Uber: Applying Machine Learning to Improve the Customer Experience Custom Case Solution & Analysis

Evidence Brief

Financial Metrics

• Total trips reached 2.1 billion in 2016.
• Customer support volume scales linearly with trip growth without automation.
• COTA v1 reduced ticket resolution time by 10 percent.
• COTA v2 improved ticket classification accuracy by 6 percent compared to v1.
• The company operates in over 600 cities globally.
• Support involves 75 million riders and 3 million drivers.

Operational Facts

• Uber uses Michelangelo as a centralized platform for machine learning development and deployment.
• Customer support tickets are categorized into thousands of specific issue types.
• COTA uses Natural Language Processing to analyze ticket text and suggest solutions to human agents.
• Manual ticket classification previously required agents to navigate deep menu hierarchies.
• Deep learning models in COTA v2 utilize convolutional neural networks for text processing.

Stakeholder Positions

• Dara Khosrowshahi: Focused on operational efficiency and path to profitability.
• Barney Harford: Prioritizes scaling operations through technology rather than headcount.
• Michelangelo Engineering Team: Advocates for centralized machine learning infrastructure.
• Customer Obsession Team: Seeks to reduce agent friction and improve the experience for riders and drivers.
• Support Agents: Require tools that reduce repetitive cognitive load without replacing human judgment on complex cases.

Information Gaps

• The specific capital expenditure required to maintain the Michelangelo platform.
• The exact correlation between reduced ticket latency and rider retention rates.
• The cost per ticket for human agents versus the cost of compute for machine learning inference.
• Potential bias metrics within the training data for different languages or regions.

Strategic Analysis

Core Strategic Question

• How can Uber decouple support costs from trip volume while maintaining service quality in a highly competitive market?
• Should machine learning resources be centralized within Michelangelo or decentralized within product units?

Structural Analysis

The application of the Value Chain framework reveals that customer service is a primary activity where Uber faces a structural disadvantage due to massive scale. Human-led support creates a margin ceiling. Analysis of the competitive landscape suggests that as ride-sharing becomes a commodity, the speed of dispute resolution becomes a key differentiator. The current bottleneck is the cognitive load on agents who must manually classify thousands of issue types. Machine learning transforms this from a labor problem into a data processing problem.

Strategic Options

Option 1: Full Automated Resolution. Remove human agents from the loop for the top 20 percent of simple ticket types. This offers the highest cost savings but carries the risk of alienating users if the model misclassifies a sensitive issue.

Option 2: Augmented Intelligence (COTA Expansion). Continue the human-in-the-loop model where machine learning provides suggestions. This minimizes errors and maintains the human touch but limits the total potential for cost reduction.

Option 3: Product-Led Error Elimination. Use machine learning insights to identify the root causes of tickets and redesign the app to prevent the issues from occurring. This requires high cross-functional coordination between support and product teams.

Preliminary Recommendation

Uber should pursue Option 3 in tandem with Option 2. While COTA provides immediate operational relief, the long-term strategy must be the elimination of the need for support tickets through product improvements. Relying solely on faster responses to avoidable problems is a suboptimal use of technical resources.

Implementation Roadmap

Critical Path

The sequence begins with the integration of COTA v2 into all global Centers of Excellence. This requires a 90-day rollout to stabilize the API and ensure latency remains under 100 milliseconds for agent interfaces. Next, the data science team must tag the 50 most frequent root causes of tickets. These tags will feed back into the product engineering backlog for app redesigns by the end of the second quarter. Finally, a pilot for full automation of refund requests for canceled trips will commence in three low-risk markets.

Key Constraints

• Data Quality: Inconsistent tagging by human agents in different regions can degrade model accuracy.
• Technical Debt: Over-reliance on the Michelangelo platform may create a single point of failure for all machine learning applications.
• Agent Adoption: Resistance from support staff who fear job displacement could lead to poor utilization of the COTA suggestions.

Risk-Adjusted Implementation Strategy

To mitigate the risk of model drift, the team will implement a continuous feedback loop where agent rejections of COTA suggestions are analyzed weekly. If accuracy drops below 85 percent in any category, the system will automatically revert to manual classification for that issue type. This ensures that the user experience is never sacrificed for the sake of automation speed. Expansion into non-English speaking markets will be delayed until local language models achieve parity with the English model performance.

Executive Review and BLUF

BLUF

Uber must transition from human-dependent support to a machine learning-first approach to achieve profitability. The COTA platform has proven it can reduce resolution time by 10 percent and improve accuracy. The company should now shift focus from assisting agents to eliminating the root causes of support tickets via product redesign. Success depends on the reliability of the Michelangelo platform and the ability to maintain accuracy across diverse global markets. The math is clear: labor-based support cannot scale with 2 billion annual trips. Automation is the only path to sustainable margins.

Dangerous Assumption

The analysis assumes that ticket volume is a direct function of trip volume. If app complexity increases or new services like Uber Eats introduce higher error rates, the current machine learning models may not scale as predicted, leading to a support backlog that technology alone cannot fix.

Unaddressed Risks

Unconsidered Alternative

The team did not consider a tiered support model where premium riders receive human-only support while standard riders are moved to a fully automated interface. This would protect the high-value segment while aggressively cutting costs in the mass market.

VERDICT: APPROVED FOR LEADERSHIP REVIEW