Most failed AI initiatives we have seen died upstream, in a schema change nobody announced. Data contracts are the cheapest insurance you can buy.
A data contract is an agreement between the team that produces a dataset and the teams that consume it: schema, semantics, freshness, quality checks and who to call when it breaks. It is enforced in CI, not in a wiki.
Why it matters more for AI
Dashboards tolerate a broken column for a day. A model retrained on silently corrupted features can ship wrong decisions for weeks before anyone notices. Contracts turn those incidents into failed builds.
Where to start
- Pick the ten tables that feed your most important decisions.
- Write the contract as code: schema, tests, owner, SLA.
- Fail the pipeline when the contract is violated, and page the owner.
Once the critical path is covered, extend contracts gradually. Teams adopt them quickly when they see the first incident that did not happen.
- #Data platform
- #Governance