Databricks Ships Genie Ontology and Says Its Agent Beat Unnamed Rivals, but Autonomy Must Wait
Databricks made Genie Ontology generally available and says its Genie agent beat two unnamed frontier rivals 91 percent to 58, but the always-on version that works unsupervised is still only headed to preview.
Ali Ghodsi runs a company of roughly 12,000 people, and he has a problem every big-company boss has. Nobody tells him the bad news anymore. So he asks a chatbot.
That was the pitch Databricks made on Wednesday, November 11, when it made Genie Ontology generally available and showed off an updated Genie agent built to take on longer, multistep work. To be clear about which Genie this is, this is the enterprise data agent from Databricks, not Google DeepMind's Genie 4 world model, which we covered when it started building playable worlds. Same name, completely different product.
In a briefing with TechCrunch, Ghodsi skipped the canned demo and opened his own agent. He showed it picking apart the weakest scores in an employee satisfaction survey, rating and discarding fixes some executives had proposed, summarizing a big sales pitch that went badly, and prepping him for a podcast by reviewing his past appearances. Databricks told the outlet that corporate leaders can now hand the agent a wide range of tasks and it will finish them unsupervised. That is the company's claim, backed by a CEO showing his own chats, not an audit.
"Anyone who has more than 150 employees, which is a lot of companies, a lot of people can get value out of this." Ali Ghodsi, CEO of Databricks, speaking to TechCrunch
The 150 is Dunbar's number, the rough limit on stable human relationships. Past it, Ghodsi argues, a salesperson will not confess a disastrous call to the CEO. An agent reading the data will.
What makes that work, according to Databricks, is the ontology. The company describes Genie Ontology as a map of how a business operates that agents build automatically from the dashboards, SQL queries, notebooks and Power BI models teams already use, while people govern the definitions that must be exact, such as recognized revenue. Databricks says it respects existing user permissions and that Unity Gateway controls connections, usage limits, guardrails and auditing. Ghodsi told TechCrunch the same gateway is the answer to runaway AI bills, since companies can set caps and swap in cheaper open-weight models.
A 91 percent score with fine print
The headline number comes from Databricks itself. In a benchmark it built with Turing, the company says Genie with Genie Ontology passed 91 percent of enterprise task runs, 1.6 times the rate of the strongest frontier agent, while using less than half the tokens. Its launch post labels the two rivals only as Agent A and Agent B, scoring 45 and 58 percent. Databricks does not say which models or products they are.
The setup matters too. The test simulated a Databricks workspace for a fictional UK energy supplier, with each agent answering the same questions five times per task. Databricks itself says the results reflect an established deployment where teams have already accumulated that knowledge. In plain terms, Genie was playing on home turf with a map, and the rivals had to explore from scratch every run. That may well be the point of the product, but it is not proof every customer job will land at 91 percent.
Then there is the line between what ships today and what does not. Databricks' press release says Genie and Genie Ontology are generally available to all customers. The headline act, the new always-on Genie that runs in the background on its own compute, delegates to subagents and emails you when it needs a decision, is only coming soon to preview and beta programs. The release puts Genie Search and the Ontology Manager curation tool in that same bucket, even though the Ontology blog post calls Genie Search generally available, a mismatch Databricks has not explained.
"Genie has made me a 10x more productive CEO, and now it doesn't wait for me to ask." Ali Ghodsi, in the Databricks press release
Nobody measured that 10x. And the version that "doesn't wait" is the one customers cannot fully switch on yet. Databricks has a genuinely interesting idea here, that agents fail at work because they do not know what a company means by its own words. Whether its map holds up outside a benchmark it designed is the question its customers will answer.
