A broken data pipeline is an operations problem, not a data team ticket
Databricks says over half of 80,000 pipeline issues needed company-specific knowledge to resolve. That knowledge sits in the operation, not in the data team.
On 19 November 2026 Databricks opened Genie ZeroOps, an agent that monitors data pipelines and proposes fixes, to public beta. The most useful line in the announcement is not about the agent. During the private preview it observed over 80,000 real customer issues, and Databricks found that "over half of issues require organization-specific knowledge to properly triage, rank, and resolve." That finding says where data incidents really come from: the business changed and nobody told the pipeline. So the fix belongs to the operation, not to a queue in the data team.
The easy take: let the agent keep the pipelines green
The appeal is real. Pipelines fail at night, tables stop updating without an error, a schema changes three tables upstream and a dashboard quietly goes wrong. An agent that watches all of it, traces the cause through lineage, tests a fix on a clone of the data and opens a pull request for approval removes a lot of tedious work. Databricks is careful here: it says ZeroOps never changes production code or data without a person reviewing it first.
Source: Databricks blog, "Genie ZeroOps, now in public beta", 19 November 2026.
Nothing here argues against the tool. The question is who reads the pull request, and what they are being asked to decide.
Most of those issues are business questions
Think about what "organization-specific knowledge" means in practice. A new product line started selling with a code the reporting table does not expect. A supplier changed the format of its file. A region was merged into another for commercial reasons. A sales team began logging discounts in a different field. A column of null values may be a bug, or it may be correct because a store closed last week.
None of those are data engineering decisions. They are decisions the business made, and the pipeline is the first place they show up. A data engineer reviewing the fix can tell whether the code is correct. They usually cannot tell whether the new number is the right one, because the answer depends on something that happened in sales, finance or operations.
That is why the knowledge is "organization-specific". It is not missing from the model. It is missing from the path between the people who change the business and the systems that record it.
An approval queue in the wrong room
Here is what happens when the incident stays a data team ticket. The agent finds the break and proposes a fix. A data engineer approves the version that makes the job pass. The pipeline is green again. The commercial change that caused the break was never discussed by the person who made it, so it happens again next month, in a slightly different shape, and the agent fixes it again.
The work looks healthy. Fewer failed jobs, faster recovery, a clean dashboard. But a pipeline that heals itself every week is telling you that a business process keeps changing without a record of the change. Faster recovery makes that signal easier to ignore.
Databricks seems to see the same thing: after the preview it added shared inboxes, custom agent instructions and notifications. Those are routing features. The question is where you route to.
Route the incident to the owner of the step that changed
Treat a data incident as an event in the operation, with an owner on the business side:
- 01DetectThe agent flags the break or the silent drift
- 02Trace upstreamLineage finds the source table and the step that feeds it
- 03Ask the step ownerWas this a business change? Is the new value right?
- 04Approve the fixData team approves the code, step owner approves the meaning
- 05Record the changeThe business rule is written where the next agent can read it
Recurring incidents on the same step go to the redesign backlog, not to the fix queue.
In practice that means four decisions, made before the agent is switched on:
- Every source table has a business owner, not only a technical one. The person who can say whether a new value is right.
- The fix needs two approvals when meaning changes. The data team approves the code. The business owner approves that the number now says what the business means.
- Every resolved incident leaves a rule behind. "Region North now includes the former West territory, from 1 November." Written where people and agents can read it.
- Repeat incidents are counted per step. The third break on the same source is not an incident. It is a process that needs to be redesigned.
Connecting sales and analytics was the work
Sentynel's Campoverde case started from a familiar problem: sales had one view, analytics had another. Two views of the same business mean every discrepancy turns into a debate about whose number is right. The system built there turned sales and analytics into a shared decision layer, so the board could ask its questions of one source. Then a Walmart campaign was built from what the data revealed. Sentynel's case record reports a +18% sales uplift; the uplift is a reported campaign outcome and the chart uses a normalized baseline.
The point is not the uplift. It is that the decision layer only works when the people who run the business and the people who run the data agree on what each number means, and keep agreeing when the business changes.
A first test
Pull the last 30 data incidents your team resolved. For each one, write who outside the data team could have explained the cause. If more than half have a name on the business side, your incidents are not a data quality problem. They are an operation that changes without telling its own records. An agent can keep the pipelines running. Deciding who owns the change is still your job.