Databricks’ new AI agent tells CEO Ali Ghodsi the unvarnished truth

Databricks on Wednesday introduced a revised version of Genie, its AI agent, which the company says can now take on multistep jobs that require extended work to complete and finish them without supervision. The capability rests on an ontology Databricks built to give the agent the context it needs from the information customers keep on the Databricks platform. An ontology, in this context, is a structured model of the entities, relationships and definitions in a domain; supplying one to an agent is a way to ground its reasoning in a company's own vocabulary and data rather than leaving it to interpret raw tables and documents unaided.
Rather than stage a product demonstration, CEO Ali Ghodsi showed a reporter his own live agent, walking through recent chats and charts. Among the examples: the agent analyzed the portions of an employee satisfaction survey where the company scored poorly, and it helped Ghodsi rate proposed fixes for those low scores suggested by executives, including discarding some of them. It also summarized a large sales pitch to a prospect that went badly, capturing the features the prospect said it needed before buying. It prepared him for a podcast interview by reviewing his earlier appearances and drafting likely questions. During the conversation it was even asked to advise him on what to say to the reporter.
The notable thread is that Ghodsi is using Genie to surface information he would be unlikely to hear directly as chief executive. At 12,000 employees, he acknowledged he cannot know everyone, and staff are unlikely to volunteer bad news to him. He invoked Dunbar's number, the theoretical ceiling of about 150 people with whom a person can maintain stable relationships, named for British anthropologist Robin Dunbar. In a small startup a CEO has direct relationships across the company and people feel safe delivering negative information straight. At Databricks' current size, a salesperson is unlikely to confess a disastrous call to the CEO, and an HR manager is unlikely to admit that an idea for lifting employee satisfaction will probably fail. Ghodsi said anyone with more than 150 employees, which he noted covers a lot of companies, could get value from agents like this.
Databricks is now positioning Genie as an always-on agent for business leaders. Ghodsi described using it in his personal life as well, in preference to alternatives he characterized as security-questionable, naming OpenClaw, Instinct and Muse. His stated view is that Genie can do for the business side what AI, agents and harnesses are doing for programmers; a harness is the software wrapped around a model that lets it use tools. On the cost risk that agents could drive up corporate AI spending, he pointed to Databricks' Unity Gateway product, which lets companies set usage limits and swap in less expensive open-weight models.
This suggests the pitch is less about a single query answering a question and more about continuous monitoring of the operational exhaust a large company already produces — survey results, call outcomes, meeting and pitch records — with an agent turning it into decisions. That is a meaningfully different product posture from a chat assistant that waits to be asked. The trade-offs are worth naming as my own reading: unsupervised multistep execution raises questions about how a leader verifies an agent's conclusions before acting on them, and grounding through an ontology only works as far as that ontology captures the business correctly. Model routing through a gateway helps on price but can mean answers vary in quality depending on which model handles a given step.
Why it matters: teams building internal analytics and agent products will see Databricks arguing that the differentiator is curated context plus long-running autonomy, not the underlying model. For organizations past the size where leaders hear bad news firsthand, an inference on my part is that demand for agents that summarize uncomfortable signals will grow quickly, and so will scrutiny of their accuracy. Platform teams, meanwhile, should expect cost controls and model swapping to be judged as core agent features rather than add-ons.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
With Databricks at 12,000 employees, AI Ghodsi is the first to admit that he can't know everyone and that they are unlikely to confess the things that go wrong to him. But he doesn't need them to. He uses Databricks' newly revised Genie agent for that.