OTTO Builds a Native Voice Shopping AI Assistant on Google Cloud's Gemini

Overview

German online retailer OTTO built a native voice shopping assistant on Google Cloud's Gemini Enterprise Agent Platform, using the Gemini Live API with native audio and Chirp 3 speech recognition to deliver conversational product advice across roughly 19 million products. Becoming the first EMEA retailer with native voice shopping in production, OTTO cut end-to-end voice response latency from 8-9 seconds to under 2 seconds and scaled expert product advisory coverage from 5 to 50 categories.

The challenge

Search answers in lists, so a customer typing a broad query like "washing machine" gets thousands of results to filter through, but a customer wanting advice needs a back-and-forth dialogue, follow-up questions, and a recommendation that explains itself — something search alone cannot deliver.

The solution

OTTO built a native voice shopping assistant on Google Cloud's Gemini Enterprise Agent Platform. After an interim cascading speech-to-text/text-to-speech layer failed to create real engagement in user tests, OTTO rebuilt the voice path on the Gemini Live API with native audio: a text assistant drives guardrails, recommendations and search while a live audio model handles the real-time conversation, with Chirp 3 handling German speech recognition. OTTO's engineers built a custom orchestration layer drawn from formal methods to make provable guarantees about conversation behaviour, and imprecise requests are treated as a retrieval-and-validation problem where candidates are re-evaluated for fit by Gemini 2.5 Flash before reaching the customer.

Reported business value

OTTO became the first EMEA retailer with a native voice shopping assistant live in production, covering roughly 19 million products. End-to-end latency dropped from 8–9 seconds in early prototypes to under two seconds, and expert product advisory coverage scaled from 5 hand-crafted categories to 50, with the next wave covering the full catalog in progress.