Databricks' AI Agent Is Learning What 'Revenue' Means. Which of Yours Wins?
Agents that learn from usage pick the popular number, not the official one
Most organizations assume an AI agent understands the business once it can see the data. Connect it to the warehouse, set the permissions, and it will work out what the numbers mean.
The larger problem is where that understanding comes from. On November 11, Databricks announced an always-on version of Genie, an agent that runs in the background on its own compute, manages long-running work and brings decisions to people, now heading into preview. Its context comes from Genie Ontology, which learns your definitions and metrics from dashboards, queries, notebooks and BI tools, and keeps them current.
That puts a quiet job on your data team's desk. If the agent learns from what your people already query, someone has to decide which of those queries reflect the official definition and which are one team's workaround from three years ago.
And it gets worse as you scale. One analyst using the wrong revenue filter affects one report. An always-on agent that learned the same filter repeats it in every weekly brief, forecast check and executive answer, without anyone asking.
As agents start working on their own in the background, every agent needs to know what it is working with, who owns it and which other agents it can trust. That is what we are building at OpenMap: The Agent Network, the connective tissue for the world's AI agents, starting with every person and every local business. We are raising a £5M seed. Meet us at openmap.ai/investors.
Why Does the Agent Learn the Popular Definition, Not the Right One?
Picture three revenue numbers. Finance reports bookings net of refunds. Sales tracks signed contract value. A regional team built a dashboard years ago that counts renewals twice, and it is still the most viewed report in the company.
An agent that learns from usage sees all three. Nothing in the queries tells it that popularity is not authority.
The weekly brief looks confident.
The number in it is wrong.
This is not a model accuracy problem. It is a definition ownership problem.
The Cost Isn't One Wrong Answer. It's Wrong Answers on a Schedule.
Databricks is clear about what the agent needs. In the launch announcement, CEO Ali Ghodsi says Genie knows "what we mean by revenue, which numbers are official, and who can see what." A Headspace data leader quoted in the same release puts it more simply: "AI is only as good as the definitions behind it."
That is the right standard. The question is who in your organization sets it. Catalog permissions decide what the agent is allowed to see. They do not decide which of two visible definitions is correct.
Before an always-on agent can be trusted with a recurring brief, someone usually has to:
Decide which version of each core metric is official, and retire the rest
Trace where every field in that metric comes from, and which reports reuse it
Name an owner who signs off when a definition changes
Resolve duplicate customer, product and account records that inflate the counts
Mark which dashboards are legacy, so nothing learns from them by accident
None of these tasks are particularly difficult. They are slow, they cross team lines, and the answers usually sit with a handful of senior analysts.
Databricks' own Ontology Manager flags concepts that "need a governed definition." Flagging is the easy part. Someone still has to own the answer.
How to Give Your AI Agents Official Definitions in 30 Days?
You do not need to redesign your data platform, or pause your agent pilots, to give agents definitions they can rely on.
Within the first month, focus on the metrics your executives ask about most: revenue, active customers, churn, margin and pipeline.
Start by mapping where each metric is calculated today and which dashboards, models and queries depend on each version. Then pick the official one, name its owner, and mark the others as deprecated.
Next, expose this context through the catalog or knowledge layer your analysts and agents already read, so whatever an agent learns from, it meets the certified version first.
DataManagement.AI accelerates this process by automatically discovering metadata, mapping lineage across the modern data stack, identifying downstream dependencies, and creating a centralized knowledge layer that stays continuously updated as your environment evolves. It resolves duplicate and conflicting records, and ownership and policy travel with every entity, so the official definition carries its owner wherever an agent finds it.
Instead of spending months documenting assets manually, or hoping the agent picks the right dashboard, your data owners decide once and every agent inherits the decision. When a definition changes, the change travels with it.
Agent Context Is Actually a Data Ownership Problem
Agent platforms will keep getting better at learning context on their own. Databricks says more than 20,000 organizations use its platform, so learned ontologies are about to become normal inside the enterprise.
But an agent can only learn what your data already says, and in most companies it says several things at once.
Organizations that scale agents successfully do not let them guess which number is official. They decide, and they make the decision readable.
Warm regards,
Shen and Team