Pular para o conteúdo

The New Databricks Genie: What Changes When Your Data Agent Starts Doing the Work

Cover – Site | | Cheesecake Labs
Summary
  • Databricks' new generation of Genie moves agentic analytics from answering questions to running jobs, with long-running work, triggers, approvals and actions in tools like Slack, Jira and Microsoft 365; the always-on autonomous mode is still coming to preview and beta, while Genie and Genie Ontology are generally available.
  • Genie Ontology infers business definitions from existing dashboards, queries and notebooks, shifting semantic-layer work from authoring to curating: domain owners still need to review what was learned and mark authoritative definitions, which decide answer quality.
  • Once an agent can act, a wrong answer becomes a wrong action, so identity, approval points, budgets and audit trails should be designed before the first job runs, with autonomy promoted in stages starting with one recurring job.
  • Genie is billed on usage rather than per seat, and SQL warehouse compute is billed separately; the promotional period can be used to measure real usage and baseline cost on one workload before scaling.

In a previous article, Agentic Analytics in Production: What Snowflake CoWork Replaces, and What It Doesn’t, we argued that Snowflake, Databricks and Microsoft Fabric are converging on the same three-part architecture: a semantic layer that defines business truth, agents that run under the user’s own permissions, and a chat interface that replaces the ticket queue.

Databricks has since pushed that pattern one step further. The new Genie doesn’t only answer a question and wait for the next one. You give it a job, the tools it needs and the rules it must follow, and it keeps working until the job is done.

What is the new Databricks Genie?

Genie is Databricks’ family of AI coworkers. The pieces that matter for this discussion:

  • Genie One is the front end for business users. The first generation of Genie only provided conversational analytics over data stored in Databricks; Genie One connects to data inside Databricks and in connected applications, and produces documents, reports and other artifacts. It runs on web, desktop and native iOS and Android apps, and inside Slack and Microsoft Teams.
  • Genie Agents are agents a team configures with its own data, files, tools and instructions, then shares with access controls.
  • Genie Ontology is the context layer. Agents build it from dashboards, queries and notebooks, keep learning from usage, and use it to ground answers in business definitions.
  • Unity Catalog and Unity Gateway are the governance layer: data permissions, tool and connector controls, approvals, budgets and audit.

The latest generation adds autonomy on top: Genie tracks progress on a job, follows up on missing inputs, asks for approval when required, and can run on a schedule or in response to a trigger, such as a metric dropping. Several jobs can run at once and be supervised, steered or stopped from a single conversation. Through MCP connectors it can read from and act in Slack, Zoom, Microsoft 365, Jira, GitHub and calendars, and teams can publish repeatable workflows as skills that are stored in Unity Catalog and run with each user’s permissions.

Databricks’ own examples show the shift well: monitoring insurance claims for fraud and loss-ratio drift and routing case files to underwriters, or running a marketing campaign, investigating underperformance and proposing weekly budget changes for approval. Those are jobs, not queries.

What’s available now, and what’s coming?

Databricks ships Genie in stages, so it is worth separating what you can put in production from what is still on the roadmap. At the time of writing:

Component Status
Genie (incl. Genie One and Genie Agents) Generally available
Genie Ontology Generally available
Always-on Genie (autonomous, long-running jobs) Coming soon to preview and beta
Genie Search Coming soon to preview and beta
Ontology Manager Coming soon to preview and beta

The headline capability, the always-on agent, is still on its way to general availability. What you can do now is prepare the two things it will depend on: a curated ontology and a governance model. Teams that do that work today will be the ones able to turn autonomy on safely when it ships.

Does Genie Ontology remove the semantic-layer work?

No. It changes its shape.

In our Snowflake work, enabling the tooling took minutes and building an accurate semantic layer took months. The semantic layer, not the model, determined whether answers were right. Snowflake’s approach is an authored one: people write semantic views that define entities, metrics and rules.

Genie Ontology takes the opposite starting point. Agents infer definitions from the dashboards, queries and notebooks your teams already use, and keep learning from usage. That is a real accelerator for the first draft of business context, especially in organizations where nobody has the time to write a semantic layer from scratch.

The catch is that an ontology learned from existing work also learns the existing inconsistencies. If finance and sales calculate “active customer” differently in their dashboards, the agent sees both. Databricks addresses this directly: teams set authoritative definitions in Unity Catalog, and the ranking algorithm (OntoRank) gives those definitions precedence over learned ones. Gap alerts flag concepts people ask about often that still lack a definition.

So the job moves from writing every definition to deciding which definitions win. That is less typing and the same amount of judgment. Someone who owns each domain still has to review what the agent learned and mark what is authoritative.

Databricks reports that, in a benchmark built with Turing, Genie with Genie Ontology completed 91% of task runs correctly, 1.6 times the rate of the stronger of two frontier agents, using less than half the tokens. It is a vendor-run benchmark, so we read it as a direction rather than a number to plan around. What it supports is the same lesson we learned on Snowflake: context quality is what moves accuracy.

Read more: Your AI Strategy Has a Data Problem

What changes when the agent can act?

When an agent only answers questions, a wrong answer costs a bad decision that a person still has to make. When an agent opens tickets, posts summaries and proposes budget changes, a wrong answer becomes a wrong action. Governance stops being a compliance checkbox and becomes the core of the design.

The controls Databricks describes are the right ones:

  • Identity. Genie acts as the user and only reaches data that user is authorized to see, enforced by Unity Catalog and by each source system’s permissions. It can draft an access request, but admins and data owners approve it.
  • Actions. Admins decide which connectors and MCP tools are available and which actions require approval. Users choose how autonomous to be within those limits.
  • Audit. Every action is traced. Agent execution traces with tool calls are available for review.
  • Spend. Admins can set workspace and per-user budgets, caps and alerts through Unity Gateway.

Two caveats from the documentation are worth knowing before you rely on budgets. They cover LLM usage, not SQL warehouse compute, which is billed separately. And enforcement is approximate: requests in flight are not interrupted, and Databricks says not to use budgets to guarantee an absolute cap on the final bill.

Industry analysts commenting on the launch pointed in the same direction. Sanjeev Mohan of SanjMo argued that enterprises need to govern what agents do, not just what they read: a distinct identity per agent, limits on actions, human approval for consequential steps and a full audit trail. Mike Leone of Moor Insights & Strategy said he would like Genie to flag when it isn’t sure of an answer, because a wrong answer surfaced proactively can spread widely.

Our own experience supports the second point. On Snowflake, the agents that earned trust were narrow ones with explicit rules about data, aggregation and output. Broad “ask anything” agents lost users after a single wrong answer. Autonomy raises those stakes. A proactive agent that is wrong once in a shared Slack channel teaches a whole team not to trust it.

Read more: Trust at Scale: Why Governance and Adoption Are the Hardest Parts of AI

What does the new Genie cost?

Genie is billed on usage, not per seat. According to Databricks’ cost documentation:

  • Usage is metered in Databricks Units (DBUs). Each user (not service principals) gets 150 DBUs free per month, reset on the first of the month.
  • Through January 31, 2027, user usage of Genie One and Genie Agents is free, and billed Genie usage gets a 25% discount. Service-principal usage is billed throughout.
  • Billed usage falls under a Serverless Real-Time Inference SKU, and can be tracked in the billing system tables by filtering on Genie as the origin product.
  • The always-on mode has no published pricing yet.

In our Snowflake piece, the lesson on cost was that the credit price is not the token cost, and that compute adds up on top. The equivalent here: free agent usage does not mean free queries. Agents still run SQL against your warehouses, and that compute is billed as usual.

The free window is also an opportunity. Every question and job your users run during the promotional period generates real usage data you can use to estimate the cost of the same workload at list price, before you commit to anything.

Snowflake or Databricks: which one should you use?

Both are strong choices, and they start from different places. Databricks is extending Genie beyond analytics into an agent layer that searches, acts and builds artifacts across the workplace, grounded in an ontology that learns from how teams already use data. Snowflake CoWork builds on semantic views that people author, which puts the emphasis on governed analytics business users can trust from the first question. Learned context gets you to a first version faster; authored context gives you tighter control over every definition. Most organizations end up needing some of both.

For most companies, the platform matters less than it looks, for two reasons.

First, the architectures keep converging. Semantic or context layer, agents under user permissions, a conversational surface: the skills a team builds on one platform transfer to the other.

Second, the agent should follow the data, not the other way around. If your governed data already lives in Databricks, Genie is the natural place to start; if it lives in Snowflake, CoWork is. Moving data to get a different agent is rarely worth the cost and risk. On either platform, the work that decides the outcome is the same: clear definitions, scoped agents and access designed on day one.

How do you roll out the new Genie without losing control?

This is our playbook for agentic analytics, extended for a platform where agents can act.

  1. Start with one recurring job, not one question. Pick work someone does every week by hand: a variance report, a pipeline review, a capacity check. Recurring work with changing parameters is where agents pay off first, and a job gives you a clear before-and-after.
  2. Curate the ontology for that domain before you add autonomy. Review what Genie learned, mark the authoritative definitions for the metrics that job depends on, and resolve conflicting ones.
  3. Write the job like a job description. Scope, data, rules, output format, and, now, which steps need human sign-off. Publish it as a skill so it runs the same way for everyone, under each user’s permissions.
  4. Design access, approvals and budgets on day one. Decide which connectors and actions are on, set per-user budget thresholds (Databricks recommends blocking on those rather than on the shared pool), and decide who reviews the traces.
  5. Promote autonomy in stages. Read-only first, then drafts a person approves, then actions with approval, and only then actions without it, for the steps that have earned that trust.
  6. Measure work removed, not chat volume. Count the tickets that never reached the data team and the hours of recurring analysis the job replaced. Use the free period to baseline cost at the same time.

FAQ

What is the difference between Genie One and the new Genie? Genie One is Databricks’ AI coworker for business users. The newer generation of Genie extends it with autonomous, long-running jobs, along with Genie Search, Ontology Manager and governance features in Unity Gateway. These capabilities are rolling out in stages.

Do I still need a semantic layer with Genie Ontology? You need the decisions a semantic layer captures. Genie Ontology builds a first version from how your teams already use data, but someone must still mark which definitions are authoritative. Those definitions take precedence and drive answer quality.

Can Genie access data a user isn’t allowed to see? According to Databricks, no. Genie acts with the user’s identity, and Unity Catalog and source-system permissions apply. It can draft an access request, but data owners approve it.

How much does Genie cost? It is usage-based, metered in DBUs, with 150 free DBUs per user per month. Warehouse compute used by agents is billed separately.

Can Genie work with tools outside Databricks? Yes. Genie connects to tools such as Slack, Microsoft 365, Jira and GitHub through MCP connectors, and Genie MCP brings its business context into other AI assistants, including ChatGPT, Claude and Microsoft Copilot.

Ready to put an agent to work on your data?

We help teams map the business context their agents need, scope a first job and design the access, approval and cost controls around it, on Databricks or Snowflake. We’ll be candid about what we’d build now, what we’d wait for, and what we wouldn’t build at all. Talk to us.