Databricks used Data + AI Summit 2026 on June 16 to push the lakehouse beyond analytics and into the daily operating layer for enterprise AI. The company announced Lakehouse//RT for low-latency queries, a broader Genie agent suite, CustomerLake for marketing data, new Unity AI Gateway controls, and a deal to acquire AI security-operations startup Panther.
The announcements are not just a routine product refresh. Databricks is trying to make one governed data platform serve real-time apps, AI coworkers, customer-data workflows, and security investigations without forcing companies to copy sensitive data into separate serving, marketing, or SIEM systems.
Lakehouse//RT is the real-time foundation
The most technically important piece is Lakehouse//RT, a new real-time warehouse powered by Databricks’ Reyden engine. Databricks says it is designed for operational analytics, BI, app serving, and observability workloads that need high concurrency and fast answers without moving data into a separate proprietary serving layer.
In Databricks’ own preview data, Lakehouse//RT delivered up to 16x better performance than separate real-time serving layers, with response times as low as 10 milliseconds on smaller datasets and sub-100 millisecond performance on larger ones. The company also points to a standard analytical benchmark result of sub-100 millisecond latency at 12,000 queries per second. Those figures are vendor-provided, but they explain the larger strategy: Databricks wants real-time workloads to stay under Unity Catalog governance instead of being peeled off into a second stack.
That matters because many agent and app deployments fail at the architecture layer before the model layer. If a support agent, revenue dashboard, fraud workflow, or security tool needs current data, teams often build synchronization pipelines into fast serving systems. Each copy adds cost, stale-data risk, access-control drift, and another place where regulated data can leak.
LTAP and Lakebase move Databricks closer to apps
Databricks also announced LTAP, or Lake Transactional/Analytical Processing, an architecture that blends transactional and analytical workloads around the lakehouse. The related Lakebase product is a fully managed, serverless Postgres database built for data apps and AI agents.
The practical pitch is copy-on-write database branching. Developers can create a branch of a production database in seconds, point coding-agent workflows such as GitHub Copilot agent mode at the temporary branch, reproduce bugs, and test fixes without giving an autonomous tool direct access to the live production system.
That is a concrete response to one of the hardest enterprise-agent problems: agents need realistic data and tool access to be useful, but live systems are exactly where unrestricted autonomy becomes dangerous. Branching does not solve all safety questions, but it gives engineering teams a cleaner place to debug and test agent-driven changes.
Genie One turns business context into an agent product
On the user-facing side, Databricks introduced Genie One, Genie Agents, and Genie Ontology. Genie began as a conversational analytics assistant, but the new version is meant to work across Slack, Microsoft Teams, mobile apps, MCP-based assistant experiences, Gmail, and other business tools.
Genie Agents let teams turn prompts into shareable agents that can reason over structured and unstructured data, use MCP connections, schedule tasks, generate documents, and write to external systems. Databricks says customers have already created more than one million Genie Spaces, which are governed chat experiences scoped to specific business topics.
The more interesting piece is Genie Ontology, a context layer that extracts business meaning from tables, dashboards, queries, pipelines, documents, and connected apps. Databricks says the system weighs sources using an authority method similar to PageRank, taking into account where a definition came from, who authored it, how often people use it, its connection to certified assets, and freshness. It also enforces source permissions, so a user should only see answers grounded in material they are allowed to access.
Databricks reports that Genie answered 84.5% of questions correctly on the first attempt in an internal 28-question enterprise data-analysis benchmark, compared with 52.4% for the strongest anonymized general-purpose coding agent it tested. The benchmark is small and company-run, but it points to a real enterprise issue: general AI agents often fail not because they cannot reason, but because they do not know which internal metric, dashboard, spreadsheet, or team convention is authoritative.
CustomerLake puts marketing agents inside the data platform
Databricks is also moving into marketing technology with CustomerLake, an agentic customer data platform embedded in the lakehouse. The product brings customer 360 profiles, identity resolution, segmentation, activation, and personalization into Databricks rather than a separate customer-data platform.
The company describes two core agent types. Profile Agents help marketers and data teams turn raw customer data into business-ready profiles. Campaign Agents help build audiences, recommend next-best actions, activate campaigns across channels, and optimize around business goals. The bet is that marketing teams will prefer agents working from governed first-party data over copying data into another martech system.
That pitch lands because customer data is both valuable and sensitive. A CDP that lives outside the core data platform can become another copy of identity data, consent status, behavioral events, and predictive scores. If CustomerLake works as advertised, the main advantage is not simply AI-generated campaigns. It is fewer uncontrolled customer-data copies at a time when privacy, personalization, and AI automation are colliding.
Unity AI Gateway is the control plane Databricks needs
Databricks also expanded Unity AI Gateway, its governance layer for models, agents, MCP services, skills, and tools. New or highlighted capabilities include unified AI spend visibility, granular cost attribution by user or team, hard spend caps, smart routing across models, runtime policies, content filtering, trace capture, and investigation workflows through Lakewatch.
This is where the announcements connect. If companies are going to let agents query data, use tools, write to systems, launch marketing actions, or investigate security alerts, they need a control plane that handles permissions, cost, audit trails, and incident review at runtime. Databricks is trying to make Unity Catalog and Unity AI Gateway do for agents what identity and access-management systems did for earlier generations of enterprise software.
The Panther deal extends the platform into security operations
The security piece came through Databricks’ agreement to acquire Panther, an AI SOC platform built around detection-as-code, cloud-native security data, and agentic investigation workflows. Databricks says Panther brings more than 100 out-of-the-box data integrations and will strengthen Lakewatch, its security lakehouse product. The proposed acquisition is still subject to customary closing conditions and any required regulatory clearances.
Panther is a logical fit for Databricks because security teams already face the same data problem as analytics teams, only with higher urgency. Logs, cloud events, identity signals, endpoint telemetry, SaaS activity, and business context often sit in expensive or fragmented systems. Databricks’ argument is that security operations should run on a governed data lakehouse with agents that can triage alerts, gather context, and propose next steps.
The risk is that agentic security workflows can create a new class of operational mistakes if actions are poorly scoped or poorly audited. For Panther and Lakewatch to matter beyond the marketing language, buyers will need clear evidence around detection quality, investigation accuracy, access controls, data-retention costs, and how human analysts approve or reverse agent-proposed actions.
What enterprises should watch next
Databricks is making a coherent argument: enterprise AI will not be won only by model choice. It will be won by platforms that can connect models to trusted data, preserve governance, deliver low-latency answers, control token spend, and keep agents inside enforceable boundaries.
The proof will come in deployment details. Customers should watch whether Lakehouse//RT performance holds under their own messy workloads, whether Genie Ontology reduces wrong answers in real business settings, whether CustomerLake can coexist with existing martech stacks, and whether Unity AI Gateway gives administrators enough runtime control across non-Databricks tools.
The strategic direction is clear even before those tests are settled. Databricks is no longer positioning the lakehouse as a better analytics repository. It is trying to make it the place where enterprise AI agents read, reason, act, get billed, and get audited.
