Spent years as a Charles Schwab customer without placing a single call about anything routine. The account manager calls me. Not the other way around. And when that call comes, it is sequential: the manager works through what he knows about my situation before he says anything. There is no script-read preamble. The call has already been worth taking before he gets to the reason for it.

That is a design decision, not an accident of good hiring. Someone at Schwab decided that the contact model should run proactively, not reactively. Customers who never have to chase their financial institution are not just satisfied, they are less likely to leave. The call that comes in from your account manager is doing something structurally different from the call you make when something goes wrong.

Hold time at Schwab, when it comes up at all, offers a real choice: Schwab network music or your own. No call routes to voicemail. These are small signals, but they are consistent ones. The same logic runs through the whole experience.

Amazon built the help link so you would use it

Amazon operates at a scale where most customer service interactions would break a lesser support model. The help link is buried. Finding it takes patience. But when you do find it, the resolution path is short enough that following it to the end is easier than abandoning it. And the resolution rate is high enough that customers who do contact support rarely have to contact support again about the same issue.

The outcome is trust. Not because Amazon has excellent agents, though it might. Because the system was built to close the loop rather than extend it.

Google Fi publishes the wait time before you commit

Google Fi does something that should be standard and is not: it tells you the current response time for each contact channel before you choose one. Chat, estimated response now. Phone, they call you back at a time you select. Email, here is the window. No channel is hidden behind another. No phone tree obscures the chat option.

The preference I have for chat is a channel preference, not a content preference. Fi accommodates it without requiring me to justify it. The interaction ends with a resolution. I have not had to return to the same issue twice.

The common thread across Schwab, Amazon, and Fi is not industry or scale. It is architecture. In each case, someone made the contact model visible, opted for proactive outreach where it mattered, and closed the loop before a second contact became necessary.

Uber failed at the worst possible moment

When time is a crunch and you need to go urgently, the app fails. This is not a routine inconvenience. A failed rideshare app at departure time is a genuine time constraint with real consequences.

The AI support response was a set of default answers that did not map to the specific failure I was seeing. The human agent, when I reached one, repeated what the AI had said. Word for word. The escalation path led back to the same information the deflection layer had already served. There was no mechanism to say: this is time-sensitive, this is an app failure, treat it differently.

I switched to Lyft. The app worked. I made the flight.

The Uber failure was not a training failure. It was an architecture failure. The system was built to deflect, not to recognize when deflection was the wrong response. At the moment the customer's stakes were highest, the service model had no way to change its behavior.

Deflection is not the same metric as prevention

The CX platform industry has spent a decade optimizing for deflection: automated resolution, reduced handle time, lower cost per contact. These are real operational metrics. They are also incomplete ones.

Deflection measures how well a system manages a contact that has already happened. Prevention measures whether the contact needed to happen at all. The companies that score highest on customer trust are not the ones with the fastest resolution times. They are the ones with the lowest unnecessary contact rates, because they built the service experience to make problems visible and solvable before a customer decides to reach out.

I have covered the CX platform market closely over the past year, including Verint's Engage 2026 conference, Cisco's Webex Contact Center announcements, Zoom CX crossing $100 million in ARR, and AWS acquiring NLX to make Amazon Connect accessible to non-engineers. The vendors in this space are investing in agent AI, workflow automation, and real-time analytics. That investment is not misplaced. But it is downstream of the architecture question.

A contact center that resolves calls faster is a better contact center. A service organization that stops unnecessary calls from arriving is a different category of operation. The platform spending follows the second goal poorly and the first goal well.

One more tell. In 2026, any company still sending emails with a "do not reply to this email" footer has already answered the architecture question. That line is not a technical limitation. It is a policy choice that says: we will initiate contact on our terms and close the channel on our terms. Calling that an AI-powered customer experience is a category error.