Zepto Details AI Support System, Says Support Costs Fell 65%

Zepto says its AI customer-support system now handles more than 100,000 tickets a day, with agents fully managing more than 80% under human oversight. Its answer to the risk of automating service at that scale is not a single model upgrade: it is a quality gate that tests agent changes before release and sends live failures back into development.

The system, disclosed Tuesday in a Databricks case study, is a multi-agent setup built with Databricks and MLflow. Zepto and Databricks describe it as evaluation-first: teams define what acceptable performance looks like, evaluate agent versions against that standard, and monitor the same standards after deployment.

The operational problem is hidden failures

A support agent does more than write a response. It can classify a request, retrieve information, analyze inputs, make a decision and call transactional tools. Zepto’s framework traces those steps—including prompts, model completions, retrieved documents, tool calls, latency and decision paths—rather than judging only the final customer-facing answer.

A dataset becomes the release checklist

Before an agent version can move forward, Zepto evaluates it against a “golden dataset”: annotated examples covering ordinary requests, edge cases and failures. Its evaluation stack combines built-in and custom large-language-model judges, which score work requiring human-like judgment, with code-based scorers for deterministic checks such as tool latency or call counts.

The framework’s core feedback loop is deliberately circular. Production failures are captured and added to future development evaluation, while a quality gate decides whether a revised agent version is deployed or returned for more work. MLflow says its tooling can reuse judges and scorers from development in production monitoring; Zepto’s case study shows that pattern applied to a customer-support workflow.

80%+Tickets fully managed by AI

Zepto reports that its agents fully manage more than 80% of support tickets with human oversight.

65%Support-cost reduction

Zepto reports a 65% reduction in support costs and payback in less than one month.

8 pointsDevelopment-to-production accuracy gap

Over six months, Zepto says it expanded its evaluation dataset from 500 to 5,247 examples while reducing this accuracy gap.

The results are promising, but company-reported

Zepto reports that the framework reduced support costs by 65%, improved customer satisfaction by 20%, raised accuracy by 8%, and shortened time to resolution fourfold. It also says development cycles became three times faster and that the investment paid back in under a month.

Those figures come from Zepto’s account in a Databricks customer story, not from an independent comparison or audit. The disclosure does not establish how the metrics were measured, which tickets remained under human handling, or how performance varied across different support categories. What it does show is a concrete operating choice: treat evaluation data, trace records and release thresholds as production infrastructure, alongside the agents themselves.