Databricks updated Genie so its CEO can read bad news directly

Databricks has published a revised version of Genie, its AI agent that runs multi-step jobs across a company's data. Chief executive Ali Ghodsi told TechCrunch on Tuesday that he uses Genie himself. He said he uses it to hear things his direct reports will not tell him. The piece was reported by Marina Temkin at TechCrunch.

What the new Genie does

The updated Genie runs on top of Databricks' Unity ontology. The ontology is a map of the company's own data and the words people use for it. With that map in place, Genie can take an open question and run several steps to answer it. It reads the data it needs. It writes small pieces of code. It calls other tools. A human does not have to name each step.

Ghodsi gave TechCrunch four live examples. He asked Genie to read an employee satisfaction survey where Databricks had scored low. Genie laid out what the survey said. Ghodsi then had Genie read and reject fixes proposed by other executives. He asked it to write a short on a sales pitch that had lost. Genie listed the features the prospect had needed. He asked it to prepare him for a podcast by reading his past interviews and listing the likely questions.

Dunbar's number as the pitch

Ghodsi framed the use case through organisational psychology. He asked TechCrunch: "Have you heard of Dunbar's number? The magic number is 150." His argument: with 12,000 employees at Databricks, a seller or a manager who finds a problem rarely tells the CEO. Genie reads the raw data and says the problem out loud.

Ghodsi told TechCrunch that anyone with more than 150 employees can get value from the agent. He added that this covers a lot of companies and a lot of people. That is his claim about the product he sells. It is not a measured result.

The cost side, and what it means for a buyer

Databricks also pointed TechCrunch at a second product called Unity Gateway. The gateway sets usage limits on Genie and can swap in a cheaper open-source model when the question allows. That is how Databricks answers the standard worry about an agent that runs many steps on big data.

For a team building software, the Genie piece is a model for one pattern. Plug an agent into a company's own ontology. Let it read what people answered to a survey. The executive layer stops being the only channel. The honest question is what that layer stops being asked to do. A head of sales would answer that reading the lost-deal note is their job. A seller who is already honest will see no change. A team rolling this out should ask who gets quieter when the CEO hears the number first.

Source

• TechCrunch