Consider Structured Data
Learning Objectives
After completing this unit, you’ll be able to:
- Identify structured data issues that can cause unreliable AI agent responses.
- Explain how duplicate or disconnected records affect an AI agent’s understanding of customer context.
Structured Data Provides Context for AI Agents
While unstructured documents can relate to a specific interaction, transaction, or policy, structured data provides the operational context for AI agents to understand customers, transactions, and business activity.
At Northern Trail Outfitters (NTO), the Case Deflection agent must evaluate customer inquiries using structured data from multiple systems. This includes CRM customer records, ecommerce transactions, store orders, and product inventory information. For example, unstructured content in shipment confirmation images might relate to a specific order, but the order record itself provides the structured data that connects that interaction to a customer profile and purchase history.
Similarly, some business policies rely on metrics that must be calculated across multiple transactions. Accommodation policies, for instance, can reference a customer’s total lifetime value. Calculating this metric requires combining data from all relevant orders—whether they were placed online, in a physical store, or through guest checkout. When this structured data is complete, consistent, and connected, the agent can accurately understand a customer’s situation and provide reliable responses.
Duplicate records, inconsistent values, missing relationships, or sensitive information stored in unexpected places can affect how an AI agent interprets the situation. If these issues are not identified, the agent can provide answers that appear reasonable but are based on incomplete or misleading data.
Common Structured Data Risks
During the case deflection proof-of-value project, Luna identifies several structured data risks that could affect the reliability of agent responses.
Risk | Example | Impact |
|---|---|---|
Duplicate customer records | The same email, rachel@mystyle.com, repeats across multiple contacts with a similar name, where both contact records have associated business transactions. | The agent might not access all related transactions and interactions, leading to the incorrect conclusion and responses. |
Disconnected transactional data | Guest checkouts are not linked to a customer. Service Cloud cases might not be linked to a contact. | The agent cannot see a customer's full transactional history, resulting in incomplete responses. Consider transactional records that contain customer-identifying information, not just customer data objects, when designing customer profile unification processes. |
Unused fields or unreliable values | Even if the HasOptedOutOfEmail Boolean field is not used, it will be populated with a null value. Fields with only one distinct value are considered unreliable. | The agent might assume the information in the fields is correct; for example, null values in binary fields mean“no.” |
Inconsistent operational values or granularity | Product code or categories (camping gear versus tents) differ between ecommerce and store inventory systems. | The agent might recommend a product that appears available online but cannot actually be fulfilled at the customer’s local store. |
Sensitive data exposure risk | Personally identifiable information (PII) exists in customer records without clear classification or access controls. | The agent might retrieve or expose sensitive information that should not be included in automated responses. |
Large language models do not inherently understand that duplicate records exist or how to resolve them. Without identity resolution and unified profiles, the agent can produce responses based on partial or overly generalized context.
By understanding these risks, you, just like Luna, can assess and prioritize remediation strategies for structured data.
Move on to the next unit to discover the important role metadata plays in AI reliability.