Databricks Launches Genie One Agent to Solve Enterprise Data Errors

Introducing Genie One: the AI coworker that understands your data (with demo) — Databricks

Databricks

Leading AI agents fail half the time on enterprise queries because stochastic reasoning loops cannot navigate unmodeled business metrics, requiring an automated context layer that mines operational logic directly from corporate pipelines and dashboards.

AI-assisted summary of Introducing Genie One: the AI coworker that understands your data (with demo) — original source. Check the original for full context.

Source published:

Key ideas

  • Databricks introduced Genie One, an agentic AI coworker that connects across corporate systems and live databases to compute answers, automate workflows, and generate artifacts rather than merely summarizing static documents. (0:00)
  • The platform is powered by Genie Ontology, a context layer that harvests knowledge from SQL pipelines, dashboards, and author expertise using the OntoRank algorithm, driving a 30+ percentage point accuracy gain over generic coding agents. (7:10)
  • Genie One is Generally Available immediately across Slack, Microsoft Teams, iOS, Android, and external tools via a Model Context Protocol (MCP) app. (20:18)
  • To drive broad enterprise adoption, Databricks is subsidizing usage by providing $10 worth of free monthly tokens for every user in an organization. (21:23)

Detailed notes

## Background Published on YouTube by Databricks on June 23, 2026, this presentation introduces Genie One, an agentic AI coworker designed to understand enterprise data and generate documents, reports, and autonomous workflows. The accompanying description highlights its foundation on Genie Ontology, a self-improving context layer that extracts organizational knowledge across internal and external data to deliver accurate, governed results for business teams.

## TL;DR Databricks product leaders Ken Wong and Elise Georis introduce Genie One, demonstrating how standard AI agents routinely fail at enterprise data tasks due to hallucinations, token-heavy loops, and fragile semantic models. To solve this context bottleneck, Databricks developed Genie Ontology and its OntoRank algorithm, which automatically extracts business logic from existing dashboards, pipelines, and expert activity to boost accuracy and cut execution time in half. Georis delivers a live demonstration building an executive OKR deck, running federated queries across Databricks and BigQuery, updating Jira tickets via Model Context Protocol, and triggering proactive mobile notifications based on live data anomalies. The presentation concludes by announcing Genie One's general availability across Slack, Teams, iOS, and Android, supported by an MCP integration app and a $10 monthly token subsidy per enterprise user.

## Atmosphere & style The presentation carries the polished, confident energy of a major tech launch while maintaining a candid, pragmatic skepticism toward current AI agent hype. Ken Wong establishes an engaging, relatable tone by openly mocking competitor hallucinations and internal engineering naming conventions, which contrasts nicely with Elise Georis's calm, deliberate live demo. The session feels refreshingly grounded and transparent, pulling back the curtain on reasoning traces, governance audits, and citation inspectors rather than relying on rehearsed conversational AI parlor tricks.

## Bulletin - [00:00](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=0s) Databricks introduced Genie One, an agentic AI coworker that connects across corporate systems and live databases to compute answers, automate workflows, and generate artifacts rather than merely summarizing static documents. - [07:10](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=430s) The platform is powered by Genie Ontology, a context layer that harvests knowledge from SQL pipelines, dashboards, and author expertise using the OntoRank algorithm, driving a 30+ percentage point accuracy gain over generic coding agents. - [20:18](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1218s) Genie One is Generally Available immediately across Slack, Microsoft Teams, iOS, Android, and external tools via a Model Context Protocol (MCP) app. - [21:23](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1283s) To drive broad enterprise adoption, Databricks is subsidizing usage by providing $10 worth of free monthly tokens for every user in an organization.

## Key takeaways - [00:42](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=42s) Existing enterprise AI agents and assistants struggle with live corporate data, frequently hallucinating numbers or stalling in runtime loops when attempting complex analytics. - [02:22](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=142s) Databricks internal benchmarks revealed that leading coding agents answer real, nuanced enterprise questions accurately only about 50% of the time. - [05:16](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=316s) Standard agentic loops rely on stochastic walks that force a punishing trade-off between computational cost, token burn, and ultimate accuracy. - [06:12](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=372s) Manual semantic modeling and prompt-engineered skills fail to scale across the long tail of edge cases and bespoke departmental metrics. - [07:10](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=430s) Genie Ontology bypasses manual data curation by automatically extracting logic, queries, and subject-matter relationships directly from operational assets. - [08:02](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=482s) The OntoRank algorithm models PageRank to determine ground truth by evaluating asset reuse frequency, data lineage, and author expertise. - [08:30](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=510s) Genie One operates under strict permission parity, inheriting existing governance rules directly from Unity Catalog without exposing sensitive context. - [10:38](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=638s) In addition to reading context, Genie One uses Model Context Protocol (MCP) actions to execute write tasks, such as posting update requests on open Jira tickets. - [11:43](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=703s) The platform supports federated query execution, retrieving and computing ground truth figures across heterogeneous sources including Databricks and Google BigQuery. - [13:07](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=787s) Every generated response provides a transparent citation drawer displaying underlying foreign catalog tables, verified glossary badges, and ontology snippets. - [15:20](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=920s) Users can package completed conversational context into certified, specialized domain agents that remain persistent and dynamically compute answers against warehouse updates. - [16:51](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1011s) Genie One shifts from reactive querying to proactive alerts by inferring employee interest areas and sending mobile push notifications when statistical anomalies occur. - [19:11](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1151s) Enterprise organizations like General Motors and Warner Music Group have deployed hundreds of production Genie agents across federated data estates. - [20:38](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1238s) A dedicated Genie MCP app allows third-party and custom-built enterprise agents to tap into Genie's underlying calculation and ontology engine. - [21:23](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1283s) Genie One is Generally Available with an allocation of $10 in monthly token credits per employee to remove initial organizational experimentation barriers.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=0s) Introducing Genie One and the Pitfalls of Current AI Agents

Ken Wong, Senior Director of Product Management at Databricks, introduces Genie One as an AI coworker capable of connecting across all enterprise data and applications. Rather than functioning simply as a conversational tool, Genie One allows anyone in an organization to uncover insights, automate workflows, and even generate autonomous agents simply by speaking with it. Anticipating skepticism about launching yet another agent in 2026, Ken addresses the elephant in the room: existing solutions fall flat when dealing with enterprise data. As Ken puts it, “The ones that you have today aren't very good with enterprise data.” They either fail entirely to reason over complex data structures or lack the accuracy required for high-stakes decision-making and operational automation.

To demonstrate the issue, Ken recounts preparing for a Databricks Product Advisory Board (PAB) meeting. He tasked Genie One with assembling a comprehensive customer profile detailing account activity, which required coordinating across fragmented personal and corporate systems: - Connecting to Ken’s personal email, Slack, and calendar to identify PAB attendees and parse recent discussion topics. - Pulling customer use-case details from Salesforce. - Interrogating Databricks consumption data to resolve nuanced questions regarding how these customers utilize Genie.

Genie One successfully returned a detailed, visual customer profile. Intrigued by how competing tools would handle the identical workflow, Ken ran the exact same prompt through five leading alternatives: two software-bundled assistants and three premier coding and coworker agents.

---

[02:22](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=142s) Benchmarking Competitors Against Real Enterprise Tasks

The benchmark against leading market alternatives revealed severe functional flaws across both bundled tools and dedicated coding agents:

1. **First bundled assistant:** Returned an answer suspiciously fast, claiming there were 24 customers in the PAB. When Ken questioned where the figure originated, the tool bluntly confessed that it had completely fabricated the number. Ken reflects on the shock of seeing a tool he frequently uses for drafting documents fail so dramatically: “It confessed that it just completely made it up.” 2. **Second bundled assistant:** Avoided outright hallucination by retrieving existing account brief documents, but the data was a quarter old, completely stale, and useless for current strategic discussions. 3. **Leading coding and coworker agents:** Attempted to query live structured data sources, but quickly became trapped in execution loops. The agent hit its execution time limit, prompted the user for permission to continue burning tokens, and resumed spinning. After 30 minutes of runtime, it produced little more than decorative ASCII art and admitted it was still far from finishing, yielding only a partial report.

Ken explains that this failure is far from an isolated edge case. Databricks Research created a benchmark from real, complex questions asked by employees (e.g., querying what percentage of Genie customers also leverage a third-party agent) rather than simple queries resolvable by basic semantic layers. Evaluated against this benchmark, the best coding agents answered accurately only about 50% of the time, taking minutes to complete. Emphasizing the inadequacy of these results, Ken states, “50% of the time for a real data question is basically unusable, right? You literally can flip a coin.”

---

[05:16](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=316s) The Flaws of the Agentic Loop and Brittle Semantic Layers

The root cause of these agent failures lies in the mechanics of the "agentic loop." Standard agents navigate an exploratory walk, relying on chance to stumble across the correct sequence of reasoning steps and context needed to resolve the question. This design forces a punishing trade-off between runtime, computational cost, and ultimate accuracy.

Historically, organizations have combated this through prompt engineering context, whether packaged as "skills" in AI tools or "semantic layers" in traditional BI. While modeling canonical, stable business concepts is a recognized best practice, Ken emphasizes that manual modeling cannot scale across the long tail of everyday work. Every marketing campaign defines "leads" differently, and individual engineering teams configure Jira fields according to unique internal norms. Manually documenting every variation is impossible. Furthermore, attempting to automate model generation with LLMs fails; citing published research from Anthropic, Ken notes that synthetic context generation tends to over-index on single narrow use cases, ultimately degrading overall model accuracy.

```mermaid flowchart TD accTitle: The Brittle Agentic Loop accDescr: Diagram contrasting how traditional agentic loops lead to high token burn and hallucination versus the scalability limits of manual semantic modeling.

subgraph Loop ["Traditional Agentic Loop"] A["User Question"] --> B["Stochastic Walk / Reasoning Steps"] B --> C{"Hits Context?"} C -- No --> D["Token Burn & Timeout Loop"] D --> B C -- Yes --> E["Coin-Flip Accuracy ~50%"] end

subgraph ContextFail ["Manual Context Limits"] F["Skills / Semantic Modeling"] --> G["Works for Stable Business Concepts"] F --> H["Fails on Long-Tail Edge Cases"] H --> I["Marketing Lead Definitions"] H --> J["Team-Specific Jira Custom Fields"] end ```

---

[07:10](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=430s) Genie Ontology and OntoRank

Databricks addresses context scalability through the Genie Ontology (referred to colloquially on stage as the Genie Anthology). Genie Ontology operates as an automated, learned context layer that sits above existing modeled frameworks in Unity Catalog or external semantic platforms. Rather than demanding exhaustive manual curation, it harvests tacit organizational knowledge directly from operational assets: - **Knowledge Harvesting:** Ingests SQL pipelines, historical queries, dashboards, and reports to detect underlying metric expressions and table relationships. - **Expertise Mapping:** Detects organizational subject-matter expertise by identifying which individuals author and maintain authoritative resources on specific domains. - **OntoRank Algorithm:** Evaluates harvested context snippets using an internal ranking mechanism modeled after PageRank. Ken jokes, “This is how you know the engineers still run the show at Databricks, we call OntoRank.” OntoRank identifies which data transformations and business formulas represent authoritative ground truth. - **Permission Enforcement:** Inherits governance policies and user permissions from the underlying data assets, guaranteeing that sensitive context does not leak across organizational boundaries during retrieval.

```mermaid flowchart LR accTitle: Genie Ontology Architecture accDescr: Flow diagram showing how internal company assets are extracted, ranked by OntoRank, and injected into the agent context loop.

subgraph Extraction ["Knowledge Ingestion"] P["Pipelines & Queries"] D["Dashboards"] E["Subject Expertise"] end

subgraph Processing ["Genie Ontology Layer"] KS[("Internal Knowledge Store")] OR["OntoRank Algorithm"] AC["Access Control & Permissions"] end

subgraph Execution ["Agent Execution"] Q["User Query"] AL["Agent Reasoning Loop"] Ans["Authoritative Grounded Answer"] end

P --> KS D --> KS E --> KS KS --> OR OR --> AC Q --> AL AC -->|"Injected Context"| AL AL --> Ans ```

By injecting this learned context directly into the agent loop, Genie One resolves nuanced questions effortlessly. For example, when asked "how many people are registered," Genie One automatically discovers the precise business logic embedded within production marketing dashboards without requiring anyone to pre-define the metric. Users can inspect the exact source assets used to generate the metric calculation. Ken jokes that the registration numbers in his recorded demo clip were already out of date because marketing had grown the figures significantly over the preceding two weeks. Across internal Databricks benchmarks, adding the Genie Ontology yielded a consistent accuracy increase of more than 30 percentage points over generic coding agents.

[09:45](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=585s) The Foundation of a Data-Smart Coworker

Building on performance benchmarks, the Genie architecture achieves half the runtime of leading agents while maintaining higher precision. Ken Wong emphasizes that as the team tunes the OntoRank algorithm and expands the overall breadth of the ontology, accuracy will be pushed even higher. The goal is moving beyond brittle prompt loops and grounding agents directly in verified organizational knowledge.

With this foundation established, organizations can deploy an AI coworker directly across the enterprise to inform, support, and even automate day-to-day business decisions. Rather than treating AI as a conversational parlor trick, Databricks positions Genie One as an active colleague capable of doing analytical heavy lifting. Ken welcomes Elise Georis, Senior Manager of Product Management for Genie, to the stage to demonstrate Genie One in action.

As Ken summarizes the broader shift:

> “So, with this level of accuracy enabled by the Genie ontology, now we can truly create a data smart co-worker.”

---

[10:38](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=638s) Live Demo: Preparing an Executive OKR Review in Genie One

Elise walks through an everyday workflow: preparing for an upcoming executive review of Genie's quarterly OKRs. The Genie One interface is intentionally straightforward, featuring a central chat area with quick actions to draft documents or create reusable skills, while a left-hand sidebar houses past chats, available agents, and enterprise data assets. To kick off the workflow, Elise pastes in a link to Databricks’ standard review template—which mandates metrics, highlights, and lowlights—and instructs Genie to draft the deck while proactively pinging colleagues who own blocking Jira tickets.

```mermaid flowchart TD accTitle: Genie One Query Execution and Verification Flow accDescr: Diagram illustrating how Genie One processes a prompt by resolving ontology snippets and governance policies before generating live data artifacts.

subgraph Input ["User Request"] REQ["Elise's Prompt & OKR Template"] end

subgraph Governance ["Governance & Context"] UC["Unity Catalog Governance"] GLOS["Glossary & Definitions"] ONTO["Genie Ontology & Snippets"] end

subgraph Action ["Live Execution"] MCP["MCP Actions<br/>(Jira Comments)"] SQL["Federated SQL<br/>(Databricks & BigQuery)"] end

subgraph Output ["Deliverables"] DOC["Populated Review Document"] CIT["Verified Citations & Lineage"] end

REQ --> UC UC --> GLOS UC --> ONTO ONTO --> MCP ONTO --> SQL MCP --> DOC SQL --> DOC DOC --> CIT ```

As the task executes, Elise opens the reasoning trace to inspect Genie's decision-making steps. Genie inspects the Unity Catalog glossary to retrieve relevant semantic definitions and ontology snippets. Crucially, the platform does not demand an isolated permission layer; it strictly inherits the data governance and access control policies already configured in Unity Catalog. From there, Genie pulls relevant Google Docs files, discovers open Jira tickets, uses Model Context Protocol (MCP) to post update requests on those tickets, and issues SQL queries across both Databricks and Google BigQuery to calculate live ground truth numbers.

Elise emphasizes that while conventional generic agents merely parrot static text, real operational workflows require computing numbers that do not yet exist in any document:

> “A lot of generic agents are just reciting what already exists in a document, but there's no document for this week's metrics. That's why I'm doing this.”

- **Speed and Efficiency:** Bypasses costly, token-heavy trial-and-error loops by navigating directly to ground-truth data using the Genie Ontology. - **Automated Document Assembly:** Yields a fully populated document arranged by region, complete with embedded visualizations, anomalies, highlights, and lowlights. - **Permission Parity:** Enforces existing Unity Catalog access controls without requiring custom administrative overhead.

---

[13:07](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=787s) Inspecting Citations, the OntoRank Algorithm, and Adding Skills

To verify the generated findings, Elise inspects the citation drawer, which displays every underlying artifact used: foreign catalog tables, enterprise documentation, tracking tickets, and glossary entries marked with verified badges. Expanding the ontology inspector reveals specific business snippets inferred and evaluated automatically by the platform—such as how the organization defines user engagement, what thresholds constitute a blocker in Jira, and the exact SQL logic required to aggregate engagement metrics across geographic regions.

Drilling into the engagement snippet reveals the strict operational metric: user sessions lasting longer than 30 seconds. The inspector indicates this rule was mined directly from an internal mobile KPIs dashboard created by a colleague named Chung. The snippet carries a high authority score determined by OntoRank, which calculates rank based on historical reuse, asset lineage, and author expertise. Checking Chung’s profile confirms his track record as a trusted authority across several relevant company domains.

```mermaid flowchart LR accTitle: OntoRank Authority Score Flow accDescr: Visual mapping of how individual employee assets contribute to snippet authority scores via OntoRank.

A["Chung's Mobile Dashboard"] --> B["Ontology Snippet:<br/>Engagement > 30s"] C["Asset Usage Frequency"] --> D["OntoRank Algorithm"] B --> D E["Author Domain Expertise"] --> D D --> F["High Authority Snippet<br/>(Trusted for SQL Generation)"] ```

Satisfied with the logic, Elise schedules the entire workflow to run autonomously each week, handling Jira reminders without manual intervention. Next, she enhances the report by instructing Genie to append a predictive forecast. Rather than building models from scratch, she invokes an existing forecasting skill authored by another teammate. The coworker constructed the skill simply by wrapping the built-in forecast function with custom instructions outlining the team's review preferences, allowing Genie to execute the computation over live operational data and present a cleanly rendered forecast visual. Once finalized, the completed document is distributed directly through Slack and email.

- **Engagement Definition:** Standardized as active user sessions exceeding 30 seconds. - **Authority Scoring:** Factored through asset usage frequency, lineage, and author relevance via the OntoRank algorithm. - **No-Code Automation:** Weekly background execution of Jira follow-ups and metrics extraction configured in a single click. - **Skill Reuse:** Composable forecasting skill applied to live warehouse data with zero custom ML infrastructure setup.

---

[15:20](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=920s) Creating Specialized Agents and Proactive Mobile Insights

Because executive metrics reviews spark continuous operational inquiries from stakeholders, Elise packages the context from the conversation into a permanent, domain-specific coworker. Within seconds, Genie auto-generates the configuration details and instructions for a specialized mobile app agent. This agent computes answers dynamically rather than letting answers grow stale in static reporting docs, allowing product managers, HR, IT, and finance teams to stand up persistent, shareable internal experts. Agents can be scoped, refined, and officially certified, which highlights them on the enterprise catalog directory and prioritizes their context inside the Genie Ontology.

```mermaid flowchart TD accTitle: Proactive Mobile Insight Pipeline accDescr: Process showing how background activity triggers proactive mobile push notifications without explicit manual queries.

subgraph Context ["Continuous Learning"] ACT["User Activity & Topics"] --> PREF["Inferred Interests"] end

subgraph Engine ["Genie Engine"] LIVE["Live Operational Data"] --> MON["Telemetry Monitor"] PREF --> MON MON --> DET["Engagement Spike Detected"] end

subgraph Delivery ["Mobile Notification"] DET --> PUSH["Proactive Push Notification"] PUSH --> APP["Genie Mobile App Chat<br/>(Drill Down & Follow-Ups)"] end ```

Transitioning from the desktop environment, Elise demonstrates how the experience translates to mobile. As she pulls up the Genie mobile app, a push alert appears notifying her of an unexpected spike in app engagement. The notification required no preconfigured threshold or monitoring dashboard; Genie inferred her interest in mobile adoption based on her past work and surface telemetry proactively.

Reflecting on the difference between passive bots and genuine autonomy, Elise highlights the paradigm shift:

> “Genie's no longer just a reactive Q&A tool. It's now a proactive AI co-worker that can bring me things that matter before I even know to ask.”

- **Self-Maintaining Intelligence:** Certified agents provide persistent domain knowledge that remains fresh against production databases. - **Proactive Interventions:** Telemetry tracking automatically alerts stakeholders to statistical anomalies without requiring manual alert rules. - **Omnichannel Access:** Workflows shift seamlessly from desktop browser interfaces to native mobile experiences.

---

[18:36](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1116s) Enterprise Ecosystem: Connectors, Governance, and Customer Impact

Ken Wong returns to summarize the core pillars that position Genie One as a data-smart enterprise coworker. Genie One pairs live analytics generation with a broad connector footprint spanning both federated queries and data ingestion pipelines. Because it natively leverages custom Model Context Protocol (MCP) tooling and the Unity AI Gateway, organizations can integrate internal systems and execute automated background schedules without needing to keep user laptops awake.

Ken emphasizes that governance and security remain paramount when deploying autonomous capabilities. All tooling, models, data connections, and access boundaries are monitored through the Unity AI Gateway, which will be covered in greater detail during a subsequent governance deep dive.

Ken illustrates the real-world impact across global enterprise customers:

> “And what this means is that customers like Warner Music Group is able to use Genie for both Databricks and non-Databricks data, and and customers like General Motors is able to deploy hundreds of Genie agents into production to improve the performance in every single one of their teams.”

- **Warner Music Group:** Leverages Genie across federated architectures encompassing Databricks and non-Databricks data stores. - **General Motors:** Deploys hundreds of domain-tailored Genie agents into production environments to accelerate team workflows. - **Foot Locker:** Equips business teams with autonomous access to live data so frontline staff can make data-driven operational decisions.

[20:18](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1218s) Ubiquitous Access Across Slack, Teams, Mobile, and Custom MCP Integrations

Databricks is expanding Genie's footprint so that every employee can harness enterprise data directly within their daily flow of work. Ken Wong announces that Genie is now natively integrated into both Microsoft Teams and Slack, launching immediately. Alongside desktop enterprise chat platforms, Genie is also rolling out as a mobile experience across both major mobile operating systems, iOS and Android. Wong notes that conversational chat has already become the primary way Databricks' internal staff interfaces with Genie, anticipating identical adoption patterns across enterprise customers: “So, as Ali mentioned, this has already become the most popular way for Databricks employees to experience Genie, and I'm pretty sure it's going to be that way for your for your custom for your employees as well.”

For companies that have already deployed other generative AI frameworks or engineered proprietary in-house agents, Databricks is introducing the Genie MCP (Model Context Protocol) app. Rather than forcing organizations into an all-or-nothing platform migration, the Genie MCP app allows existing third-party agents and workflows to query Genie directly as an underlying computational engine. Wong explains that this integration provides external agents with the data grounding needed to run accurate calculations over enterprise data: “And that gives your your existing agents or existing tools the ability to tap into Genie's ability to compute results accurately.”

```mermaid flowchart TD accTitle: Genie One Access Surfaces and Integrations accDescr: Diagram showing client chat apps, mobile devices, and external MCP agents connecting into Genie One and its underlying ontology.

subgraph Surfaces["Workplace Surfaces"] Slack["Slack Integration"] Teams["Microsoft Teams Integration"] Mobile["Mobile Apps (iOS & Android)"] end

subgraph External["External AI Stack"] Agents["Existing / Custom AI Agents"] MCPApp["Genie MCP App"] end

subgraph Platform["Databricks Genie Engine"] GenieOne["Genie One Agentic Worker"] Ontology["Genie Ontology & Compute Layer"] end

Slack --> GenieOne Teams --> GenieOne Mobile --> GenieOne Agents --> MCPApp --> GenieOne GenieOne --> Ontology ```

- **Collaborative Chat Integration**: Direct availability within Microsoft Teams and Slack channels. - **Native Mobile Support**: Immediate deployment on both Google Android and Apple iOS devices. - **Genie MCP App**: Standards-based connectivity enabling external and custom-built agents to leverage Genie's analytical accuracy.

---

[21:23](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=1283s) General Availability and Monthly Token Credits for Every Employee

Wong underscores the core architectural distinction that separates Genie One from conventional enterprise assistants. While standard AI agents are limited to indexing and summarizing text from pre-existing documents, Genie One utilizes the Genie Ontology to execute computations and synthesize trusted, data-backed deliverables: “With the ontology else the only AI agent that allows everyone in your organization to actually consistently make results and not just summarize summarize information from existing documents.”

To encourage widespread enterprise adoption and remove initial experimentation friction, Databricks is subsidizing usage by providing $10 worth of free tokens every month for every single user in an organization. Wong officially announces that Genie One is Generally Available (GA) immediately, ending his address with an open invitation for organizations to connect their entire tech stack and distribute the AI coworker to all team members across the business.

- **General Availability**: Genie One enters GA status across Databricks workspaces. - **Monthly Token Subsidy**: $10 worth of monthly tokens allocated per individual user across the organization. - **Core Value Proposition**: Moves beyond static document summarization to verifiable computational execution against live enterprise data.

## Notable quotes > [10:38](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=638s) A lot of generic agents are just reciting what already exists in a document, but there's no document for this week's metrics. — Elise Georis

> [15:20](https://www.youtube.com/watch?v=6W3BnCgoYVA&t=920s) Genie's no longer just a reactive Q&A tool. It's now a proactive AI co-worker that can bring me things that matter before I even know to ask. — Elise Georis

## Mentioned Genie One, Genie Ontology, OntoRank, Databricks, Unity Catalog, Unity AI Gateway, Model Context Protocol (MCP), Ken Wong, Elise Georis, Salesforce, Slack, Microsoft Teams, Jira, Google BigQuery, Google Docs, Anthropic, Warner Music Group, General Motors, Foot Locker

## Video description Introducing Genie One, the data-smart AI worker.

Genie One goes beyond just answering questions. The newest member of the Genie suite of products, Genie One is an actual agentic co-worker that can produce documents, reports, and artifacts.

It's powered by Genie Ontology, a new self-improving context layer that extracts information from across all a company's data — inside and outside of Databricks — for trusted answers and actions. Genie One helps marketing, finance, sales, and other teams automate and orchestrate work.

Learn more: https://www.databricks.com/blog/introducing-genie-one-genie-ontology-and-genie-agents

00:00 — Introducing Genie One 00:42— Why AI agents struggle today 05:16 — How the "agentic loop" impacts accuracy, cost, and runtime 07:10 — Introducing Genie Ontology 09:31 — Benchmarking Genie One against other leading AI cowork agents 10:21 — Demo of Genie One and Genie Ontology with Elise Georis, Senior Manager, Product Management, Databricks 17:32 — The shift to proactive AI coworker 19:08 — Genie One's custom MCP tools, skills, and connectors 19:51 — How Warner Music Group, General Motors and Foot Locker use Genie 20:24 — New Genie One availablity on Slack, Teams and mobile devices

Presenters: Ken Wong, Senior Director, Product Management, Databricks Elise Georis, Senior Manager, Product Management, Databricks

[Source](https://www.youtube.com/watch?v=6W3BnCgoYVA)

Reading summary

Databricks introduced Genie One, an enterprise AI coworker designed to compute live analytical answers and automate cross-platform workflows using an automated context layer called the Genie Ontology. Product leaders Ken Wong and Elise Georis argue that the tool overcomes the high error rates and execution loops of conventional AI agents by harvesting verified business logic directly from operational assets. Failure of Current AI Agents on Enterprise Data (2:22) Existing AI assistants struggle when handling complex enterprise data because standard agentic loops rely on exploratory, stochastic reasoning walks. In a live evaluation assembling an executive customer profile across email, Slack, Salesforce, and Databricks telemetry, Databricks found that bundled software assistants either outright fabricated attendee counts or returned quarter-old static documents. Premier coding agents timed out after 30 minutes of token burn without completing the task. Across Databricks Research benchmarks testing complex employee queries, leading coding agents answered accurately only about 50% of the time. Genie Ontology and the OntoRank Algorithm (7:10) While organizations traditionally attempt to resolve AI failures through manually curated semantic layers or LLM-generated context, Databricks argues manual modeling cannot keep pace with team-level definitions, while synthetic context degrades accuracy. To address this, the Genie Ontology automatically mines SQL pipelines, dashboards, reports, and organizational subject-matter expertise from Unity Catalog and external repositories. An internal ranking algorithm modeled after PageRank, dubbed OntoRank, evaluates the historical lineage, reuse, and author expertise of operational assets to determine authoritative formulas. Databricks claims this injected context increases query accuracy by over 30 percentage points over generic coding agents while cutting runtimes in half. Live Operational Workflows and Autonomous Actions (10:38) During a live demonstration preparing an executive quarterly OKR review, Genie One drafted a review deck by executing federated SQL queries across Databricks and Google BigQuery while inheriting access controls directly from Unity Catalog. Using Model Context Protocol (MCP), the agent identified blocking Jira issues and posted comments requesting updates from assignees. The platform provides full citation lineage down to individual dashboard rules—such as standardizing user engagement as sessions exceeding 30 seconds based on an internal mobile KPI dashboard. Users can also attach reusable analytical skills, such as peer-authored forecasting models, and schedule workflows to run autonomously without keeping local machines active. Proactive Alerts, Specialized Agents, and Integrations (15:20) Genie One allows users to convert ad-hoc analytical workflows into certified, domain-specific agents for ongoing stakeholder queries. The platform also monitors live operational telemetry against inferred user interests, delivering proactive mobile push notifications when metric anomalies occur rather than functioning as a purely reactive Q&A interface. Enterprise customers cited by Databricks include General Motors, which has deployed hundreds of production agents, and Warner Music Group, which runs queries across hybrid Databricks and non-Databricks stores. Genie One is accessible through Slack, Microsoft Teams, native iOS and Android apps, and external workflows via an MCP connector. Genie One is now generally available, with Databricks offering a $10 monthly token subsidy per user to accelerate enterprise deployment. While Databricks claims the platform resolves the trade-off between computational cost and accuracy, success hinges on the continuous tuning of the OntoRank algorithm and the platform's ability to interpret diverse, messy enterprise environments without manual semantic intervention.

Source quotes

[10:38] Elise Georis: A lot of generic agents are just reciting what already exists in a document, but there's no document for this week's metrics. [15:20] Elise Georis: Genie's no longer just a reactive Q&A tool. It's now a proactive AI co-worker that can bring me things that matter before I even know to ask.

Original transcript

[0:00] Welcome to the stage, Databricks Senior [0:03] Director of Product Management, Ken [0:06] Wong. [0:13] >> All right. Good morning. [0:16] So, today I get to introduce you to [0:18] Genie 1. Genie 1 is an AI co-worker that [0:23] can connect to all of your data and all [0:25] of your apps so that everyone in your [0:27] organization can get insights and even [0:30] automate action. You can even create [0:32] autonomous agents just by talking to [0:34] Genie. [0:36] But, I know what you're all thinking. [0:38] In 2026, [0:40] why do we need another one of these AI [0:42] agents? [0:44] Well, the answer is actually very [0:45] simple. The ones that you have today [0:47] aren't very good with enterprise data. [0:50] They either can't connect and reason [0:52] about data at all, or they're just not [0:54] accurate enough for you to actually [0:57] trust them to make real decisions or let [0:59] alone automating your work. [1:02] So, let me show you what I mean. So, [1:04] yesterday we had our product advisory [1:06] board meeting at Databricks. Um, shout [1:08] out to the PAB. And I wanted to [1:11] prepare myself for it by getting Genie 1 [1:14] to help me with a comprehensive profile [1:16] of all the customers we have, what's [1:18] going on in their account. So, this [1:20] required Genie 1 to connect to all the [1:22] systems that we have including my my [1:24] personal email and Slack [1:27] and my calendar to figure out who's in [1:29] the PAB and what we've been talking to [1:30] them about. And that required Genie 1 to [1:32] connect to Salesforce to pull in the use [1:34] case information. But, it also meant [1:37] that Genie 1 needed to connect to [1:40] Databricks in order to pull all the [1:42] consumption data that they had because I [1:44] had a lot of nuanced questions that I [1:46] wanted to include in the brief about [1:47] what they're using Genie for. [1:49] And this is what Genie 1 was able to [1:51] come up with. Um, as you can see it gave [1:53] me a comprehensive profile. I had to [1:55] blur some of it so you don't see the [1:56] details. Um, but it after you know, [1:59] querying Databricks and querying these [2:01] systems, they able to give me a [2:02] comprehensive and visual profile of all [2:05] of our customers. So, what I wanted to [2:06] do was try this exact same prompt with [2:10] five of the leading AI agents that were [2:12] out there. I tried two of the AI [2:14] assistants that came bundled with some [2:15] of the software we had and three of the [2:17] leading coding and co-work agents. And [2:20] this is what I saw. [2:22] So, the first AI assistant is actually [2:23] one that I use a lot when I used to [2:25] write docs and things like that. And it [2:29] it came back with a response to my [2:30] prompt very, very quickly, actually a [2:32] little bit too quickly. And there was [2:34] something a little bit off about it. And [2:36] the first thing I noticed was that it [2:37] said that we had 24 customers in the [2:40] pub. And I knew this was not true. So, I [2:42] asked it and said, "Hey, like where did [2:44] you get this 24 number from?" [2:47] And [2:48] guess what it said? [2:49] It confessed that it just completely [2:51] made it up. I was stunned, right? Cuz [2:54] this is actually I use this thing a lot. [2:56] Uh, not just for at work, but also have [2:58] a personal subscription. Um, but it just [3:01] shows to goes to show you how important [3:03] in enterprise grounding actually is. [3:06] Now, the other AI assistant that I used [3:07] actually did a lot better. It actually [3:09] searched existing documents and pulled [3:11] out account briefs. Um, but they were [3:13] all sort of stale, you know, they were [3:15] all from a quarter ago, information that [3:18] was no longer relevant to the [3:20] conversation I wanted to have. [3:22] So, the results of this thing completely [3:24] unusable. The good news it was very, [3:25] very quick. [3:27] Now, the the coding agents and the [3:29] co-work agents actually made a genuine [3:31] attempt and connected to all the all the [3:34] systems and the structured source in [3:36] order to pull live data out. But if [3:38] you've ever used one of these things, [3:40] you've probably experienced what I [3:41] experienced, which was that it went off [3:43] and tried very, very hard to pull this [3:45] information out. It actually hit its [3:47] time limit um, where it said [3:49] asked for permission to continue burning [3:51] lord knows how many tokens in the [3:52] process. And then after I allowed it to [3:55] to continue, it went [3:58] away and it [4:00] iterated again and again, and it came [4:01] back with some cool ASCII art. Um but [4:04] then [4:05] confessed again that actually we were [4:07] far from done, you know, we had only [4:09] gone partway through. So, we're 30 [4:11] minutes into this thing at a partial [4:13] report. It's like a huge contrast with [4:15] what Genie 1 was able to do. [4:17] So, you might say, "Hey, this is sort of [4:19] a one-off task. Um you know, you [4:21] cherry-picked an example." But you know, [4:23] it's actually very very consistent with [4:26] the research that Databricks Research [4:27] had been doing on the ability of coding [4:30] agents to answer novel questions. [4:32] Actually, what we saw, we created a [4:34] benchmark out of real questions that our [4:36] employees were asking of Genie, and we [4:38] asked coding agents to solve the same [4:40] problems. Now, these weren't [4:41] run-of-the-mill simple questions that [4:43] could be easily, you know, um solved [4:46] with some semantic layer. These were [4:48] like nuanced questions like, "Hey, with [4:49] all the all the customers who use Genie, [4:52] what percentage also use a third-party [4:53] agent?" Questions that, you know, hadn't [4:55] been asked before. And we saw that [4:58] coding agents were able to do it [5:00] the best of them about half the time, [5:02] and it took minutes. And remember this [5:04] chart, we'll come back to it. But 50% of [5:07] the time for a for a real data question [5:10] is basically unusable, right? You cannot [5:12] make you literally can flip a coin. [5:14] Um [5:15] and if you know how these agents work, [5:17] and I already explained this a little [5:19] bit already, you're [5:21] it shouldn't be surprising because they [5:22] basically go through this process of the [5:25] agentic loop where really getting to the [5:27] right answer is a matter of it, you [5:29] know, kind of making sure happening upon [5:32] like the right sequence of reasoning [5:34] steps to chance upon like the right [5:37] context it needs to ask answer the [5:40] question. So, it kind of goes through [5:41] this walk where it it just does its [5:43] best. And there is a very real trade-off [5:45] between the accuracy that you get and [5:48] the cost and run time of these things. [5:50] And And this trade off results in what [5:52] we see. [5:53] Now, our current best approach to [5:55] solving this problem is to inject some [5:58] context, right? Like to give it some [5:59] information that gives it hints on the [6:01] right thing. And if you come from the AI [6:03] world, you might call these things [6:04] skills. And if you come from the BI [6:06] world, you might call these things [6:07] semantic layers, but they're really the [6:09] same idea. It's like, "Hey, we can just [6:10] tell it the things that it need to do [6:12] and then it can do a better job." And [6:14] the problem with this is that, well, [6:16] while it's totally reasonable and in [6:17] fact best practices to do it for the [6:20] most stable business concepts you have, [6:22] it's unreasonable to do it for the [6:24] breadth of things that you do within [6:26] your organization. Think about how each [6:28] and every single marketing campaign [6:30] defines leads a little bit differently, [6:32] or how every single scrum team uses all [6:34] the fields in Jira a little bit [6:36] differently. You just can't [6:38] realistically write all this stuff down. [6:41] And before you say, "Hey, can't we just [6:42] use AI to solve this problem by [6:45] generating these models?" Well, it [6:46] actually just doesn't work. Actually, [6:48] Anthropic [6:49] published a great blog on the limits of [6:51] this, talking about that the fact that [6:53] if you do this, what you end up doing is [6:56] just encoding or like anchoring the AI [6:58] on one specific use case, and then that [7:01] actually [7:02] negatively result affects the results [7:04] overall. [7:06] So, what is our solution to this [7:08] problem? Well, I mean, Ali stole my [7:10] thunder a little bit. It is the Genie [7:12] Anthology, right? The Genie Anthology is [7:15] an automatic context layer that adds a [7:18] learned layer on top of any modeled [7:21] context that you have, whether that's in [7:22] Unity or in one of your existing [7:25] semantic model or uh modeling tools. [7:28] Um it And it basically allows Genie and [7:31] any agents attached to the anthology to [7:33] do a much better job of accurately [7:36] answering questions with data. Let me [7:38] tell you how it works. [7:40] So, first, with the Genie Anthology, [7:41] what we do is connect to the systems [7:44] that you have that contain knowledge. [7:47] So, for example, using the Databricks [7:49] system, you have a ton of information in [7:51] your pipelines, queries, dashboards, and [7:53] things like that that tells an agent how [7:56] they can reason about data. And we [7:57] extract all of those things, things like [8:00] expressions and relationships, but also [8:02] expertise, like the idea that this [8:04] person is really knowledgeable about a [8:07] specific topic. And we extract all of [8:09] this into our internal knowledge store. [8:12] And then we go over this knowledge with [8:15] our an algorithm which, you know, this [8:17] is how you know the engineers still run [8:18] the show at Databricks, we call [8:20] OntoRank. And it's a PageRank out like [8:23] algorithm that helps us determine which [8:26] snippets are actually authoritative. [8:28] And then with with this information, [8:31] when a when a question comes into Genie, [8:33] we're able to look up this information, [8:35] apply permissions because we also [8:38] extracted out the permissions of the [8:39] underlying source access, so that we [8:41] don't have to you don't have to worry [8:43] about leaking permissions, and then we [8:45] just inject that context into the agent [8:48] loop so that Genie is able to answer [8:50] much better [8:52] questions much more accurately. [8:53] And the results of this is so that if if [8:55] I ask a question like how many people [8:57] are registered, um without ever having [9:00] to model it, Genie knows about the [9:01] nuances of how um [9:04] how we captured registrations in our [9:07] internal systems because it was able to [9:10] extract that information out of some [9:11] source asset. And if you want to know [9:13] where it came from, you can just inspect [9:14] it and see how it actually learned about [9:16] this these calculations from our, you [9:19] know, underlying dashboards. Uh [9:20] marketing team's a little bit mad at me [9:22] about this because I took this video [9:24] from 2 weeks ago, um but the numbers are [9:26] a lot higher now. [9:28] So, [9:30] coming back to this chart. So, we saw [9:32] what happened uh [9:33] with generic coding agents, how does [9:35] Genie with the ontology do? Well, in our [9:38] internal testing, what we've seen is a [9:40] consistent 30 percentage point plus [9:42] improvement in in accuracy and roughly [9:45] half the the run time of the leading [9:48] agents, right? And we're really [9:50] confident that as we continue to tune [9:53] the onto rank algorithm and expand the [9:56] the the the breadth of the ontology that [9:58] we're able to push this accuracy much [10:00] much higher. [10:03] So, with this level of accuracy enabled [10:05] by the Genie ontology, now we can truly [10:08] create a data smart co-worker. One that [10:11] you can give to all of your employees to [10:13] actually make decisions with data and [10:14] even automate decision-making, right? [10:17] It's the it's the foundation. [10:19] So, [10:20] I'd like to welcome now Elise Joris onto [10:23] the stage to show you what you can do [10:25] now that you have this foundation and [10:26] what you can do with Genie 1. [10:33] >> [music] [10:38] >> Hello everyone. [10:40] My name is Elise. I'm a product manager [10:42] here at Databricks for Genie and I use [10:45] Genie in a lot of my day-to-day work. [10:48] We actually have an exact review for our [10:50] Genie OKRs that's coming up and I'm [10:52] going to use Genie to prepare for it. [10:55] I'm starting here in Genie 1 and as you [10:58] can see, it's a pretty simple interface. [11:01] I have my chat here in the middle. We [11:03] have some [11:05] buttons that we can use to draft a doc [11:07] or create a skill. [11:09] On the left side, I have my prior chats [11:12] as well as all the agents and the assets [11:14] that are available to me. But, more on [11:17] that in just a second. [11:19] I'm going to start by asking Genie to [11:21] create a document for my review. Every [11:23] team at Databricks uses the same [11:25] template. It's got metrics, highlights, [11:28] lowlights. So, I've given Genie a link [11:30] to that. [11:31] I've also asked to poke the people who [11:32] have open Jira tickets so that we can [11:34] keep those moving. [11:36] Now, as Genie is running, we can [11:37] actually click into what it's thinking. [11:40] So, as you can see here, it clearly [11:42] found something interesting in the Unity [11:44] catalog glossary. [11:45] Seems like it found some potentially [11:47] relevant ontology snippets. And this is [11:50] actually all from data that I do already [11:52] have access to. Genie isn't forcing you [11:55] to create a separate permission system. [11:57] It's enforcing the permissions that you [11:59] already have through Unity catalog. [12:02] Looks like it also grabbed some Google [12:04] files and found those Jira tickets. [12:07] It used MCP to comment on those tickets [12:10] just like I asked. And now it's actually [12:12] running some SQL across both Databricks [12:15] and BigQuery to get the latest ground [12:17] truth. [12:18] And that's actually really important. A [12:20] lot of generic agents are just reciting [12:23] what already exists in a document, but [12:25] there's no document for this week's [12:27] metrics. That's why I'm doing this. [12:29] Genie understands the numbers and is [12:32] actually able to compute the results [12:33] based on our live operational data. And [12:36] to that end, it looks like I already [12:38] have a doc for my review. [12:40] Looks like Genie pulled the metrics and [12:43] broke them down by region. [12:45] Gave me a bunch of nice visualizations. [12:48] Grabbed the highlights, the lowlights, [12:50] anomalies. And notice that it actually [12:52] did that fairly quickly. Rather than [12:55] just grinding through an agentic loop of [12:57] trial and error and eating up a lot of [13:00] tokens along the way, Genie just went [13:02] straight to the ground truth. And that [13:04] is really the Genie ontology at work. [13:07] So, how do we know that we can actually [13:09] trust this? [13:10] Well, for that we're going to pop down [13:11] here into the citations. This is [13:13] everything that Genie actually used to [13:15] answer my question. And again, we have [13:16] tables, foreign catalogs, tickets, docs. [13:20] Looks like it did actually use the [13:22] definition from the glossary. You can [13:24] see here that it's verified. [13:26] But I actually want to click into the [13:27] ontology. I want to make sure that Genie [13:29] inferred everything correctly. [13:32] So, what you're seeing here are ontology [13:34] snippets. These are again the facts [13:36] about our business that Genie has [13:38] learned and evaluated automatically. [13:42] It looks like there's a snippet for how [13:43] we define engagement, a snippet for what [13:46] qualifies as a blocker in Jira, and one [13:50] for the sequel that we can use to pull [13:52] engagement by region. [13:54] I'm going to click into this first [13:55] snippet and see Here we see the full [13:58] definition. We're looking at sessions [14:00] that lasted longer than 30 seconds. [14:03] And it looks like Genie pulled this from [14:05] our mobile KPIs dashboard, which was [14:07] authored by our colleague Chung. [14:10] It's also got a high authority score. [14:12] This is a function of things like how [14:14] often the snippet is used or the asset [14:16] that it was pulled from. It goes back to [14:18] that OntoRank algorithm that Ken was [14:20] talking about. [14:21] And I can actually click into Chung's [14:23] profile and here I see that he's engaged [14:26] with a lot of the domains our team cares [14:27] about. And yeah, it actually looks like [14:29] Genie has pulled quite a few snippets [14:31] from the assets that he's created. [14:34] Now, while I'm here, I might as well [14:36] schedule this to run every week, so I [14:38] don't have to manage Jira myself. [14:41] Great. And now I'm just going to take [14:43] this one step further. I'm going to ask [14:45] Genie to add a forecast to my doc. One [14:48] of my teammates already has a [14:49] forecasting skill, so I'm just going to [14:51] use that. And she didn't do anything [14:53] fancy. She just used the built-in [14:54] forecast function and added a few [14:56] instructions for how we like to review [14:58] these forecasts. [15:00] And again, this is really only possible [15:02] because Genie's able to compute these [15:04] results based on our live operational [15:07] data. [15:08] And it looks like it added my forecast. [15:10] It also gave me a pretty nice [15:11] visualization here. [15:13] So, now I'm going to ship this. I'm [15:15] going to send it over email and Slack. [15:19] Now, this OKR review isn't a one-time [15:21] thing. People are asking me for updates [15:23] on the mobile app literally all the [15:25] time. So, I'm going to create an agent [15:27] that can answer questions using context [15:30] from this conversation. [15:34] I'm going to name my agent. [15:41] Great. Now, I have an agent. This is an [15:44] expert coworker that can answer [15:46] questions and take action related to the [15:49] Genie mobile app. [15:51] And unlike an OKR doc, which is probably [15:54] going to get stale pretty quickly, this [15:56] agent stays up to date and again, it can [15:59] compute results on the fly. So, [16:01] while I'm here, might as well share this [16:03] with the other PMs at Databricks. [16:05] And now I'm going to click into this [16:08] agent. [16:09] So, as you can see, I'm actually [16:11] chatting directly with my agent. I'm not [16:13] going to have to specify what product [16:15] we're talking about because it's already [16:17] baked into the agent. It's customized to [16:19] the mobile app. And you can also see [16:21] that in these agent details that Genie [16:23] has automatically generated. [16:26] And more broadly, you can probably [16:27] imagine all the different teams that [16:29] might use something like this. Think [16:31] about any function that has to answer a [16:33] bunch of questions, HR, IT, finance. [16:37] Now, everyone in the organization can [16:39] create and collaborate on these types of [16:42] expert coworkers in one place. [16:45] Now, from here, I can go in and manage [16:47] my agent, you know, tighten the scope, [16:49] sharpen the instructions, add some [16:51] context, or I can certify it. And this [16:54] is going to appear on our organization's [16:56] agent list page as a certified agent, [16:59] and the Genie ontology is also going to [17:01] prioritize it. [17:03] Now, I don't do all of my work on the [17:06] desktop. I actually have to do a lot of [17:09] work on the go, which is why I have the [17:12] Genie mobile app. [17:14] And if I open my app, [17:17] on my phone that is [17:19] it looks like I just got a notification. [17:21] Uh engagement with the app is spiking. [17:24] It's all very meta when you have to work [17:26] on Genie with Genie. And I can click [17:28] into the app to learn more or maybe ask [17:31] a follow-up question. [17:33] And to be clear, I didn't actually have [17:35] to set up this insight. Genie knows that [17:38] I care about engagement with the app [17:40] because of my activity and it can [17:42] actually proactively notify me because [17:44] something interesting happened. And I [17:47] think that's actually really powerful. [17:49] Genie's no longer just a reactive Q&A [17:51] tool. It's now a proactive AI co-worker [17:55] that can bring me things that matter [17:56] before I even know to ask. So in just a [17:59] couple of minutes we moved the project [18:01] along, prepped for an exact review, got [18:03] some colleagues questions answered, and [18:06] even learned something that we didn't [18:07] know to ask about. That is a super [18:10] high-quality AI co-worker and it's [18:12] really only possible because Genie [18:14] understands our business through the [18:16] ontology, it can compute results rather [18:18] than just reciting documents, and it can [18:21] enforce our organization's governance [18:23] along the way. So we'd encourage you to [18:26] try this on your own data and share your [18:28] feedback. And with that Ken is going to [18:30] come back out. Thank you all. [18:35] >> [music] [18:36] >> All right, what do you think? [18:39] So [18:40] Genie 1 is a data smart AI co-worker. It [18:44] has full core capabilities. You can [18:47] create documents and again compute them [18:49] live, not just summarize and recite [18:51] existing things that you have, but net [18:53] new analysis. You can connect it them it [18:56] to all of your systems thanks to the [18:57] connectors that we have. You can [18:59] schedule it to work when you're back in [19:01] the background and you can schedule it [19:03] without having to keep your laptop open. [19:06] Um and you can integrate it to all of [19:07] your systems even custom ones because we [19:09] have custom MCP tools all managed to uh [19:13] Unity AI Gateway as well. And you can [19:15] develop custom skills that you can share [19:17] and create with your co-workers, even [19:20] create a dedicated domain specific agent [19:23] just by talking to Genie, so you can [19:25] improve the performance of that focused [19:27] agent over time. [19:30] And Genie, again, works on all of your [19:32] data, thanks to the large and growing [19:34] ecosystem of connectors that we have [19:36] through federation and ingestion. And [19:39] customer And it's all managed by Unity. [19:41] So And we'll have a deep dive on all the [19:44] governance features of the Unity AI [19:45] Gateway tomorrow. And what this means is [19:48] that customers like Warner Music Group [19:50] is able to use Genie for both Databricks [19:53] and non-Databricks [19:55] data, and and customers like General [19:58] Motors is able to deploy hundreds of [20:00] Genie agents into production to improve [20:02] the performance in every single one of [20:04] their teams. And it allows customers [20:07] like Foot Locker to be able to think [20:09] about how to transform their business to [20:11] a world where every employee is armed [20:14] with the ability to make decisions with [20:16] data. [20:18] We really want every single user to be [20:20] able to leverage data to do their jobs, [20:23] which is why we're striving to make [20:25] Genie available everywhere. So, I'm [20:27] pleased to announce that Genie is [20:28] integrated now into Teams and Slack, [20:31] available today. [20:33] And it'll be available on your mobile [20:35] devices on both both major platforms on [20:39] Android iOS again today. So, as Ali [20:42] mentioned, this has already become the [20:44] most popular way for Databricks [20:45] employees to experience Genie, and I'm [20:47] pretty sure it's going to be that way [20:49] for your for your custom for your [20:51] employees as well. [20:53] Now, for those of you who are using one [20:55] of these existing AI agents or have [20:57] developed your own and would like to [20:59] benefit from the super [21:01] superior accuracy and data smarts of [21:03] Genie, we're also pleased to announce [21:06] the availability of the Genie MCP app. [21:09] And that gives your your existing agents [21:12] or existing tools the ability to tap [21:14] into Genie's ability to compute results [21:17] accurately. [21:18] And all and that is also available [21:21] today. [21:23] So we really believe that Genie 1 is [21:25] unique. With the ontology else the only [21:27] AI agent that allows [21:29] everyone in your organization to [21:30] actually consistently make results and [21:33] not just summarize summarize information [21:35] from existing documents. And we want to [21:37] make sure it's available to to [21:39] absolutely everybody. So to in order to [21:41] to kind of encourage that to happen, [21:43] we're giving $10 worth of tokens [21:46] for every single user inside your [21:48] organization every single month. So [21:51] there's really no reason not to get [21:53] started with with Genie Genie 1. And [21:56] we're going to make Genie 1 generally [21:59] available today. [22:04] >> [applause] [22:06] >> Right. So please connect Genie 1 to [22:08] every everything inside your [22:09] organization and give it to everybody. [22:15] >> [music]