One Person Owns the Data or the AI Lies

Your CRM says a customer is active. Your finance tool shows their last payment was fourteen months ago. Your AI-powered churn model never flags them because it trained on the CRM record. Nobody is lying. Nobody is even wrong, exactly. The systems just drifted apart, and no single person had the job of noticing.

The defects tools do not catch

There is a reasonable argument that tooling solves this. Required-field validation stops blank entries at the point of creation. Nightly deduplication removes duplicate records before a model trains on them. Automated syncs between systems eliminate stale statuses without anyone monitoring lag.

This argument is correct about structural defects. A required field stops a record from saving without a customer name. It does nothing about a staff member entering "Active" for a customer who churned last month, because "Active" is a valid value. It is just wrong in context. Deduplication catches records sharing the same identifying attributes. It does not catch two records for the same customer entered under slightly different company names, each carrying different status values, because both pass validation.

Tools catch what violates a rule. Semantic drift — the gap between what a record says and what the business now means by it — is invisible to any automated check.

What actually happens when no one owns the domain

When multiple staff members enter data into a shared system without a named owner, each person makes local decisions: which fields matter, how to handle exceptions, when to bother updating a record. Nobody audits completeness. Nobody defines accepted values. Nobody tracks whether the update cycle still matches how the business operates.

The first month after deployment, records stay clean. Later, other tasks compete, and updates become sporadic. A sales team renames customer categories in the CRM without telling the analytics team. A finance change renames revenue fields in a spreadsheet but never reaches the AI training pipeline. No observability tool catches either of these, because every individual system is valid. The mismatch between them is invisible until a model starts producing outputs that contradict what the team knows to be true.

This is the failure mode the research identifies: not missing fields, but conflicting values and outdated statuses that pass every validation check and still mislead every model trained on them.

The numbers behind the pattern

Dresner Advisory Services reported in 2024 that only 31-32% of organizations have a formal data governance structure. A 2026 report on database environments and AI adoption found that only 23% of firms use formal governance or quality frameworks, despite active AI investment. Separately, surveys on data governance challenges identify unclear ownership and accountability as the leading barrier to scaling governance.

85% of AI projects fail. Data quality drives roughly 70% of those failures.

The tooling the counterargument describes has been widely available throughout this period. The failures persist anyway. The gap between them is not a missing software feature.

What a named owner actually does

The role is not auditor. One person owns a data domain, which means they define which sources qualify as training inputs, align schemas and terminology across systems, and track changes over time so that training data respond when the business changes. When a new product line launches and existing category codes no longer apply, the owner updates the validation rules. When a finance team renames a revenue field, the owner connects that change to downstream pipelines before a model trains on the old label.

Access policy sits in the same role. AI training needs exposure to sensitive data — customer interactions, transaction histories, support logs. Without a named owner, extraction scripts appear ad hoc, access spreads without oversight, and the quality and security problems compound each other.

Governance frameworks have recommended this structure for years. The reason only 23% of firms with active AI investment have implemented it is not that founders disagree with the logic. It is that naming an owner feels like adding overhead to an already stretched team. That calculation ignores the cost of the alternative: models trained on data nobody is responsible for, producing outputs nobody trusts, in a company that spent real money on AI and got noise.

Name one person. Give them explicit authority over accuracy, updates, and access for the data feeding your models. The role does not require a title or a budget line. It requires a name.