Delivery and mobility / Ride-hailing
Lyft: Running a platform that lets frontline staff build customer support AI agents for riders and drivers
- Company
- Lyft
- Country
- United States
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
According to a post Lyft's engineering team wrote on the LangChain blog on May 27, 2026, Lyft's AI Assist handles rider and driver support requests such as account access, damage claims, charge reviews and earnings disputes. After it began in 2023, every agent required months of work by machine learning engineers, and as demand grew in 2026 with new user segments, new issue types and autonomous vehicle support, a setup in which the frontline staff who best understood customer issues had to go through engineers to change an agent became a bottleneck.
Technology and data
Lyft built a router-based multi-agent architecture with LangGraph. A top-level agent classifies incoming requests and hands them to specialized subagents for riders, drivers and damage claims, and at every turn malicious intent detection and safety issue detection run in parallel before any LLM reasoning. Ops teams, voice of customer leads and product managers define new agents using only prompts written to a set template and JSON configuration, while engineers hand-build high-stakes workflows such as damage claims that need image processing and fraud detection. New agents first serve only 5% to 10% of real traffic, every run is traced in LangSmith, and samples of real conversations are scored automatically by LLM judges before the rollout widens.
Results
According to the Lyft team, development that took about 6 months for the first driver agent fell to about 2 weeks for new configurable agents, and 100% of production agents now have automated evaluations that score real conversations. Guardrails based on the evaluation metrics cut hallucination and contradiction rates by 20%, and the AI resolution rate rose 16% after a few agents were launched on the self-serve platform. Many non-engineering team members now build and refine agents themselves.
Limits and open questions
The baseline for the AI resolution rate, the number of requests handled and changes in cost are not in the source.
Sources
Compiled from public sources. These are not results from ATF Works customers.