Managing the Transition From Predictable API-Based Software Toward AI Agents

More companies are shifting from API-based software to autonomous AI agents in order to stay competitive. However, this transition demands careful planning to address potential issues, such as hallucinations and their consequences. Traditionally, businesses relied on predictable platforms and APIs, which followed established rules for authentication, authorization, and transactions.

In contrast, agentic AI operates by interpreting a user's intent and independently deciding which actions to take, introducing uncertainty and the need for oversight.

The key question for businesses considering AI adoption is how to leverage autonomous systems while maintaining accuracy, control, and accountability. APIs remain essential to agentic AI, as agents depend on them to access information and execute actions. Ashay Satav, a technology and product leader with experience at Blackhawk Network, Intuit, and eBay, emphasizes that platform architecture plays a crucial role in determining which actions are available, permissions are enforced, and systems respond at scale.

Satav argues that the platform layer, not just the underlying model, determines an agentic system's trustworthiness with real business consequences. He stresses that even the most intelligent model can become a liability if the underlying API layer fails to enforce allowed actions, scales appropriately, and provides fallback mechanisms.

Hallucinations, where AI systems generate inaccurate or misleading information, pose a significant challenge for AI adoption. In e-commerce, for example, an AI listing tool might correctly identify an item's brand but embellish its condition or features in the description, potentially misleading customers. This can lead to business consequences like returns, customer complaints, reduced trust, and negative seller performance metrics.

Ashay highlights the importance of controlling hallucination rates to minimize buyer remorse, which is effectively measured by the "Seriously Not As Described" (SNAD) metric on e-commerce platforms like eBay.

Retrieval-Augmented Generation (RAG) is one approach to mitigate hallucinations by retrieving relevant information from approved sources before an AI model generates an answer or takes an action. For instance, an e-commerce RAG system could examine historical listings and verified product information to prevent the model from inventing details.

However, RAG effectiveness depends on factors such as the accuracy of source data, the use of metadata filters, and the retrieval of semantically related records. To optimize RAG systems, teams must experiment with chunk size, metadata, similarity thresholds, retrieval depth, and ranking methods. Ultimately, finding the right balance between retrieval and generation is critical, as poor retrieval can still provide irrelevant or misleading context.