Two Ways to Deploy AI in Customer Service. One Works Better.

Uber cut 10% of its global customer support organization in July 2026. Bank of America announced the same month that it had upgraded the AI tools used by more than 18,000 customer service representatives, helping each agent resolve calls nearly a minute faster.

Same moment. Same technology. Opposite philosophies.

This comparison is worth studying because it is not a debate about whether to use AI. Both companies are all-in. The question is what you use it for, and who bears the cost of getting it wrong.

Uber: Restructure First, Scale Second

Uber’s internal messaging from VP of Global Community Operations Megha Yethadka was direct: “We cannot scale frontier technology on top of fragmented processes.”

That is a correct observation. AI layered on top of broken workflows produces faster broken workflows. The restructuring logic is sound.

What is less clear is whether the sequencing is right. Uber already automates roughly two-thirds of its half-billion annual customer inquiries with zero human involvement. Agent-assist tools have reduced handling time by about 25% on the issue types where they have been deployed. The system is working. The cuts are designed to simplify the organization so AI can scale further, not because the current headcount has run out of work.

The risk: the remaining 150 million-plus contacts that require human involvement are, by definition, the hard ones. The disputes, the safety issues, the frustrated customers who have already tried the bot and failed. Those interactions do not get easier with fewer agents.

Public reviews and complaint data going into 2026 consistently flag the same problems: chatbot loops, template responses, difficulty escalating to a human, low resolution rates on anything complicated. Analyst commentary around the July cuts noted that customer service is already a frequent friction point for Uber, and that the impact of further reductions will take time to assess.

Bank of America: Augment First, Always

Bank of America’s Erica launched in 2018. It has now handled more than 3 billion client interactions. The self-service success rate runs near 98-99% for users who engage with it. That is not an experiment. That is infrastructure.

The July 2026 EricaAssist upgrade extended generative AI to the human agents themselves. When a client calls, the agent sees a real-time summary of why the client is calling, relevant account context, and recommended next steps, all delivered in under three seconds. Average call time dropped by nearly one minute per interaction. Agents spend less time searching for information and more time actually helping.

Bank of America CEO Brian Moynihan has been clear that AI is reducing headcount over time, primarily through attrition and selective non-replacement rather than targeted layoffs. Roughly 1,000 fewer roles in one recent quarter. The pace is deliberate.

This is the “high tech, high touch” philosophy in practice: AI handles the volume, humans handle the relationship, and the two work together on everything in between.

What the Research Says

The evidence on this is not ambiguous.

A large-scale empirical study published in the Quarterly Journal of Economics on generative AI deployed in a customer service setting found that agent-assist tools raised productivity, particularly for newer and lower-skilled agents, without degrading customer satisfaction scores. The AI improved human performance rather than replacing it.

Meta-analyses across customer service AI deployments consistently show: pure AI replacement reduces positive customer emotion and intention, especially in complex, high-stakes, or emotionally charged interactions. Hybrid configurations that keep humans in the loop preserve quality while capturing efficiency gains.

The pattern is consistent. Augmentation protects satisfaction. Replacement creates risk, particularly at the tail end of the contact distribution where the hard problems live.

What This Means for How You Deploy AI

If you are building an AI strategy for customer-facing operations, the Uber/BoA comparison offers a useful forcing question: are you using AI to eliminate human judgment, or to make human judgment faster and better?

Both approaches deliver cost and efficiency gains. The difference shows up in the interactions that matter most: the frustrated customer, the complex dispute, the moment where someone needs to feel heard before they will accept a resolution.

Three principles that hold up across both cases:

Fix the process before you layer in the AI. Uber’s instinct to simplify fragmented processes before scaling AI is right. AI amplifies what is already there. Broken workflows at scale are more expensive than broken workflows with humans.

Protect the escalation path. The value of automation depends partly on the quality of the handoff when automation fails. If customers cannot reach a knowledgeable human when the bot fails them, your automation rate becomes a liability.

Measure what the AI does to the hard cases, not just the easy ones. Deflection rates and handling time are easy to track. Resolution quality on complex issues and emotional satisfaction scores are harder, but they are where the real strategic risk lives.

Bank of America has nearly a decade of iterative feedback loops built into Erica. That institutional knowledge compounds. Uber’s more aggressive restructuring may close the gap if escalation paths become seamless and remaining agents become more effective. Current public signals do not yet show that outcome.

The cleaner strategic model, based on available evidence, is the one that makes humans better rather than replacing them. The efficiency is real either way. The difference is in what you leave behind.