Enterprise AI Agent: Databricks Genie Goes Always-On

Enterprise AI Agent: Databricks Genie Goes Always-On

The Ontology Is the Point: Databricks Makes Its Enterprise AI Agent Always-On

Databricks has announced the second generation of Genie, which it now describes as an always-on, autonomous enterprise AI agent. Given a goal, Genie takes ownership of the work, delegates to sub-agents, tracks progress, follows up on missing inputs and reaches out when a person needs to decide. It runs in the background on its own compute, opens to a priority inbox rather than a chat history, and sends each alert as its own email thread so an inbox does not fill with updates. The architectural claim underneath all of that is Genie Ontology, which the company describes as an agentic ontology maintaining an understanding of how a business operates across its data, people and processes. Genie and Genie Ontology are generally available. Genie Search, Ontology Manager and the new always-on Genie are coming to preview and beta programmes.

What ships today and what does not

The availability paragraph deserves stating before anything else, because the headline describes a capability most readers will assume they can buy.

Genie and Genie Ontology are generally available to all customers. The always-on autonomous mode, Genie Search and Ontology Manager are not: the release places them in preview and beta programmes with no date given. That is a reasonable way to ship a large platform change and it is worth being precise about, because the difference between a generally available product and a preview is the difference between something a team can plan on and something it can experiment with.

The same paragraph names the second generation as the announcement, so the substance of what Databricks is describing is a direction of travel with a working implementation behind it rather than a finished product. That distinction does not diminish the engineering. It does change what an enterprise should do this quarter.

Why the ontology is the real architecture

An enterprise AI agent that can answer a question needs a model and a retrieval path. One that can act on a business needs to know what the words mean, and that is the problem Genie Ontology addresses.

The chief executive’s quotation is the clearest explanation in the release. Ali Ghodsi describes Genie knowing what the company means by revenue, which numbers are official, and who can see what. Those three things are not data and they are not code. They are the shared agreements a business runs on, and they usually live in people’s heads, in a definitions document nobody reads, or in a BI tool’s semantic layer that covers only the analytical half of the estate.

An ontology that captures definitions, metrics, the people behind them and the processes they belong to gives an agent something to reason with. Without it, an agent asked to keep a quarterly forecast on track has to guess which revenue figure counts, and a guess in an autonomous system is a liability rather than a mistake.

This is what Databricks means by horizontal context, and the phrase is worth keeping. Most enterprise software delivers vertical context: a sales tool knows about accounts, a finance tool knows about ledgers, and neither knows how the two connect. An enterprise AI agent that operates across a business needs the connective layer, and building it is a data governance problem before it is a model problem.

Three ways to connect, and why the choice matters

Genie Search extends across more than 100 enterprise data sources, including the lakehouse and business applications such as Google Drive, Jira and Confluence. The mechanism is the more interesting detail: Databricks offers three connection strategies, being ingest into the lakehouse through Lakeflow Connect, index in place, or federate queries live through the Model Context Protocol.

Offering all three is a practical decision with real consequences. Ingesting copies the data, which is fastest to query and worst for freshness and for governance surface, because a copy is a second thing to secure. Indexing in place keeps the source authoritative and leaves the data where its access controls already sit. Federating through MCP queries the application live, which means the answer is current and nothing moves at all.

A customer can now choose per source rather than accepting one architecture for everything. A Confluence space might be indexed in place; a Jira project might be queried live because its state changes hourly; a large analytical table might be ingested because it is queried constantly. That flexibility is what makes a hundred connectors useful rather than a hundred liabilities.

Governance is the unlock, not the compliance page

Every action Genie takes stays within Unity Catalog permissions and Unity Gateway approval rules and budgets, and the release says each is fully auditable.

That is the sentence that decides whether an autonomous enterprise AI agent is deployable in a regulated company. An agent that can act is an agent that can act wrongly, and the controls that make wrongness survivable are permissions, approval thresholds, spend limits and a record. Running inside the customer’s existing Databricks environment means those controls are the ones the company already operates rather than a new set to configure and audit separately.

The sub-agent design interacts with this in a way worth noting. Genie delegates to multiple sub-agents to work in parallel, which is how long-running work becomes feasible. It also multiplies the number of actors taking actions, and a governance model that covers one agent has to cover all of them. Databricks says the controls apply by default rather than by configuration, which is the right answer to that problem.

The customer evidence and the one hard number

Three customers are quoted, and the third is the most useful because it contains a figure. Alan Khalil of HelloFresh says Genie’s analysis of performance signals improved programmatic ad return on investment and reduced the time to identify ad winners on Meta by 75 per cent, and that Unity Gateway gives the team flexibility to use the best model for each job without managing separate integrations.

A 75 per cent reduction in time to identify winning creative is a specific operational claim about a specific platform, and it is the kind of number a media buyer can check against their own experience. Marissa Wong of Headspace makes a different point, that building on agreed governed definitions means every answer matches, and that plain-language access put data in the hands of teams beyond the analyst group. Jaime DeLanghe of Slack frames the collaboration case, that teams go from question to action without leaving the conversation.

The chief executive’s own claim, that Genie has made him a ten-times more productive chief executive, is a self-assessment from the person best placed to be enthusiastic about it, and a reader should take it as such.

Genie everywhere, including inside competitors

One detail deserves separating from the rest. The release says teams can interact with Genie wherever they work, naming a web browser, mobile app, email, and collaboration and AI tools including ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot.

Putting an enterprise agent inside other vendors’ assistant interfaces is a notable strategic choice. It treats the assistant layer as interchangeable and the governed context layer as the asset, which is consistent with everything else in the release: the ontology, the search and the permissions are Databricks’ contribution, and the surface where a person meets them matters less than what sits behind it.

The open ecosystem point follows the same logic. Genie MCP lets other AI tools and agents use Genie Ontology and Genie Search to find enterprise knowledge, which means Databricks is positioning the context layer as infrastructure rather than as an application.

What would complete the picture

Four things would let an enterprise plan against this rather than admire it.

How the ontology is built and maintained is the first, and it is the largest question in the release. Databricks says the ontology keeps itself current as the business changes and that Ontology Manager flags frequently asked concepts needing refinement, which implies a combination of automation and human curation. The proportion between the two determines whether this is a product an organisation can run or a consulting engagement wearing a product’s clothes.

Preview timelines are the second. Three of the headline capabilities are in preview or beta with no date, and a date is what converts interest into a plan.

Cost and compute are the third. Genie runs in the background on its own compute, which is exactly the kind of architecture whose value depends on what the meter reads at the end of the month.

And the governance detail is the fourth. Unity Gateway approval rules and budgets are named without describing how a budget is set, what happens when an agent hits one, and who is notified. Those are the operational questions a finance team will ask first.

Enterprise AI Agent: Databricks Genie Goes Always-On
Photo by Ferenc Almasi on Unsplash

What to watch

Whether the always-on mode reaches general availability on a published timeline, because that is the milestone that turns a direction into a product. Whether a customer describes how much of their ontology they had to build by hand, since that answer determines the real cost of adoption. And, whether the ontology becomes something other vendors can read, because a context layer that only Databricks can consume is a moat and one that others can query is infrastructure. And whether the assistant layer does become interchangeable, because if it does, the value in enterprise AI moves decisively to whoever holds the definitions.