AI doesn't have a conscience, but it will absolutely bankrupt you out of politeness.
An e-commerce brand deployed a Retell AI voice agent to handle customer service calls. They didn't lock down the function-calling parameters with strict Standard Operating Procedures.
A customer asked for a refund on a used, out-of-warranty item. The agent wanted to be helpful. It hallucinated a "lifetime satisfaction guarantee" and authorized a $300 return on the spot.
Word got out in a forum. By the time the founder caught the anomaly, the agent had cheerfully approved thousands of dollars in invalid refunds.
The cost to fix it wasn't just the lost revenue. It was the engineering hours spent ripping the agent offline, auditing the damage, and manually reversing the approvals.
This is what happens when you treat AI like an employee who "understands the vibes."
AI needs rigid boundaries. If you give an LLM a blank check to solve problems without an operator's oversight, it will solve them with your money.