# Data Product

**TL;DR:** Data packaged with a contract: defined schema, named owner, service levels, access policy, and documentation, so consumers can depend on it.

A data product is a dataset plus commitments. The data is the smaller half. What makes it a product is the contract around it: a schema that will not change without notice, a named owner accountable for it, a stated freshness guarantee, a documented access policy, and documentation that explains what it means and what it excludes. Consumers can build on it the way you'd build on any internal service, which is the entire point.

The idea came out of the data mesh conversation, where the argument was that centralizing all data engineering into one team creates a bottleneck that scales badly, and the alternative is domain teams owning and publishing their own data with product-grade rigor. You can disagree with the wider mesh architecture and still find this piece useful, because the underlying insight is about accountability rather than org structure: data with no owner degrades, and degradation is invisible until something downstream breaks.

The contract usually specifies a versioned schema with a deprecation policy, freshness and availability SLAs, quality expectations expressed as testable assertions, a documented access and PII classification, and a defined channel for consumers to report problems. That last one sounds minor and isn't. A dataset with no support channel means problems get discovered by whoever's dashboard looks wrong, weeks later.

Compared to a plain dataset, the difference is entirely about whether anyone made promises. Same rows, same columns, potentially the same physical table. One has an owner who gets paged when freshness slips; the other has an owner in the sense that someone created it once. That gap is why the same table can be perfectly adequate for exploratory analysis and completely unsuitable as the foundation of a production AI system.

For AI teams this is becoming the practical prerequisite for RAG and agent systems that stay correct. If your retrieval corpus is a data product, someone is accountable for it being current, changes are announced before they land, and quality is monitored. If it's a dataset, your system's accuracy silently depends on someone's forgotten cron job. Treating the retrieval corpus as a first-class data product with an owner and a freshness SLA is one of the least glamorous and highest-return investments in production AI.

## Why it matters

The gap between a working AI prototype and a system the business can rely on is mostly a data reliability gap, not a model gap. Data products make dependencies explicit and failures attributable: when freshness slips, there is an owner, an SLA, and an alert rather than an archaeology project. As AI systems consume more internal data and act on it automatically, undocumented dependencies stop being technical debt and start being operational risk.

## Example

The customer support team publishes a "resolved tickets with verified solutions" data product. The contract: refreshed hourly, schema versioned with 30 days notice before breaking changes, PII redacted at write time, owned by the support operations team with a Slack channel for issues, quality tests asserting that no resolution field is empty and that no ticket is older than its stated close date. Three AI systems consume it. When the ticketing vendor changes a field name, the owner gets the failing test, publishes a v2 schema, and gives consumers a month, rather than three teams discovering it through broken retrieval.

## Related terms

- [Dataset](https://www.maximem.ai/glossary/dataset)
- [Data Catalog](https://www.maximem.ai/glossary/data-catalog)
- [Data Observability](https://www.maximem.ai/glossary/data-observability)
- [Data Lineage](https://www.maximem.ai/glossary/data-lineage)
- [AI Data Governance](https://www.maximem.ai/glossary/ai-data-governance)
- [Enterprise AI Stack](https://www.maximem.ai/glossary/enterprise-ai-stack)

---

Source: [https://www.maximem.ai/glossary/data-product](https://www.maximem.ai/glossary/data-product)