A data contract is a promise with consequences

A data contract is an enforceable agreement about what a dataset contains: its schema, value constraints, freshness, and who is accountable when those break. The useful version has a test behind every clause, so the contract is evidence that ran last night rather than a document somebody wrote once.

By turning each clause into a test that runs on a schedule and reports pass or fail. Enforcement is the difference between a contract and a wish. Without the test, a breach is discovered by whoever consumes the data, which is exactly the situation the contract was meant to prevent.

Because teams treat them as legal negotiations rather than as receipts. Six weeks of meetings produce a PDF nobody checks. Starting with one dataset, encoding what you already promise informally, and shipping the passing test alongside the data teaches more than a program that tries to contract everything at once.

Which columns exist and their types, which fields are never null, acceptable value ranges, expected freshness, and expected volume. Each of those maps to a test. Anything you cannot express as a test is a preference rather than a contract term.

Yes, and it is the case that matters most. AI consumes data without a human reviewing it, so there is no step where somebody notices a column changed meaning. A contract is the checkpoint that catches a schema change or shifted value range before it reaches training or inference.

Pick one dataset you already ship to somebody and write down what you implicitly promise about it. Turn each promise into a test. Publish the result with the data. One working receipt is more persuasive than a governance framework covering everything.