Uber Assistant, powered by OpenAI, makes earning and booking smarter!

Hey everyone, it's Shiichan! Today I'm bringing you a story about Uber teaming up with OpenAI. I got so curious about how frontier models work behind that ride-hailing app you tap every day.

OpenAI News openai.com

What was announced?

This is a customer story published on OpenAI's News. It's about how Uber, one of the world's largest ride-hailing and delivery platforms, is building AI features for drivers and riders using OpenAI's frontier models.

Uber's scale is wild: 40 million trips a day, and 10 million drivers and couriers across 15,000 cities in more than 70 countries. Traffic, weather, airport arrivals, local events, demand... Uber wants to read that tangle of real-time signals faster, and LLMs are how it's getting there.

Why it matters

Uber has leaned on machine learning for a long time, but adding large language models and OpenAI's frontier models unlocked things that used to feel out of reach: reasoning across complex signals, fast conversational responses, and voice inside the app.

What's just as big is who gets to build. Instead of one specialized team owning innovation, engineers all over the company now work with prompting, retrieval, evaluation pipelines, and orchestration, collaborating closely with legal, operations, and design teams.

What changes

The first result is Uber Assistant, an AI helper that supports drivers through their whole lifecycle, from onboarding and first trips to day-to-day earnings optimization. It turns complex data like earnings trends and heatmaps into simple, actionable guidance on where and when to earn, and you can ask follow-up questions in plain language.

The best part is how much faster new drivers ramp up. Instead of learning marketplace dynamics through hundreds of trips of trial and error, they get AI summaries of Uber's real-world data. And even though it was expected to help newcomers most, experienced drivers keep coming back to optimize their time, so it's a long-term utility, not just an onboarding tool.

Dive Deep

Uber Assistant is built around three principles: safety, trust, and low latency. At its core is a multi-agent architecture that routes each request to the most suitable specialized system, so earnings questions are handled separately from onboarding ones. Uber uses faster nano/mini models for lightweight classification and larger reasoning models for complex tasks.

Safety is handled by AI Guard, an internal governance layer that screens prompts and responses to protect safety, privacy, and security, enforce policy, and cut down on hallucinations. Trust really matters here, and Parikh puts it this way:

If users don't trust the system, you lose them quickly. But when they see value, they return.

The other highlight is voice. Uber uses OpenAI's Realtime API so you can tap the microphone on the "where to" search bar and request a ride in natural speech. The system interprets intent, draws on your saved places and context, and makes recommendations, like suggesting UberXL for luggage-heavy trips or recognizing a saved "home".

Voice removes the multi-tap barrier because you can say multiple things.

Voice widens accessibility too: drivers can go hands-free, and riders get simpler interactions. Just note that Voice Booking is rolling out over the coming weeks.

Wrap-up

• Uber is building the Uber Assistant driver helper and voice experiences on OpenAI's frontier models
• A multi-agent setup routes tasks to the right model, mixing nano/mini with larger reasoning models
• An internal governance layer called AI Guard handles safety, privacy, and hallucination control
• Voice booking via OpenAI's Realtime API rolls out over the coming weeks, and the Uber Assistant beta is expanding to hundreds of thousands of U.S. drivers

If you want a real-world example of how to fit AI into a large, real-time marketplace, with multi-agent routing and model specialization, this one is for you.