Most hands in a room of 31,000 stayed down when Databricks CEO Ali Ghodsi asked whether artificial general intelligence had already arrived. Then he read a graduate-level topology problem, confirmed that every leading frontier model solves it without hesitation, and made his point: the models are already past the threshold. What hasn't crossed it yet is the connection between those models and your organization's data, processes, and institutional knowledge. That gap, not model capability, is what is holding enterprise AI deployment back.

The announcements drew attention: Genie One as a universal data co-worker, the Raider real-time engine, Unity Artificial Intelligence Gateway for cost and governance control, and the acquisition of Panther Labs for agentic security operations, Databricks' third cybersecurity deal in just over a year. The company's acquisition of Neon , its serverless Postgres play announced in May 2025, provided the technical foundation Ghodsi drew on repeatedly when discussing agentic database provisioning.

"Raider" is also Databricks' internal shorthand for "Raiders of the Last Data Silos," a nod to Lakehouse Federation and the founders' apparent fondness for naming things that annoy them.

The context gap is the actual bottleneck

The assumption in most enterprise AI deployment planning is that model selection determines outcome quality. Pick the right model, tune it, deploy it. Ghodsi's argument, backed by internal benchmark data his team presented, cuts against that assumption directly.

When Databricks tested the best available coding agents against a set of real enterprise data questions, questions that actual employees submitted through Genie, the most accurate agents answered correctly roughly half the time. A coin flip. The questions weren't exotic. Examples given included: what percentage of customers who use one product also use a third-party service; what is engagement broken down by region this week. The kind of questions a CFO or a product lead needs answered before a board meeting.

The problem isn't reasoning. Frontier models can reason at graduate-level mathematics. The problem is that these agents have no reliable map of how your organization captures data, what your internal terminology means, which tables hold which numbers, or who the authoritative source on a given metric is. Faced with a question, they go searching live, consuming time and tokens, and they frequently surface stale documents or hallucinate figures when the search fails.

Genie Ontology is the bet on a different architecture

The Databricks answer is Genie Ontology, a background knowledge graph construction system that runs continuously against an organization's data assets rather than at query time. The distinction matters. Current agent architectures search for context when a question arrives, which is why they are slow, expensive, and unreliable on proprietary enterprise questions. Genie Ontology indexes data pipelines, dashboards, Jira tickets, calendar data, SharePoint, Google Drive, and any connected source into a knowledge graph that agents can query in milliseconds.

The ranking algorithm the research team developed, described as functioning on organizational data the way Google's PageRank algorithm worked on the web, surfaces high-authority snippets based on who created an asset, how frequently it is accessed, and what domains a given employee has expertise in. A definition of engagement stored in a mobile key performance indicators dashboard, authored by a data scientist with demonstrated authority in that domain, gets weighted accordingly.

A demo using Databricks' own internal instance showed more than 4.5 million ontology snippets already indexed for a single company deployment. No live search agent could traverse that corpus at query time and return a trustworthy answer in seconds. The ontology makes the traversal unnecessary.

The accuracy improvement claimed is a consistent 30% lift over coding agents on the same questions, with the gap widening as the ontology is refined. That number, presented by Senior Director of Product Management Ken Wong during the Genie One session, deserves stress-testing in any proof of concept before budget commitments follow. The underlying mechanism is sound: an organization that encodes its semantic layer, business definitions, and institutional context into an automatically-maintained knowledge graph gives agents a reference that static prompting cannot replicate.

The Customer Evidence Behind the Accuracy Claim

The customer cases on stage were not chosen at random. PepsiCo's Global Chief Data and Artificial Intelligence Officer Magesh Bhagavathi described six years of consolidation from more than 60 data silos to roughly 8, with 90 percent of the company's data now on Databricks. Within the first weeks of deploying Genie, the company recorded 30,000 query engagements from procurement and indirect spend leaders who had previously depended on dashboards and reports. The direction of the shift, from static reports toward live conversational queries against governed data, is exactly what Ghodsi is predicting becomes standard.

Mastercard presented a different dimension of the same problem. Operating across more than 200 countries, processing over 150 billion transactions per year, the company is building a shared agent foundation on Lakebase, the company's term for serverless Postgres on top of the lakehouse. The requirement is simultaneous isolation and learning: each issuing bank's data must remain separated for regulatory and trust reasons, while the agents must accumulate insights across the full portfolio.

That tension, isolation at the data layer and shared learning at the intelligence layer, is one of the harder unsolved architectural problems in enterprise AI right now. The Mastercard team built a concept-to-demo MVP in weeks rather than months, which they attributed to governance and guardrails being baked into the architecture before the first agent was deployed.

The agent ops problem most teams haven't priced in

One announcement that got less attention than it deserves is Genie Zero Ops, a background agent that monitors data pipelines and machine learning models, detects anomalies, traces root causes through data lineage, and proposes fixes, including by writing and testing code in a shadow branch of production data before presenting it for human approval.

The case rests on a known cost: data engineering teams report spending more than half their time on incident response and maintenance rather than building new pipelines. As coding agents write more of those pipelines faster, the surface area for failures grows in proportion. Zero Ops runs inside the data plane, where it has access to production data lineage, compute logs, and the Unity Catalog governance layer. That governance boundary is what makes autonomous write access to production feasible at all, because the operations agent cannot exceed permissions that a human engineer would not have.

The cost argument Ghodsi made for the full platform is worth taking seriously. As agentic workloads scale, token costs that seem manageable in a pilot can become prohibitive at production volume. He referenced an unnamed chief executive who depleted a full year's artificial intelligence budget ahead of schedule. Unity Artificial Intelligence Gateway, which provides a single control plane for model spend, identity management, and compliance across all agents and all model providers, is Databricks' answer to that problem, and it is priced as a platform play rather than a premium add-on.

What the portability argument still leaves open

Earlier this year, in coverage of the Delta Lake and Apache Iceberg format convergence announced ahead of this summit, I traced a portability thesis: the format question is now resolved, but governance portability and performance portability remain open problems. That characterization held at Monday's keynote.

Ryan Blue, creator of Apache Iceberg and Chief Executive Officer of Tabular, joined Ghodsi on stage to confirm that Iceberg V3 support is generally available, and that a unified metadata layer built into Delta 5 and Iceberg V4 is targeted for later this year. The format argument is closing. But Genie Ontology opens a new portability question: the knowledge graph. If an organization builds several years of indexed institutional context inside Databricks' Genie system, what does migration of that ontology look like? This was not addressed on stage, and it is the question procurement teams should be asking now.

The Raider real-time engine, announced as the technology powering the new Lakehouse Real Time warehouse, adds a third layer to the portability question. Reynold Xin, who co-founded Databricks, demonstrated benchmark results showing Raider sustaining 12,000 queries per second at a 37-millisecond P90 latency under load that caused competing serving engines to crash or spike. The engine uses machine learning trained on Databricks' own query traces to predict and implement the right algorithmic approach before executing, and that learned optimization accumulates only inside the Databricks environment, with no equivalent available if you move to another platform.