Databricks Genie wants to tell the CEO what nobody else will
Past 150 people, the boss stops hearing bad news. That is Ali Ghodsi's pitch for the new Genie, which Databricks launched Wednesday (https://www.databricks.com/blog/new-genie-autonomous-ai-complex-enterprise-work). Ghodsi cited Dunbar's number to TechCrunch and noted that at 12,000 employees no salesperson is going to confess a disastrous call to the CEO. So the CEO asks the agent instead (https://techcrunch.com/2026/11/11/databricks-new-ai-agent-tells-ceo-ali-ghodsi-the-unvarnished-truth/).
The demo was not a demo. Ghodsi opened a real Genie history for the reporter: an analysis of where Databricks scored badly in its employee survey, a ranking (and rejection) of fixes executives had proposed, a post-mortem of a sales pitch that went poorly including the features the prospect wanted, and prep for a podcast built from Ghodsi's past appearances. Ghodsi also says the same agent covers personal tasks, instead of what Ghodsi called security-questionable options like OpenClaw, Instinct or Muse.
The product shift is from chat to jobs. Databricks describes Genie as an autonomous coworker that runs long jobs continuously: watching insurance claims for fraud patterns and routing case files to underwriters, running marketing campaigns and proposing weekly budget changes for approval, firing on triggers like a metric dip or more than 10 customers asking for the same feature. It works through Genie One, email, Slack and Teams. Context comes from Genie Ontology, a map of business definitions and metrics; control comes from Unity Catalog permissions and Unity Gateway, which sets usage limits and can swap in cheaper open-weight models. Databricks also published a benchmark with Turing on a fictional UK utility: Genie at 91% pass rate against 45% and 58% for two unnamed frontier agents, using under half the tokens (https://www.databricks.com/blog/genie-genie-ontology-delivers-16-accuracy-less-half-tokens). Vendor benchmark, vendor-built company, rivals unnamed; treat it as a claim.
The interesting bet is not the agent, it is who owns the context. Every lab wants to be the CEO's agent; Databricks is betting the agent that wins is the one sitting on the warehouse, the metric definitions and the permissions. Whether a CEO really wants an agent grading their own executives' ideas is a culture question no ontology answers.
← Back to all articles
The demo was not a demo. Ghodsi opened a real Genie history for the reporter: an analysis of where Databricks scored badly in its employee survey, a ranking (and rejection) of fixes executives had proposed, a post-mortem of a sales pitch that went poorly including the features the prospect wanted, and prep for a podcast built from Ghodsi's past appearances. Ghodsi also says the same agent covers personal tasks, instead of what Ghodsi called security-questionable options like OpenClaw, Instinct or Muse.
The product shift is from chat to jobs. Databricks describes Genie as an autonomous coworker that runs long jobs continuously: watching insurance claims for fraud patterns and routing case files to underwriters, running marketing campaigns and proposing weekly budget changes for approval, firing on triggers like a metric dip or more than 10 customers asking for the same feature. It works through Genie One, email, Slack and Teams. Context comes from Genie Ontology, a map of business definitions and metrics; control comes from Unity Catalog permissions and Unity Gateway, which sets usage limits and can swap in cheaper open-weight models. Databricks also published a benchmark with Turing on a fictional UK utility: Genie at 91% pass rate against 45% and 58% for two unnamed frontier agents, using under half the tokens (https://www.databricks.com/blog/genie-genie-ontology-delivers-16-accuracy-less-half-tokens). Vendor benchmark, vendor-built company, rivals unnamed; treat it as a claim.
The interesting bet is not the agent, it is who owns the context. Every lab wants to be the CEO's agent; Databricks is betting the agent that wins is the one sitting on the warehouse, the metric definitions and the permissions. Whether a CEO really wants an agent grading their own executives' ideas is a culture question no ontology answers.
Comments