The reviews said "portions got smaller" for three months before anyone connected it to the new supplier.

The situation

Feedback everywhere, intelligence nowhere

Each property manager checks their own Google reviews when they remember to. TripAdvisor gets attention when a bad review stings. Booking.com feedback lives in the extranet nobody exports. The in-house comment cards get typed into a spreadsheet quarterly, if the front office has time. Nobody sees the whole picture, so recurring themes — a specific room category's AC, a signature dish that changed, slow weekend check-ins — surface only when they've done visible damage to ratings.

On the margin side, the F&B director changed suppliers twice this year and re-costed the menu once. Portion and ingredient changes ripple into guest perception with a lag, and by the time a rating dips, the causal thread back to "we switched the seafood supplier in March" is gone.

The group doesn't want an AI that auto-replies to reviews — they've seen competitors' generic AI responses and hate them. They want to know what the reviews are saying, property by property, dish by dish, week by week, and whether operational changes correlate with sentiment shifts.

The work

What Taz builds for this group

Taz designs and prototypes a bounded review-intelligence workflow — ingestion through weekly digest — with every categorization rule visible and every output reviewed by a human before it drives action.

• Ingestion design: A repeatable path for pulling reviews from all four platforms plus the internal forms into one normalized record format — review text, rating, property, date, stay type, language — using platform exports and approved access, not scraping that violates terms.
• Categorization rubric: A written taxonomy built from the group's actual reviews — rooms, F&B, service speed, cleanliness, value perception, specific dishes and room categories — so the AI tags reviews against categories the operators recognize, and the rubric itself is a reviewable document the group can amend.
• Sentiment and theme extraction: Each review gets category tags, sentiment per category (not one blended score), and quoted evidence. A review praising the room but flagging the restaurant counts in both ledgers, with the exact phrases preserved.
• Margin-drift correlation: Menu changes, supplier switches, and re-costings are logged as dated events. The weekly analysis overlays sentiment trends per dish and category against those events, flagging correlations for human judgment — the workflow surfaces the pattern; the F&B director decides what it means.
• Weekly digest artifact: One document per week: top themes per property, movers since last week, flagged correlations, and the specific reviews behind each claim. Every number traces to source reviews.

The deliverable

What the group receives

• Working prototype pipeline — ingestion, normalization, tagging, and digest generation running on the group's real review data, handed over with setup and run instructions.
• Categorization rubric document — the taxonomy with definitions and examples, versioned so the group can evolve it without breaking historical comparability.
• Event log template — the structured format for recording menu, pricing, and supplier changes so correlations stay possible.
• Weekly digest template with two live examples — real digests produced from the group's data during the engagement, demonstrating the format on actual reviews.
• Operating guide — who runs it, weekly time cost, how to amend the rubric, and what the workflow deliberately does not do (auto-reply, auto-publish, guest profiling).