What problem does a data contract solve?
The silent breaking change. A producer renames a column, changes a unit, or starts writing nulls, and nothing fails loudly: the pipeline runs, the numbers drift, and the error is discovered weeks later in a business review. A contract converts that silent drift into a loud, early failure: the producer’s change violates a published commitment, and automation catches it before consumers do. It is the data equivalent of an API spec between services.
What goes into a data contract?
Five sections recur. Schema: columns, types, nullability. Semantics: what each field means and how it is derived, which overlaps with what a semantic layer encodes. Quality rules: ranges, uniqueness, completeness thresholds. Delivery: update frequency, latency, retention. Change management: how breaking changes are announced and versioned. The Open Data Contract Standard gives this a shared YAML shape (Bitol ODCS, read 12 October 2026). A contract nobody enforces is documentation; the checks are what make it a contract.
Why do AI systems make data contracts urgent?
Because models consume data without skepticism. A RAG system grounded in a table whose meaning drifted will answer fluently and wrongly; an agent acting on a broken feed acts wrongly at machine speed. Press coverage of agent governance keeps landing on the same point: agents need governed data with known lineage before they deserve autonomy. Every dataset an AI system reads or acts on deserves a contract first, which is also the concrete half of data readiness for AI.
How does a team adopt data contracts without a big program?
Start at the painful interface, not everywhere. Pick the one dataset whose breakages cost the most, write the contract down with the producer, and wire the checks into their CI so violations block the change. Expand only along real dependencies. One enforced contract on one critical table beats a company-wide standard on paper. Ownership follows naturally, and the upstream discipline keeps downstream models honest: unannounced input changes are a leading cause of the decay that model drift describes.