8.3.2.2 - Policy Hallucinations: When Bots Promise Refunds or Discounts You Can't Honor (Difficulty: Advanced | Path: Scale)

Imagine waking up to find that your customer support agent has issued 50 full refunds for products that were marked "Final Sale," promised a 75% discount code to a disgruntled influencer, and guaranteed next-day delivery on a backordered item. Now imagine that agent isn't a rogue employee you can fire, but the Artificial Intelligence you just installed to "save time." This is not a glitch; it is a fundamental characteristic of how Large Language Models (LLMs) function. It is called a Policy Hallucination.

In the rush to automate customer service, brands often deploy generative AI with the instruction to "be helpful." Unfortunately, LLMs are probabilistic engines, not truth machines. They do not "know" your return policy; they predict what a return policy sounds like based on the millions of websites they were trained on. If the model decides that a "helpful" response involves granting a refund, it will confidently fabricate a policy that allows it. In the eyes of the law—as confirmed by recent high-profile court cases like Moffat v. Air Canada—your business is liable for these promises. If your bot says it, you bought it.

This masterclass is your defense strategy. We are moving beyond simple "prompt engineering" into the architecture of Grounded Generation and Deterministic Routing. You cannot simply ask the AI not to lie; you must strip it of the authority to invent. We will explore why "pure" generative AI is dangerous for financial queries and how to implement a hybrid architecture where the AI handles the conversation, but rigid, hard-coded logic handles the wallet.