Two-way reviews, model-drafted replies

Context

The platform consumed third-party reviews via a scraping pipeline for purposes of display, but operators could not reply to those reviews from within the platform itself. Reply at scale also introduces a quality problem of its own: a generic response is worse than no response, and a templated reply that fails to address the actual complaint actively damages operator trust.

Design

• A migration off the prior scraping pipeline onto the official first-party business-profile API, restoring read-write access at the protocol level.
• A model-drafted response generator constrained by business-specific tone, brand-specific vocabulary, and the original review itself as grounding context.
• A draft-then-approve user experience. Operators retain control over what is published under their name. A clear path is reserved for automatic publication of replies to high-confidence positive reviews once a per-business trust threshold is met; everything else passes through review.

Outcome

Reviews become a two-way operator channel inside the platform rather than a passive feed, without surrendering tonal control to a generic chatbot. The same draft-then-approve pattern has subsequently been re-applied to other domains in which the cost of a bad outbound message is asymmetric with respect to the cost of a delayed one.

Stack

Go services (integration management and storage) · LLM gateway · third-party business-profile API · React administrative client.