The mechanism is three steps — classify, draft, route. A classifier scores each review for sentiment and severity; the agent drafts a reply in the right voice — Eleanor for food, Tommo for drink; a routing rule decides whether a human must see it first. Praise and mild service notes become one-click drafts. Allergen, legal and staff complaints are never drafted and never auto-posted. Everything passes a human gate before it goes live.
On the roadmap
- Close the loop with pricing. Aggregate sentiment per dish, not just per review. A plate that's consistently adored is one the menu-engineering tool can reprice with confidence; one that's quietly disappointing is a reformulate-or-cut signal. The reviews are a pricing input, not just a reputation chore.
- Surface the patterns. One slow-dessert review is noise; the same complaint six times in a fortnight is a fixable process problem. Trend detection turns a queue of replies into a short operations brief — dessert timing, Friday noise, whatever keeps recurring.
- Answer in the guest's language. Drafts in English, Mandarin, Bahasa Melayu and Tamil, matched to the language of the review.
- More than two sources. Google leads in Singapore, but the same triage and approval gate should sit behind Tripadvisor and Facebook feeds too.
- Everything logged. Every classification, draft and escalation decision is recorded for the quality layer — so the safety net (the allergen review that never auto-posted) is auditable, not merely asserted.