AI-drafted review replies, in the house tone
The problem
Review replies written by hand, one by one, across languages.
The solution
An AI pipeline drafts replies in the house tone: a human reviews and publishes.


The client came from a DIY attempt with an unconstrained autonomous agent, with no control over costs or permissions: a governed approach was needed. Before automating anything we analysed the full historical record of reviews and the owner's replies: from those, Claude distilled a tone and style guide (register, opening and closing formulas, complaint handling), validated by the client on sample drafts before the pipeline went live.
Every morning the pipeline checks the public page for new reviews, discards the ones already seen and drafts a reply for each in the review's own language, with an Italian translation alongside. Sensitive cases, such as very low ratings or concrete complaints, are highlighted and never published blindly.
The operator finds everything in a shared worksheet: they review the draft, publish it via the platform's management console and can leave feedback on every reply.
Once a week a second process re-reads the operator's feedback and the prompt in use: when needed, it rewrites a consolidated version, archives the previous one and reports the differences. The system improves itself, with a human in the loop at all times.