Databricks unveils new data architecture and warns companies must change processes for AI

Databricks CEO Ali Ghodsi told a Data + AI Summit audience on June 16 that companies trying to add AI into their existing workflows without a firm grip on costs and their data are heading for trouble. He warned that the model at the heart of an AI agent can disappear overnight, making process change non-negotiable.

The company simultaneously announced a new architectural approach called LTAP, short for Lakehouse Transactional/Analytical Processing, which merges the two databases every company runs: the operational database for live transactions and the analytical database for business intelligence. Databricks also launched Lakehouse//RT, a product that delivers millisecond query latency directly on governed Delta and Iceberg tables without data movement.

LTAP collapses decades-old OLTP/OLAP divide

For decades, data professionals have kept operational systems separate from analytical systems to avoid performance degradation. Agents that reason continuously and act on live data cannot tolerate a pipeline between themselves and the information they need, Databricks argued. The LTAP architecture aims to eliminate that latency by providing a single lakehouse that supports both transactional and analytical workloads on the same data. Key features include:

• Unified governance across Delta Lake and Apache Iceberg tables, removing the need to copy data between SQL and analytics environments.
• Real-time ingestion and query support with Lakehouse//RT, which promises millisecond latency directly on governed tables.
• Native support for AI agent workloads that require continuous access to live operational data without batch pipelines.
• Compatibility with existing Spark, SQL, and ML workflows, reducing migration risk for current Databricks customers.

Process change more important than technology

Ghodsi emphasized that the technology shift alone will not deliver results. Companies must redesign their processes to manage costs per query, govern data access for agents, and handle model churn when a foundational model is replaced. He noted that agents built on top of a single model face existential risk if that model is deprecation or replaced by a competitor. The Databricks platform now supports multi-model orchestration to mitigate that risk.

The LTAP announcement follows months of industry debate about the best data architecture for AI agents. Databricks is betting that most enterprises will not want to maintain separate operational and analytical systems as they deploy agents that both read and write live data. The company faces competition from Snowflake, which has its own transactional ambitions, and from startups building purpose-built agent databases.

What comes next: Databricks plans to roll out LTAP in preview during Q3 2026. Early access partners include financial services and ecommerce firms that need sub-10-millisecond latency on transactional data for fraud detection and real-time personalization. The company expects general availability by early 2027.