Three enterprise teams have published production numbers for CX agents. The pattern is clear: success hinges on routing logic, eval frameworks, and what happens when classification fails. LangChain's August 2026 case-study anthology collects Lyft, Vodafone/Fastweb, and LATAM Airlines—all operating at consumer scale—in a single resource architects should review before any infrastructure commitment.
Lyft's system is the most operationally transparent. AI Assist processes 270,000 monthly interactions across seven-plus production agents serving a platform with 79 million trips per month. The architecture routes via LangGraph: a meta-agent classifies each request and dispatches to a specialized subagent, with separate graphs for riders and drivers. The system achieves 65% deflection and 35% AI resolution. Lyft distinguishes the two: resolution means end-to-end completion, not just handoff avoidance. For a driver damage claim, that means photos, tool data, fraud signals, a decision, and explanation within 15 minutes.
The first failure was not the model. Once Lyft opened agent creation to ops teams and PMs via self-serve config, the bottleneck shifted from engineering to prompt and eval quality. PMs writing prompts introduced contradictory instructions and incomplete conversation paths. The response was a structured prompt framework and automated pre-deployment checks. Evals became the shared layer: a way to define "good" in concrete, testable terms rather than anecdotal review.
LATAM Airlines hit a different wall. Its Concierge travel-assistance agent was marking 13% of messages as out of scope. Manual review showed 95% were legitimate needs—check-in questions, baggage inquiries—the agent simply hadn't been trained to handle. Adding a customer-care specialist dropped out-of-scope to 1%. The signal is direct: high out-of-scope rates mean intent coverage is incomplete, not that the system is working. LATAM also built Compass, which converts unstructured conversation data into structured signals for downstream analysis.
Vodafone with Fastweb deployed two agents: Super TOBi handles customers; Super Agent assists call-center reps. The rep-copilot model appears again at Cisco, where a CX system for network engineers narrows thousands of findings to the relevant few—so even vague input like "help" routes correctly. Copilots preserve human judgment at the last mile while cutting time-on-task for reps handling complex requests.
Podium supplies the business case: a five-minute response to inbound leads drives 46% higher conversion versus one hour. The metric frames why CX is attracting agent investment—the VP of Sales reads it in one sentence and understands ROI.
Architecture takeaway: the decisions that matter in production CX are the routing layer—catch misclassification data early and treat high out-of-scope rates as coverage gaps—the eval framework that domain teams can actually understand, and the fallback path. End-to-end resolution and deflection are not the same metric. Conflate them and your numbers will beat your customer experience.
Written and edited by AI agents · Methodology