AI Customer Support at Scale: Lessons from Uber

When your business runs tens of millions of on-demand trips a day, “customer support” stops being a department and becomes core infrastructure. Every confusing fare change, missing grocery item, or delayed delivery is a moment where a customer decides whether your product feels dependable—or chaotic.

Uber’s approach to AI-driven customer service is a strong case study for U.S. digital services: it shows what happens when a company treats AI not as a novelty chatbot, but as a system that reduces friction across a massive, real-time marketplace. In a conversation with Uber’s Jai Malkani (Head of AI and Product, Customer Obsession), the throughline is clear: AI is being used to make support faster, more accurate, and more consistent—while keeping humans in the loop where it counts.

This post is part of our “AI in Customer Service & Contact Centers” series. The goal here isn’t to recap Uber’s story—it’s to extract what’s practical: the patterns, metrics, and implementation moves that other technology and digital services teams in the U.S. can apply.

Why Uber’s AI support strategy matters for U.S. digital services

AI in customer service works best when it’s tied to operational reality. Uber sits at the intersection of the digital and physical worlds, where support isn’t just answering questions—it’s resolving disputes that involve GPS routes, tolls, photos, receipts, inventory, and human behavior.

That’s exactly why it’s relevant beyond rideshare. Many U.S. companies now operate “real-world” digital services: delivery, telehealth, field services, marketplaces, travel, fintech disputes, and subscription businesses with complex billing. The same pressure shows up everywhere:

• Volume spikes (holiday travel, weather events, promotions)
• High customer expectations (fast answers, clear resolutions)
• Policy complexity (regional rules, eligibility logic, fraud controls)
• Multi-party outcomes (customers, contractors, merchants, partners)

Uber’s stance is blunt and correct: AI isn’t optional if you want to compete on experience at scale. The companies winning in 2026 won’t be the ones that “added a chatbot.” They’ll be the ones that redesigned support workflows around automation + human judgment.

How AI resolves real customer issues (not just FAQs)

The most useful way to think about AI in a contact center is as a decision-and-evidence engine. Uber’s examples are revealing because they’re not simplistic.

Rides: handling fare changes with context and transparency

Fare disputes aren’t rare in mobility. Routes change. Traffic forces detours. Tolls appear. A good support experience depends on two things: correctly interpreting what happened and communicating it clearly.

Uber describes using AI to:

• Interpret trip events (route deviation, toll triggers, time/distance deltas)
• Determine what’s equitable for riders and drivers
• Provide transparent explanations within support interactions

The lesson: resolution quality depends on contextual data, and AI is increasingly the layer that can synthesize it quickly enough to matter.

Delivery: using photos and item “lifecycles” to pinpoint the error

In Uber Eats, the problem often isn’t “Where is my order?” It’s “Why is this wrong?” That requires evidence.

Uber highlights AI that can:

• Analyze photos (e.g., delivered items)
• Track an item’s lifecycle (handoffs, packaging steps, delivery confirmation)
• Decide whether to refund, reorder, or escalate

If you run any delivery-like operation, this is the play: connect support decisions to proof, and let AI do the first pass of triage and verification.

Grocery: real-time inventory and preference-based alternatives

Grocery is messy because the world is messy—inventory changes minute to minute. AI can reduce back-and-forth by helping shoppers and customers agree on substitutes quickly.

Uber’s approach includes:

• Facilitating real-time communication between shoppers and customers
• Suggesting next steps based on customer preferences
• Increasing satisfaction while reducing handling time

The takeaway: automation isn’t just deflection. Sometimes it’s faster coordination, which is a different kind of customer service win.

Personalization in AI customer service: helpful or creepy?

Personalization in customer support should be judged by one standard: does it reduce customer effort without surprising them?

Uber describes “hyper-personalization” using customer preferences, past transactions, and context. A concrete example: helping drivers handle document updates before expiration with guidance tailored to region and account requirements.

That kind of personalization works because it’s:

• Expected (people know documents expire)
• Relevant (regional rules differ)
• Actionable (clear next steps)

Here’s what I’ve found in other customer service AI rollouts: personalization breaks when teams chase novelty (“We know you like X!”) rather than usefulness (“Here’s what you need to do next, and why”). If you’re adopting AI in a contact center, put personalization behind customer intent:

• Billing issue → explain charges, show evidence, offer resolution options
• Account risk/fraud → clarify required verification steps, provide timeline
• Delivery problem → confirm what happened, propose next-best actions

Personalization should feel like competence, not surveillance.

AI as an agent co-pilot: the highest ROI use case

Most companies underestimate this: the fastest path to impact is often agent assist, not full automation.

Uber describes AI as an “intelligent co-pilot” used internally across teams—and especially in customer support. Examples include:

• Conversation summaries for faster handoffs and less repetition
• Automated investigations to gather evidence (trip data, order events, policy criteria)
• Empathetic next-best responses to improve tone and clarity
• Translating complex policies into actionable routines for resolutions

This is the modern contact center stack:

1. LLM summarization (reduce after-call work)
2. LLM response drafting (consistent language, fewer mistakes)
3. Workflow automation (auto-pull data, pre-fill forms)
4. Decision support (recommended actions within policy)
5. Selective automation (handle low-risk cases end-to-end)

A strong stance: if you’re choosing between building a customer-facing chatbot and building agent assist, start with agent assist. Customers notice fewer transfers, faster resolution, and clearer explanations—without you betting the brand on a fully automated front door.

What to measure: the metrics Uber uses (and what you should add)

AI initiatives die when teams can’t prove impact. Uber’s measurement approach combines experience, speed, and business metrics.

They cite signals such as:

• Customer experience scores
• Resolution speed n- Automation rates
• Productivity gains
• Quality of LLM responses
• Controlled experiments comparing AI-augmented workflows to traditional ones
• Engagement and incremental gross bookings across regions

That mix is smart because it avoids the trap of measuring only “containment rate” (how many contacts didn’t reach humans). Containment is easy to inflate and often correlates with unhappy customers.

If you’re running AI in customer service, I’d add a few metrics that tend to predict long-term success:

• Recontact rate within 7 days (did we actually solve it?)
• Escalation quality (did the AI package the right context for a human?)
• Policy compliance rate (did the recommended action stay within rules?)
• Time-to-evidence (how long to retrieve the data needed to decide?)
• Tone adherence for regulated or sensitive workflows

AI support should be measured like a product: outcomes, not activity.

A practical adoption playbook (steal this)

Uber’s advice to product leaders is direct: the debate about whether to adopt AI is over. For most U.S. digital services companies, the real question is how to adopt AI without breaking trust.

Here’s a playbook that matches what Uber describes—and what tends to work in the field.

1) Start where the pain is expensive

Pick 1–2 high-volume contact types that cost you real money or churn. Examples:

• Fare/billing disputes
• Missing/wrong items
• Account access and verification
• Refund eligibility

If it doesn’t move cost-to-serve or customer satisfaction, it’s a demo—not a deployment.

2) Build an “evidence bundle” before you generate text

Support is decisioning. Decisioning needs inputs.

Before an LLM drafts a response, your system should assemble an evidence bundle:

• User/account state
• Event logs (trip/order timeline)
• Policies that apply (region, product type)
• Risk signals (fraud, abuse patterns)

Do this well and your AI gets safer and more accurate fast.

3) Put humans where judgment matters

Not every case deserves the same automation level. A good pattern:

• Low-risk + clear evidence → automate resolution
• Medium-risk or ambiguous → AI recommends, human approves
• High-risk (fraud, safety, regulatory) → human-led, AI assists

This keeps the “AI customer support” experience fast without becoming reckless.

4) Run controlled experiments like a product team

Uber calls out controlled experiments. That’s essential.

A simple A/B you can run:

• Group A: traditional workflow
• Group B: AI summaries + AI drafted responses + auto-evidence retrieval

Measure handle time, recontacts, CSAT, refunds accuracy, and escalation rates. If Group B isn’t better in at least two customer-facing metrics, pause and fix the workflow.

Where AI customer service is heading in the on-demand economy

Uber points to autonomous vehicles and robotic delivery, plus new work categories like data annotation and sentiment analysis. That’s directionally right, but the nearer-term shift is more immediate: support is becoming predictive, not reactive.

In on-demand services, the next wave looks like this:

• Proactive outreach when a trip or delivery is likely to fail
• Real-time resolution offers before a customer contacts support
• Voice support that can authenticate, summarize, and resolve within guardrails
• More “policy-to-workflow” automation (policies converted into executable routines)

The companies that win will treat customer service like a real-time operations product—not a cost center.

What you can do next

If you’re leading a contact center, CX, or digital operations team, Uber’s case study offers a practical message: AI pays off when it’s tied to evidence, workflows, and measurable outcomes. Chatbots are the least interesting part.

For this “AI in Customer Service & Contact Centers” series, we’re tracking what actually works across U.S. technology and digital services: agent assist, smarter triage, policy automation, and better measurement.

If you had to pick one support workflow to redesign with AI in Q1, which would it be—and what evidence would your system need to resolve it confidently?