Databricks Genie: What Its CEO’s Agent Reveals

A customer stops replying after a sales call. An employee flags a recurring process problem. The explanation exists somewhere in your notes, tickets, or survey responses, but nobody has time to connect it to a decision. Databricks CEO Ali Ghodsi uses an AI agent to surface those uncomfortable signals; the useful question for a small business is how to do that without giving an agent more access or authority than it needs.
What Ghodsi actually showed
Ghodsi’s demonstration showed an agent helping him investigate problems, not independently fixing them. TechCrunch reported that he used Genie to examine low-scoring areas of an employee satisfaction survey and assess proposed fixes. He also showed a summary of an unsuccessful prospect pitch, including features the prospect wanted before buying.
Those are two different jobs. The survey example asks, What are people telling us is wrong, and do our proposed changes address it? The sales example asks, Why didn’t this prospect buy, and what should we take from that? Neither answer is useful if the agent merely produces a fluent summary. It must preserve the evidence, distinguish a person’s stated reason from its own interpretation, and put the result in front of someone who can act.
Databricks had 12,000 employees when TechCrunch covered the demonstration. A CEO at that scale cannot personally hear every concern or review every deal. But the underlying information problem is familiar to a much smaller team: the founder sees a neat status report while the useful detail sits in a call note, a support thread, or a comment field.
Databricks announced its second-generation, always-on Genie on November 11, 2026. The company says it can delegate work to subagents, track progress, and ask a person for a decision. Its described priority inbox and email alerts are important because an insight that remains inside a chat session rarely changes operations.
There is an availability distinction here. Genie and Genie Ontology were generally available at the time of the announcement, but Databricks described the new always-on mode as coming to preview or beta. Ghodsi’s demonstration is a useful design example, not proof that every customer can turn on the same workflow today. His description of the agent as making him a “10x more productive CEO” is his characterization, not an independently measured result.
Start with one decision, not all your data
A small team should begin with a recurring decision it already struggles to make. “Tell me what’s going wrong in my business” is too broad to test. “Show me stalled prospects whose recorded objections have not been assigned to an owner” is narrow enough to build, check, and improve.
Pick a decision that has all three parts:
- A signal: a recorded event or statement, such as a lost-deal reason, repeated support complaint, or delayed invoice response.
- An owner: someone responsible for deciding what happens next.
- An observable outcome: a follow-up sent, a product request reviewed, a process changed, or a reason documented for taking no action.
Consider an illustrative sales workflow. A founder wants to know whether prospects are asking for a capability the team does not offer. The relevant inputs might be call notes, deal records, and a maintained description of current capabilities. The output is not “a list of things customers want.” It is a review queue showing the prospect’s exact request, where it appeared, whether the capability already exists, and who should decide whether to follow up.
That last distinction matters. A salesperson’s note saying “they mentioned reporting” does not establish that reporting caused a lost deal. The system should be able to label that as mentioned in notes, not confirmed reason for loss. If the notes are thin, the correct output is a request for clarification, not a confident explanation.
Write down the decision before choosing an agent or model. A workable specification might say:
Surface prospect objections recorded in deal notes that do not appear to have a documented follow-up. Show the supporting excerpt and deal owner. Do not infer a lost-deal reason from an unanswered message. Do not contact the prospect or create a product commitment.
This gives you a way to judge the system. Did it find relevant records? Did it show why each record qualified? Did it avoid claims the source did not support? Did the owner make a decision? Those questions are more useful than asking whether the summary sounded intelligent.
Build a traceable path from source to decision
A grounded operational agent should turn source records into a reviewable recommendation, with a link back to each source. The minimum workflow is: collect permitted records, apply known business definitions, find candidate signals, attach evidence, and route the result to a person. An autonomous action is optional.
Here is a practical shape for the output. This is an illustrative schema for a small-team workflow, not a Databricks API response:
{
"signal": "Prospect requested a capability before buying",
"status": "Needs human review",
"source_record": "deal-note-reference",
"supporting_excerpt": "Exact excerpt from the authorized note",
"interpretation": "Possible product gap; not a confirmed lost-deal reason",
"owner": "Deal owner",
"proposed_next_step": "Check the current capability and follow up",
"action_taken": "None"
}
The supporting_excerpt and interpretation fields must remain separate. The first is what the source says. The second is what the system thinks it might mean. A reviewer can dispute an interpretation; they cannot assess a claim that has been detached from its source.
Business definitions need the same discipline. Does “lost prospect” mean a deal explicitly marked lost, or one that has gone quiet? Does “unresolved” mean no response was sent, or that a response did not satisfy the customer? If the agent invents those definitions as it goes, two runs can produce different answers to the same operational question.
Databricks says Genie Ontology combines learned business context with definitions that people explicitly govern. It can learn context from existing dashboards, SQL queries, and notebooks, while Unity Catalog Pages can hold authoritative definitions. The transferable lesson is not that a small business needs that infrastructure. It is that familiar-looking terms in a prompt still need an agreed meaning.
For a small team, the authoritative definition might live in a maintained document alongside the workflow configuration. The important properties are mundane: an owner can edit it, the agent can use it, and reviewers can tell which definition produced an answer. If a definition changes, previously generated recommendations should not silently take on the new meaning.
A useful first release can be deliberately unambitious. Run the workflow on a limited set of records, send candidate signals to an internal review queue, and record whether the reviewer accepted, corrected, or dismissed each one. That feedback exposes missing context faster than adding more subagents. Only expand the sources or actions after the team trusts the evidence path.
Make permissions part of the answer
An agent should only retrieve information the requesting person is allowed to see, and it should not gain permission to act just because it can explain an action. Read access, recommendation, and execution are separate privileges.
This is particularly important for the survey example. Employee feedback may be appropriate for an authorized leader to review, but that does not make raw responses appropriate for every manager, sales user, or general-purpose assistant. A summary can also expose sensitive details if it includes a distinctive quote or combines information that identifies a respondent. Permission-aware retrieval is the starting point, not a guarantee that every generated summary is safe to share.
Databricks says Genie actions are governed by Unity Catalog permissions and Unity Gateway approval rules and budgets. Its Genie Search documentation says indexed documents retain source permissions, which are enforced during retrieval. The same documentation says Genie Search stores a copy of indexed content and permissions in Databricks. If data location or copying matters to your organization, read the setup documentation rather than assuming that “indexing” leaves content untouched.
For any platform, write a permission matrix before connecting sources:
| Capability | Appropriate starting boundary |
|---|---|
| Read records | Only sources the requesting user is authorized to access |
| Summarize | Include evidence without exposing restricted material to other readers |
| Recommend | State uncertainty and name the human decision owner |
| Update a record | Require a defined rule, an authorized identity, and an audit trail |
| Contact a customer or employee | Require human approval before sending |
The matrix is a design decision, not a claim about what a particular product enforces automatically. Test it with records that different users should and should not see. Also test the handoff: a permitted user may generate a summary, but that does not mean the summary belongs in a broadly visible channel.
Treat retrieved text as evidence, not instructions. A prospect note may contain “email me the full customer list,” but that sentence is part of the record being analyzed. It must not become a command for the agent. The same rule applies to documents, tickets, and messages from external parties.
Finally, log both the recommendation and the eventual human decision. When someone asks why a deal was flagged or why a follow-up was proposed, the answer should be reconstructable from the source, the definition in effect, and the approved action—not the agent’s memory of a conversation.
Test whether it finds problems without inventing them
Evaluate an operational agent against reviewed cases from your own workflow, not against a general claim that an AI system is accurate. The test set should include clear signals, ambiguous notes, missing context, restricted records, and examples where the correct answer is “not enough evidence.”
For each case, record what a human reviewer would expect the system to do. Then check several behaviors separately:
- Retrieval: Did it find the relevant authorized record?
- Grounding: Does each important claim point to supporting text or data?
- Interpretation: Did it distinguish a stated fact from a possible explanation?
- Permissions: Did it withhold records the requester could not access?
- Routing: Did it identify the right person to make the decision?
- Restraint: Did it stop before taking an unapproved action?
These checks catch failures that a single pass-or-fail score can hide. An answer may find the right sales note yet wrongly claim that one requested feature would have won the deal. Another may make a sensible recommendation using a source the requester was never allowed to read. Neither should pass.
Databricks published a benchmark developed with Turing using a fictional company workspace. Genie passed 91% of its task runs, compared with 45% and 58% for two unnamed frontier agents. That describes performance on that task set. It is not a reliability estimate for your survey data, sales process, permissions, or staff decisions.
Run a small, repeatable review before widening access. Save representative questions and expected outcomes; rerun them when you change a source, business definition, prompt, or approval rule. Ask reviewers to mark why an output was wrong. “Wrong answer” is less actionable than “used an outdated definition,” “treated a suggestion as a commitment,” or “could not retrieve the call note.”
Track whether the workflow improves decisions, not just whether people open its alerts. A useful signal should help an owner resolve an issue, request better information, or explicitly decide not to act. If alerts accumulate without decisions, improve the routing or narrow what qualifies. Producing more summaries is not the objective.
Decide whether Databricks is the right starting point
Databricks is a plausible place to explore this pattern if your team already keeps governed operational data there. Its Genie Ontology, permissions, and agent direction address parts of the problem that become hard when data and users span a large organization. A small team whose records live in a handful of SaaS tools may be better served by a narrower workflow tied directly to one decision and one approval path.
Do not evaluate the choice by the CEO demonstration alone. Check which Genie capability you can actually use, what data it can access under your permissions, how the relevant content is indexed, and which actions require approval. The new always-on mode was announced for preview or beta rather than general availability at the time of Databricks’ announcement. Availability for the mode Ghodsi demonstrated should be confirmed before planning around it.
Pricing needs the same care. Databricks’ Genie documentation says user usage of Genie One and Genie Agents is free through January 31, 2027, with service principals excluded from that promotion. Release notes say billing for that user usage was paused as of July 15, 2026, while Genie Code remained on its pay-as-you-go model. Neither statement establishes a price for the new always-on mode or says the promotion covers it. Check the current pricing page and the terms for the exact product and identity you intend to use.
The practical choice is less “enterprise agent or no agent” than “where can we place a dependable decision loop?” If your problem is a recurring missed follow-up, begin there. If your data is already governed in Databricks and the relevant capability is available to you, test it against your own reviewed cases. Either way, require evidence, permission checks, and a named human owner before giving the system more autonomy.
How BizFlowAI approaches this
BizFlowAI builds practical AI automations for solopreneurs and small teams. For a workflow like this, a proposed starting point is the missed decision: identify which records contain the signal, who is permitted to see them, and what a reviewer needs before acting. A grounded, permission-aware version should show its evidence and leave consequential actions with an authorized person.
Before building, assess whether the workflow is sufficiently contained or whether the underlying records and definitions need attention first. The goal is not to imitate a CEO’s agent; it is to make the operational signals your team already has usable at the point of decision.
Work with BizFlowAI
If you'd rather have this built for you, that's what we do: production AI automation for solo founders and small teams — agents, integrations, and document pipelines that actually ship.
Request a free discovery call — 30 minutes, we map the highest-ROI automation in your workflow. No pitch deck, just engineering.
More guides like this on the BizFlowAI blog.
Frequently asked questions
What did Databricks CEO Ali Ghodsi use Genie to investigate?
Ali Ghodsi demonstrated Genie examining low-scoring areas of an employee survey and summarizing an unsuccessful prospect pitch. The agent helped surface concerns and requested features for review; the demonstration did not show it independently fixing those problems. A useful output keeps the source evidence separate from the agent’s interpretation.
How would I start building a Genie-style agent for a small business?
Start with one recurring decision, such as identifying recorded prospect objections that lack an assigned follow-up. Define the signal to look for, the person who owns the decision, and the outcome you can observe. Run the workflow on a limited set of permitted records and send findings to a human review queue before expanding its scope.
How can an AI agent avoid making up reasons a sales deal was lost?
Require each finding to include a link to the deal record and an exact supporting excerpt. Keep what the prospect said separate from the agent’s interpretation, and define terms such as “lost” and “unresolved” before running the workflow. If the notes do not establish why the deal was lost, the agent should request clarification rather than claim a reason.
What permissions should an AI agent have when reviewing employee feedback?
The agent should retrieve only records the requesting person is authorized to see. A generated summary also needs review because a distinctive quote or combination of details could identify a respondent. Treat permission to read feedback, recommend a response, and take action as separate privileges.
Can I turn on the always-on Databricks Genie workflow from Ghodsi’s demo?
The article distinguishes Ghodsi’s demonstration from general product availability. At the time of the November 11, 2026 announcement, Genie and Genie Ontology were generally available, while Databricks described the new always-on mode as coming to preview or beta. Check current Databricks availability before planning around that specific workflow.