Data Ownership Without an On-Call Rotation Is Just a Spreadsheet

Walk into almost any enterprise and ask who owns customer master data. You will get an answer within thirty seconds, usually a name, sometimes a RACI chart, occasionally a glossary entry in a catalog tool that cost seven figures. Now ask a harder question: when that dataset silently changed shape last quarter, who got paged? Who decided whether the three downstream consumers could absorb it? That question tends to produce silence. The gap between those two answers is where most AI programs quietly fail, and no amount of additional strategy documentation closes it.

The numbers are hard to argue with. Gartner has predicted that through 2026 organizations will abandon 60% of AI projects unsupported by AI-ready data, and found that 63% of organizations either lack or are unsure of the data management practices needed to support AI. Informatica's CDO Insights 2025 survey of data leaders puts data quality and readiness at the top of the obstacle list, and a recent Cloudera report found that roughly four in five enterprises say data access problems are holding AI back. None of these are modelling problems. They are operating model problems.

Governance documents degrade; governance operations don't

Traditional data governance was built for a world of quarterly reporting cycles, where a stale definition caused an awkward conversation in a steering committee two months later. A retrieval pipeline or an agent calling a product catalog consumes that same data thousands of times a day and has no capacity to notice that the definition drifted. Documentation has a half-life. The moment a schema changes, a source system is re-platformed, or a region adds a field, the written artifact starts decaying — and nothing in the process forces it to be updated, because nothing breaks loudly enough.

Software engineering solved the equivalent problem decades ago. We stopped writing documents describing how services should behave and started instrumenting them, alerting on them, and putting a human on-call. Data governance for AI needs the same move: fewer artifacts, more running feedback loops. The practical test of whether a dataset is governed is not whether it appears in the catalog. It is whether someone's phone buzzes when it misbehaves.

What an operational owner actually owns

If you want to know whether ownership is real, check for these five properties. Most enterprises have one or two.

• A named on-call, not a named department. "Finance owns it" is not ownership. A rotation with an escalation path is.
• A published freshness and completeness expectation. Consumers need to know whether the dataset is hours stale or days stale before they build an agent on top of it.
• A deprecation calendar. Owners must be able to retire fields and versions on a schedule, with notice. Without that, every dataset accumulates compatibility debt forever.
• A refusal right. An owner who cannot say "no, this dataset is not fit for that use case" is an administrator, not a steward.
• A funded maintenance line. If upkeep is unpaid work squeezed between project deliverables, it stops the first time a release date slips.

Scope governance to the AI surface, not the estate

The second failure mode is ambition. Enterprise-wide governance programs that attempt to classify everything deliver their first usable output eighteen months in, by which time the sponsoring executive has moved on. The alternative is to work backwards from the AI surface: list the use cases actually heading to production, trace the specific datasets they read and write, and govern that set properly. It is usually a surprisingly small list — a few dozen entities carrying most of the risk.

This is where governance stops being a cost center. The same owner who can certify a dataset for an AI use case can also certify it for regulatory reporting, partner exchange, and analytics. In our architecture and AI advisory work, the fastest-moving clients are the ones who treat the AI backlog as a prioritization function for data work rather than a separate workstream competing with it.

Make the lineage question answerable in minutes

When an AI output is challenged — by a regulator, an auditor, or a customer — the question is always the same: where did this come from and who approved it? Organizations that can answer in minutes treat lineage as telemetry emitted by pipelines. Organizations that answer in weeks treat it as a diagram someone drew. The difference shows up most sharply in complex ERP and commerce landscapes, where a single customer attribute may traverse five systems before reaching a model; our SAP practice sees this pattern constantly.

The organizations that will scale AI through the next two years are not the ones with the most sophisticated governance framework. They are the ones who staffed it, scoped it to what matters, and made data ownership an operational commitment with a pager attached. Strategy decks age badly. Rotations don't.