Databricks has put an updated version of its Genie agent on a shared layer of company knowledge, called an ontology, which learns what a business's terms and numbers mean from the work its teams already do. Its chief executive, Ali Ghodsi, showed a reporter how he uses Genie on the company's own records: it picked apart the low scores in an employee survey, rated the fixes his executives had suggested, and summarized a sales pitch that failed, including the features the buyer said it needed.
Ghodsi's reasoning is that a person can keep real relationships with about 150 others. Beyond that, the people below a boss stop volunteering anything that makes them look bad. A salesperson does not confess a disastrous call to a CEO with 12,000 employees, and an HR manager does not announce that their own idea will probably fail.
The truth is already written down
A lost pitch leaves a call recording and notes. A survey leaves a table of scores and a pile of comments. A weak plan leaves a proposal. All of it exists, but it reaches the top softened, because everyone who handles it has a reason to soften it. An agent has no career to protect, so it reads the same material and reports what is in it.
The ontology is what makes that reliable. Without it, an agent has to guess what "a qualified lead" or "late delivery" means in your company. Databricks' layer learns those meanings from the dashboards, queries and documents teams already use, and people only have to fix the few definitions that must be exact. That is the same problem we described when we wrote that AI answers are trusted only when someone defines revenue, solved by letting the agent learn most of it.
The limits are plain. It reads only what the company has recorded and connected, and only what its user is already allowed to see. A grievance nobody wrote down stays invisible.
For example, an operations director at a distributor with 400 staff wants to know why on-time delivery fell. Each site manager has an explanation that points at someone else. An agent that reads the shift logs, the customer complaints and the dispatch dashboards can show that most late orders were picked on Monday mornings, and can test each manager's explanation against that. The director still has to decide what to do and walk the floor, but the question is already answered when they arrive.
Staff will change what they write
Reading the records changes the records. Once people know the boss's agent reads call notes and survey comments, they will write them with the agent in mind. Notes get shorter and safer, comments get vaguer, and the second way of hearing the truth slowly turns into the first.
Aviation has been through this. Planes already record their flight data, and airlines can analyze it routinely to spot risks before an accident. That program, known as FOQA, is voluntary, and it took about 10 years of talks between the US regulator, airlines and pilot unions to agree the legal protections before the rules came out in April 2004. Under them, what reaches the regulator is summarized with names removed. The pilots had a say in how the data could be used before the reading became routine.
A manager who points an agent at a team's records can borrow that idea in an afternoon. Say beforehand what the agent is for and what it is not for: patterns across the team, yes; a list of who wrote what, no. If staff believe that, the records stay honest, and the agent keeps hearing the truth.