Know Your Metadata
Learning Objectives
After completing this unit, you’ll be able to:
- Describe why metadata is important in enterprise AI reliability.
- Identify common metadata gaps that can negatively impact AI responses.
- Understand how data dictionaries, classifications, and field-level governance contribute to trustworthy agent behavior.
Previously, you learned the importance of having reliable unstructured and structured data for AI agents to work properly. The agent must be able to search and find the right data to complete its instructed tasks.
Why Metadata Matters
Metadata is structured data that describes other data and provides context about it. With clear metadata, an agent can search for and find the right data to complete its tasks.
Gen AI relies on metadata like field names, descriptions, and classifications. When key metadata is incomplete or misleading, agents can interpret fields incorrectly or surface information a user shouldn’t see.
As an example from Salesforce CRM, the custom field Contact Owner Email is missing metadata values for the description, help text, field usage, data sensitivity, and more.
Luna knows that in the past year, the NTO team ensured any new or updated CRM object field is well-documented with a clear description, help text, business owner, and if applicable, sensitivity classification. However, she suspects that older fields might be missing this documentation.
Also, since the Case Deflection agent brings together data from multiple sources, including orders containing personally identifiable information (PII), Luna adheres to essential data governance practices. She ensures that data dictionaries and sensitivity classifications are in place so that agents are grounded in reliable, compliant data.
Common Metadata Risks
Luna reviews three common metadata issues that can lead to inaccurate and unhelpful AI agent output.
Metadata Risk |
Example at NTO |
Impact on Agent Responses |
|---|---|---|
Misleading or incomplete field labels and descriptions |
NTO’s standard CRM fields do not have descriptions, as they are available in help and training. Many custom CRM fields are missing field descriptions or help text. How critical a field is to the use case determines which fields need descriptions. NTO also uses Informatica data catalog as an enterprise repository. Luna identifies this as a trusted source to augment the Agentforce reasoning engine. |
If metadata is unclear or missing, the agent can misinterpret the purpose of a field or use the wrong data when generating a response. Clear metadata helps ensure the agent retrieves and uses the correct information. |
Missing sensitivity classifications and security metadata |
Some external datasets include personally identifiable information (PII) but lack embedded sensitivity classifications or compliance metadata. Luna uses Data 360 governance capabilities to classify fields using labels such as personally identifiable information (PII) or payment card information (PCI) and ensures proper field‑level access controls are applied. |
Without sensitivity classifications, the agent can retrieve or expose restricted information. Proper metadata ensures the agent respects governance policies and avoids referencing sensitive data in responses. |
Metadata drift and unexpected changes |
Over time, field usage or governance rules can change. Luna monitors critical fields—such as customer communication preferences—to track changes in fill rates or value distributions. Automated monitoring alerts her when unexpected changes occur. |
If metadata or field usage changes without oversight, the agent can begin relying on outdated or unreliable fields. Monitoring helps detect drift early and keeps agent behavior aligned with current data governance practices. |
After reviewing metadata risks in NTO’s data, Luna makes two recommendations:
- Use Data 360’s data sensitivity classification feature to assess all data sources, not just CRM data.
- Evaluate data dictionary completeness in NTO’s Informatica data catalog. If descriptions and metadata are complete, incorporate the data catalog as a data source.
With the pilot findings at her disposal and important data reliability risks called out, Luna can begin the data readiness assessment.
To continue Luna and Charlotte’s journey in preparing NTO’s data for AI, check out AI Data Readiness Assessment Fundamentals.