Lyft's agentic AI cuts resolution times and brings always-on support
Lyft, working with the AWS Generative AI Innovation Center, launched an "intent agent" that uses natural conversation to understand what riders and drivers need and then takes action. The result: a 87% reduction in average resolution time and support that's available 24/7 in English and Spanish.
For customer support teams, this is a clear signal: conversational systems that tie directly into backend workflows can remove wait time, shrink call volume, and free human agents to handle the hard stuff.
What changed
The intent agent doesn't force users through menus. It asks clarifying questions, searches Lyft's systems for context (ride history, account status, previous interactions), and decides whether to resolve the issue automatically or hand it to a human specialist.
More than half of customer and driver requests now finish in under three minutes. Drivers spend less time on the phone and more time earning. As Lyft's VP of safety and customer care, Ameena Gill, put it: "This creates a positive cycle: supported drivers are more engaged and available to deliver great services."
Why it works for support leaders
- Natural dialogue replaces menu mazes, cutting friction from the first message.
- Multi-turn conversations disambiguate vague requests before resolution starts.
- Direct access to account data enables accurate, personalized answers.
- Built-in workflows handle routine fixes instantly; complex cases route to specialists with context attached.
- Round-the-clock availability reduces backlogs and protects SLAs during spikes.
How to apply this in your operation
- Pick high-volume, low-variance intents (refunds, account holds, trip issues) as your starting lane.
- Map the backend steps for each intent so the AI can actually complete the task, not just answer questions.
- Give the system scoped access to customer history and prior interactions for true personalization.
- Design clarifying prompts to resolve ambiguity in two to three turns.
- Define a clear handoff policy: confidence thresholds, red flags, and which queues receive escalations.
- Instrument everything: time to resolution, containment rate, CSAT by automation vs. human, reopens.
- Train human agents as "specialists" for edge cases, recovery, and accountability moments.
- Run weekly quality reviews on transcripts to tune prompts, workflows, and guardrails.
Metrics that tell the real story
- Average resolution time (ART) and distribution by intent.
- Containment/automation rate vs. assisted rate, with quality checks on both.
- CSAT/DSAT by channel and by handoff outcome.
- Escalation and reopen rates to spot hidden failure modes.
- Agent productivity: cases per hour for specialists and reduction in hold time.
Guardrails to keep in place
- Scoped data access with audit logs and automatic redaction for PII.
- Instant human opt-out for sensitive or high-stakes issues.
- Policy checks for refunds, safety reports, and identity-related actions.
- Continuous accuracy testing on critical intents to catch drift.
Tech note
This approach blends conversational intelligence with executable workflows and a clean escalation path. For background on the platform side, see the AWS generative AI resources.
Upskill your team
If you're building an automation-first support model, curated training by role can speed up adoption. Explore options here: AI courses by job.
Bottom line: let AI handle the repetitive, high-volume requests with context and execution. Put your best people on the nuanced cases where empathy and accountability really matter.
Your membership also unlocks:
