depends heavily on how data reaches the system. Teams must decide which operations should remain deterministic (handled by coded business logic) and which should be probabilistic (handled by LLMs).
Choosing the Right Data Source#
Architects face a hybrid choice: transactional systems or data platforms. Each serves different needs:
Access the transactional system when:
- Writing data back to the system
- Requiring real-time, fresh data
- Executing business rules embedded in production code
Access the data platform when:
- Using slightly stale data is acceptable
- Performing historical analysis
- Running semantic search or data enrichment
- Combining external data sources
The answer is rarely "one or the other"—successful designs draw from both, routing requests based on workflow requirements.
