Databricks bet the enterprise AI race on knowing the business, not on the biggest model

The coworker pitch

Databricks introduced the new Genie as an autonomous AI coworker for enterprise work, and the pitch is not about a bigger model. Genie Ontology provides a map of business definitions, metrics, relationships, and organizational knowledge, and Genie Search combines content indexing from business apps with live search through MCP. Genie One is an enterprise home for AI cowork, and it uses Genie Ontology for enterprise context. m:…4911 m:…4916

The product is built to act, not just answer. Genie One can read and update systems like Jira or Asana via MCP, and Genie Search indexes knowledge from Google Drive, Jira, Confluence, Workday, and SharePoint, supporting over 100 MCP connectors. Genie retains personal memory of user preferences and instructions, and can build artifacts like slide decks or dashboards directly in the conversation. Genie Agents can work across structured tables, PDFs, and slides. m:…4916 m:…4911

Governance is part of the pitch, not an afterthought. Unity Gateway governs Genie within existing permissions and policies, and Genie Search enforces source permissions through Unity Catalog connections. Genie Ontology is an agentic ontology that learns from enterprise data assets, and it is built automatically from existing dashboards, queries, and notebooks. OntoRank evaluates and ranks knowledge by authority, so agents prioritize certified definitions over learned context. m:…4911 m:…4912

The benchmark behind the claim

Databricks developed a new enterprise data-agent benchmark in collaboration with Turing, and the headline result is striking. Genie achieved a 91% pass rate in the new benchmark, compared to 45% and 58% for two other frontier agents. The benchmark results show that Genie's correct answer pass rate is 46% better than Agents A and 33% better than Agents B. m:…4915

The benchmark is designed to be hard in the way enterprise work is hard. It includes notebooks, dashboards, saved SQL queries, and documents in addition to table schemas, and it uses a fictional company, Calderwick Energy, to simulate real-world data team challenges. Genie uses fewer than half the tokens compared to the other frontier agents in the benchmark, and Genie is 4x faster than Agents A and 3x faster than Agents B. The argument is that context, not compute, is what makes an agent useful inside a company. m:…4915

The same benchmark result is reported across Databricks' own announcements: a benchmark with Turing showed Genie achieved a 91% pass rate, and Genie achieved a 91% success rate in a benchmark test against frontier agents. Genie Ontology is now generally available, and it is used by 7-Eleven, Virgin Atlantic, and General Motors. m:…4912 m:…4911 m:…4915

Routing is the other half

Glean is solving the same enterprise problem from the routing side. Glean auto-routing matches requests to models based on task complexity, and it evaluates models using 5,000 real enterprise tasks across 37 model-and-reasoning configurations and 11 different models. Approximately 25% of Glean requests are simple lookups, routed to fast, lightweight models, while approximately 75% of requests are handled by a main agent that receives context from previous searches, files, and tool results. m:…9260

The routing layer also knows when to hand off. Glean identifies experts by comparing model performance on specific enterprise use cases, and frontier models were preferred 68% of the time for exhaustive analysis. An AI judge verifies every statement in the response during evaluation. The pattern is the same as Databricks': the model is not the product. The system around the model, the context, the routing and the governance, is the product. m:…9260

The money behind it

Databricks is putting real weight behind the bet. The company opened a new 235,000-square-foot headquarters at One Sansome Street in San Francisco, and is investing over $1 billion into its San Francisco operations by early 2028. Databricks reported a $7 billion revenue run-rate in August 2026, and the 2026 Data + AI Summit drew over 30,000 attendees. Genie One is free for users through January 31, 2027, which is the classic move: get the coworker into the building first, and let the ontology learn the business before the bill arrives. m:…5218 m:…4916

The company is also building the pieces around Genie. KARL is a custom enterprise model that outperforms frontier models on inference cost, latency, and quality, according to Databricks, and Lakebase is a serverless Postgres database built for AI agents. MemCurrent is a persistent memory architecture for production agents that maintains information consistency. Databricks has over 12,000 employees worldwide, and has acquired companies including Mosaic, Neon, Tabular, Row Zero, and Panther. m:…5218