Databricks has shipped the rulebook and is still advertising the employee.
The company wants Genie to do more than answer enterprise questions. Its announced second-generation direction is a system that pursues goals in the background, delegates tasks to subagents and asks for decisions through a priority inbox or email.
Only part of that exists. Genie and Genie Ontology are generally available. Always-on Genie, Genie Search and Ontology Manager are headed for preview or beta programs. (Databricks announcement)
CTOL's view is that Databricks' most credible near-term position is defining the work before an agent performs it. Genie Ontology offers a shared account of business metrics, definitions, people and processes. Databricks proposes wrapping autonomous execution in permission, approval, budget and audit controls. That is a valuable seat in an enterprise's operating architecture, though Genie has yet to show it can own a workflow.
A weekly variance report shows the gap
Take one recurring process. A weekly forecast-variance investigation is an illustrative candidate, and no customer has been reported running it. Today a human analyst checks an agreed metric, reads the relevant documents and issue records, then asks an owner to sign off on the explanation.
The available Genie components can be judged now on one thing: whether they interpret the business consistently. The proposed always-on system has three more jobs. It must perform the handoffs, request approvals and leave an auditable result.
The unit of value is the correctly completed, approved investigation. Prompts, subagent calls and jobs started measure activity. The all-in cost per successful case adds compute, integration, human review, retries and correction work, and it belongs next to the same process performed today. A buyer should also test whether a failed or unauthorized action can be stopped and reversed.
The announcement gives no job-level results, so CTOL makes no savings forecast.
Authority is the product
The proposed sequence is easy to follow.
- Genie Ontology supplies business meaning.
- An always-on agent applies that context to a goal, spreads work across subagents and asks humans when a decision is needed.
- Genie Search extends discovery across enterprise applications.
- Databricks says Unity Catalog and Unity Gateway provide the permissions, approvals, budgets and audit controls around all of it. (Databricks architecture)
The hard problem sits at the handoffs. A stale definition can make several subagents repeat the same wrong interpretation. Access to a document in one system gives zero authority to change a record in another. A budget rule for compute is a separate check from an approval to send an instruction or close a ticket. A user experiences one task, and the system runs several checks.
So the procurement question is where authority lives. For each proposed action, a buyer needs to know whose permission applies, where approval is enforced, what gets logged and how an incorrect change is reversed. The announced architecture names the components of such a control path. Whether every external handoff enforces it is still unproven, which is the principal technical caveat on a sound direction.
If the control chain works, business definitions and action rules could make the platform harder to remove from a recurring process. That goes beyond a better answer to a data question. If it fails, governance becomes a human workload layered onto an assistant, and the automation savings never arrive.
Two of five is a count, and function is the score
Counting generally available offerings misleads. Genie and Genie Ontology give buyers something to examine today. The background execution loop, Genie Search and Ontology Manager are on the announced preview-and-beta path, and the workflow-ownership claim depends on exactly those three.
Databricks says Genie Search can connect to more than 100 enterprise sources, including Google Drive, Jira and Confluence. That widens the proposed reach of an investigation. Every added destination also brings its own access and approval boundary, so source coverage is a distribution claim and completed cross-system work is the only completion metric. (Databricks announcement)
Buy the context layer, audition the agent
That sets the buying rule, and CTOL holds it with high confidence while making no claim about eventual market-wide workflow ownership. Evaluate semantic consistency and permissions now. A CIO can test the available context and question-answering tools on a bounded business task, and keep approvals and changes to systems of record under existing authority. An investor should read the announcement as a path to deeper enterprise attachment, since no estimate of revenue from displaced work has been given.
CTOL's 31 January 2027 checkpoint is an editorial date and no Databricks launch forecast. By then two things together would upgrade the call: a dated general-availability notice for the always-on execution path, and a documented customer comparison of correctly completed, auditable jobs and their all-in cost against the previous workflow. Evidence that permissions hold through cross-system handoffs would matter as much as the cost result, and general availability alone is too little.
Until that proof appears, Databricks' opportunity is to become the shared context and governance layer around enterprise work. Becoming the operator of the work is the harder proposition, and the enterprise keeps responsibility for outcomes either way.