“You own your data” is easy to put in a contract. The real test begins when a customer leaves, changes a provider, questions a derived score or wants to understand why an automated decision was made. Ownership without practical control creates dependence disguised as reassurance. A data platform should therefore be judged by the exit it makes possible, not only the service it provides while the customer stays.
There are at least four distinct questions. Who is permitted to collect an event? Who can interpret and enrich it? Who may reuse a derived insight? And who can demand deletion, portability or a complete audit trail? These rights need not belong to the same party. Collapsing them into one vague ownership sentence hides the very negotiations that determine trust.
The architectural consequences are concrete. Keep tenant boundaries explicit. Separate source records from computed features. Track lineage, retention and the version of definitions used to create a metric. Make exports machine-readable; test restore to another system. An access token must not imply permission to train a model, and a data-processing agreement must not silently become a permanent license to business knowledge.
A powerful test is to simulate a difficult departure. Can the customer retrieve their original data and usable context? Can they prove which parties received it? Are derived artifacts governed by clear terms? The future data platform will compete on its ability to be trusted after the relationship ends. Portability is not only compliance. It is product quality.
Four rights that should not be confused
Data access is the ability to see or retrieve a record. Data control is the ability to define who else can use it. Data portability is the ability to take the content and its operational meaning elsewhere. Model-use rights concern what can be inferred from information and for which purposes. These are separate relationships. Organizations should document each one and make their consequences testable rather than describing all of them with the word ownership.
A useful departure rehearsal
Ask a customer to export its original records, documentation of fields, relevant relationships, history of changes, and access logs. Then ask a different team to reconstruct the information in a new environment. If the export cannot be interpreted outside the former vendor, the organization has transferred bytes without transferring usable control.
The difficult counterargument
Full portability is not costless. Derived models may contain mixed contributions, security logs may need regulated retention, and exposing raw infrastructure metadata can create abuse risks. The solution is not unconditional disclosure of everything. It is a contract that distinguishes customer data from operational security evidence and proprietary provider methods, with a documented process for resolving conflicts.
What I would measure
Time to complete a verified export; percentage of business concepts represented in an open schema; latency of access revocation; coverage of lineage records; and whether the customer can independently validate deletion or the documented legal reason for retention. These are better indicators of practical control than a checkbox on a settings page.