Zepto Adopts Evaluation-First Multi-Agent Systems to Scale Customer Support

Zepto, an Indian quick-commerce platform operating in over 60 cities, uses a multi-agent artificial intelligence system to manage customer support. [1] The platform processes more than 100,000 support tickets daily. [1] Customer behavior shifts, expansion into new categories like electronics and beauty, and seasonal events like Diwali drive spikes in ticket volume. [1] At this scale, a one percent error rate results in thousands of bad outcomes and actual revenue leakage each day. [1]

To maintain reliability, Zepto partnered with Databricks to adopt an evaluation-first approach to building, testing, and operating its agents. [1] The company identified an "assurance gap" in its multi-step workflows, noting that failures can occur during intent classification, knowledge retrieval, or tool calling rather than just in the final generated response. [1] Before implementing this evaluation framework, failures remained invisible until customers complained, internal errors were hidden by final answers, and fixes were slow. [1]

The transition from simply shipping agents to adopting an evaluation-first methodology highlights a critical maturation point for enterprise AI operations. When multi-agent systems reach the scale of hundreds of thousands of daily interactions, traditional end-point monitoring fails to capture intermediate logic errors in reasoning or retrieval. Zepto’s reliance on Databricks and MLflow indicates that infrastructure tooling is shifting from basic deployment capabilities toward robust, step-by-step observability. As platforms expand across languages and product categories, managing the assurance gap becomes a primary operational requirement rather than an edge case. What the sources don't address: How the evaluation-first framework specifically impacts the latency of customer support responses during peak volume events.