Uber boosts driver earnings and dispatch speed with OpenAI

Uber is using OpenAI's frontier models and the Realtime API to build the driver-facing Uber Assistant and in-app voice experiences. On top of a real-time marketplace that handles 40 million trips a day, 10 million drivers and couriers, 15,000 cities, and 70+ countries, concurrent rides worldwide reach 1.7 million.

Uber Assistant answers driver questions like where to position themselves right now, whether the airport is worth it, whether to focus on rides or deliveries during lunch hours, and why today's earnings differ. It helps new drivers ramp up faster, while experienced drivers also return to it repeatedly to optimize their time and earnings.

To meet the bar for accuracy, safety, reliability, and latency, Uber built a multi-agent architecture that splits requests across specialized systems.

• Nano/mini models handle lightweight classification and fast responses.
• Larger reasoning models handle complex tasks.
• An internal governance layer called AI Guard reviews prompts and responses to uphold policy, privacy, and security, and to reduce hallucinations.

The voice feature starts from the microphone icon in the app's where to search bar. Users can speak complex requests in natural language, such as an airport trip with a lot of luggage, and the system uses saved locations and customer context to suggest options like UberXL or destinations like home, matching voice and on-screen responses together. Drivers can operate the app hands-free, and accessibility improves for blind or elderly users as well.

The way development happens has also changed. Engineers work on prompting, retrieval systems, evaluation pipelines, and orchestration frameworks, while product, legal, operations, and design teams jointly define policy boundaries and output quality. Rather than a single centralized AI team, the structure now allows many teams to build model-based features, and hundreds of thousands of U.S. drivers are using the Uber Assistant beta.

• Early-stage drivers ramp up faster.
• Strong repeat usage is observed.
• Time utilization on the platform improves.
• Model specialization and continuous evaluation speed up product iteration.