Uber found that rule-based AI agents failed because their internal policy documentation was incomplete and designed for human interpretation. Their new approach scraps the rules and instead provides the AI with desired outcomes (e.g., "keep this customer happy"), letting the model determine the best action.

The most significant weakness of a multi-component model isn't price sensitivity, but the deep customer resentment it fosters. This reputational damage is difficult to quantify on a balance sheet but leads to long-term customer churn and incentivizes users to find alternatives.

Government campaigns asking for better public behavior, such as in air travel, are pointless when the underlying system is fundamentally broken. Passenger rage is a rational response to systemic failures like shrinking seats, chronic delays, and rolled-back consumer protections. Fixing the system, not lecturing the user, is the only real solution.

When customers blame your product for external failures you can't control (e.g., an SMS isn't delivered), don't dismiss the feedback. This often signals a need for better error handling or resilience. Use it as a prompt to build fallback mechanisms or better user notifications, thereby improving the overall experience.

Gamification backfires when it rewards unintended actions. For example, when Visual Studio's badge system inadvertently incentivized developers to write curse words in code comments. This shows the need to understand the second-order effects of any incentive system before implementation.

Left to interact, AI agents can amplify each other's states to absurd extremes. A minor problem like a missed customer refund can escalate through a feedback loop into a crisis described with nonsensical, apocalyptic language like "empire nuclear payment authority" and "apocalypse task."

An attempt to use AI to assist human customer service agents backfired, as agents mistrusted the AI's recommendations and did double the work. The solution was to give AI full control over low-stakes issues, allowing it to learn and improve without creating inefficiency for human counterparts.

When gathering direct customer feedback, it's easy to over-anchor on a single negative comment. Founders must implement a disciplined process to collect all feedback and analyze it for recurring themes. This prevents making reactive changes based on one-off opinions versus addressing true patterns.

Counterintuitively, Uber's AI customer service systems produced better results when given general guidance like "treat your customers well" instead of a rigid, rules-based framework. This suggests that for complex, human-centric tasks, empowering models with common-sense objectives is more effective than micromanagement.