Agent Ready Data Series :

01  From AI Ready Data to Agent Ready Data

02  Is a Data Product Ready for an AI Agent

03  Data Products Need Owners Not Only Schemas

When a customer says a data product is ready, I like to ask who will answer when an agent gives the wrong answer. The schema may be clear and the pipeline may run every day. But if a business definition changes or a source starts arriving late, somebody must own that change.

Figure: The owner remains accountable for changes and exceptions after a data product is published.

Treat the interface as a product

Take a supplier performance data product. Its consumers may include a dashboard, a sourcing analyst and an agent that recommends which supplier needs attention. Each consumer needs to know the unit of analysis, refresh schedule, permitted use, expected quality and contact for an issue. A table name and column list tell only part of that story.

I would define a small product agreement: business purpose, owner, source systems, grain, measure definitions, freshness target, quality rules, sensitivity and access method. I would also describe what happens when the schema or a business rule changes. These are design choices for the team that publishes and consumes the data, not properties I would assume exist automatically.

The agent changes the support model

A human analyst may notice a missing country or an unusual supplier name and investigate. An agent might continue to the next step before anyone notices. That makes ownership and change communication more urgent. If a quality rule fails, should the agent stop, answer with a warning or route the case to a person? The data product owner and process owner should decide together.

SAP Business Data Cloud provides a data product approach, and the catalog helps users find and inspect the assets available to them. My recommendation is to add a clear operating agreement around each high-value product used by agents. Start with one use case, test the agreement with its consumers, then reuse what works in the next domain.

The strongest sign of a mature data product is not that it has been published. It is that a consumer knows what it means, whether it can be trusted today and who will resolve an issue tomorrow.

This also changes how I measure success. I would track whether consumers can find the product, understand its scope and resolve an issue without a long chain of emails. An agent pilot is a good stress test because it exposes unclear ownership quickly.

The product owner should also hear how the agent is being used. If users repeatedly ask questions the product cannot answer, that is input for the product roadmap. If an agent frequently stops because a field arrives late, the owner can prioritize that issue with the source team. Consumption feedback should improve the product.

What to read next

Source link