I reserved a ride from Dulles to home. The reservation tracked the flight and carried a premium for predictability. At the terminal I marked myself ready. The app showed the driver about fourteen minutes out.

I messaged. No reply. I called. No answer. After thirty minutes I contacted support. The trip never closed. Hours later the app still showed a driver arriving or already arrived, even after I had reached home another way. New drivers kept being matched to a pickup that no longer existed.

None of this needed a model. It needed ordinary rules:

• If the rider is ready and the driver makes no progress toward the pin past a fixed window, reassign.
• If the driver does not respond after the ready signal, reassign or release the trip.
• If the rider leaves the airport geofence and the trip stays open past a hard timeout, close it.
• Keep a working cancel path even when the status display is wrong.

These are checks on location, time, and messages. They require a state machine that cannot freeze, not an agent or a reasoning chain.

Support added steps instead of removing them. There was no direct escalation. Reaching a human meant backing out and starting earlier in the flow. The AI layer was a short sequence of fixed prompts, not open conversation. It took roughly fifteen minutes, then passed the case to two human agents who gave generic replies. Only an escalation on X produced a useful response.

The total was two AI passes, two human contacts, and a public escalation. One direct path to someone who could read the trip logs would have been shorter. The AI layer increased the work required to reach a resolution.

The cost did not stay with the rider. Engineering can report token usage and model activity. Those numbers are easy to show and are not automatically tied to resolution time, number of touches, or satisfaction. When the trip freezes and support multiplies contacts, customer success absorbs the load: repeated hand-offs, generic replies, and escalations that should not have been necessary.

Vendors often note that Uber uses their software. In this case the relevant metric is likely call deflection: the interaction was handled by prompts or a first-line agent, so it counts as deflected. There was no resolution. The trip stayed open and further contacts were still required. Deflection without resolution transfers work to the customer and to the next human. When dashboards treat “handled” as the outcome, they record activity while the underlying state remains broken.

An engineering manager who cannot show that their work reduced resolution time, reduced touches, or improved outcomes on the cases their system creates is not managing the system. They are managing a dashboard that excludes the people who clean up after it.

Uber Reserve is sold on flight tracking, locked pricing, and recovery when something goes wrong. When the trip-state rules and the escalation path are missing, the premium buys a more expensive version of the same failure. The system already had the rider’s location, the driver’s location, the original ETA, and the lack of response. It did not act on them. The metrics did not surface the cost.

Rules engines, clear escalation paths, and metrics tied to actual resolution are not glamorous. They make higher-level AI useful instead of costly. Without them, each added layer of intelligence becomes another place the system can lose the thread and another set of steps the customer and support team must complete.

The test is simple. For any workflow with a bounded wait, a location, a status change, or a support hand-off, ask whether a clear rule or direct path already exists. If it does and it is missing, an AI layer will usually lengthen the path to resolution and leave customer success carrying the difference. Any metric that counts an interaction as handled while the underlying problem remains open is measuring the wrong thing.

The systems that still work under stress keep their deterministic foundations, their escalation paths, and their outcome metrics intact. The ride that never ended, the support sequence that multiplied contacts, and the deflection rate that recorded a handling without a resolution show what happens when those foundations are absent.