Production CX Agents: Lessons from Lyft, Vodafone, and LATAM Airlines

With Customer Experience (CX) agents entering production, the core challenge has shifted from building to operating and continuously improving. Fast responses increase conversion rates, reduce escalation and support costs, and improve issue resolution rates, thereby impacting customer retention.

Teams in the production phase treat agents as systems requiring continuous testing, deployment, monitoring, and iterative improvement. Key usage patterns include:

• Customer-facing self-service agents: Handle payments, account access, insurance claims, reservations, etc. Podium's AI Employee showed a 46% higher lead conversion rate when responding within 5 minutes of an inquiry compared to responses within 1 hour.
• Agent and field employee copilots: Cisco filters critical issues from thousands of potential findings for network engineers and suggests next steps.
• Self-service agent platforms: Lyft built a platform that allows operations teams and product managers to launch support agents using only prompts and configuration files, without assistance from machine learning engineers.
• Semantic-based routing and triage: Connects incomplete or ambiguous customer requests to the appropriate agents and workflows.

In LATAM Airlines' travel support agent Concierge, 13% of initial messages were classified as out-of-scope. However, analysis of conversations revealed that 95% of these were actual passenger needs such as check-in and baggage inquiries. After adding a dedicated customer care agent, the out-of-scope rate dropped to 1%.

Fastweb and Vodafone built Super TOBi and Super Agent to handle both customer conversations and internal call center support. As multiple teams participated in agent development, evals became crucial as common criteria for defining what constitutes a good response and determining release readiness.

After opening up agent creation to non-engineers, Lyft confirmed that prompt and evaluation quality, rather than the platform itself, were the main constraints. Consequently, they introduced a structured prompt writing framework and automated checks to identify conflicting instructions and incomplete conversation paths in advance.

These operational practices extend to continuous testing and improvement across the entire Agent Development Lifecycle using LangSmith, Deep Agents, and LangGraph. Systems like LATAM Airlines' Compass, which transform unstructured conversations into structured signals, are also utilized as a foundation for improving customer experience.