Guest review triage

A scheduled trigger fetches new reviews from booking platforms and Google every hour. Each review is classified with structured output into sentiment, urgency, a fixed topic list and the department responsible. Urgent issues (safety, hygiene, discrimination) alert the duty manager immediately, all reviews are stored with tags for dashboards, and the model drafts a reply in the hotel's voice for a manager to approve before posting.

Thank the guest, address their specific points, apologise sincerely for real problems without admitting liability, mention concrete fixes where true, invite them to contact the hotel directly for unresolved issues, keep the brand voice and avoid repeating generic templates.

Aggregate topics by property, department and week, track sentiment trends, highlight rising complaint topics (for example Wi-Fi or breakfast), link to specific reviews as evidence, and share weekly reports with department heads.

Combine keyword rules for critical terms (bed bugs, injury, theft, discrimination) with model classification, prefer false alarms over misses for safety topics, and review urgent tags weekly to tune thresholds.

Store each review's platform ID and status (new, drafted, approved, posted), skip already processed IDs, and post replies idempotently using that ID.

Flag suspected fake reviews (no matching booking, unusual patterns) and abusive content for the manager, who can report them to the platform under its policies rather than replying in kind.

Use official APIs (Google Business Profile, booking platform partner APIs) or the review aggregator your hotel already uses, respecting rate limits and terms, and handle missing APIs with manual export.

Response rate and response time to reviews, rating trends, time saved per manager, share of drafts approved without edits, and resolution of recurring issues.

Provide a style guide and example replies in the prompt, include property-specific details (names of amenities, manager sign-off), and review a sample of drafts regularly.

Never include booking details, room numbers or personal data in public replies, and move sensitive conversations to private channels.

Store the platform, review ID, property, date, rating, text, language, detected topics, sentiment, urgency, assigned department, draft reply, approval status and posted reply ID. The review ID plus platform is the unique key that drives deduplication and status tracking.

Count tagged topics per property per week, compare with the property's own baseline, alert when a topic rises well above normal, and include sample reviews so managers see what guests are saying.

Public replies speak for the brand; a wrong apology, promise or tone can make things worse or create legal risk. Drafting saves most of the time, and a quick approval keeps quality and accountability.

Use a fixed list of topics in the classifier's structured output, review and change it deliberately, and evaluate the classifier on a labelled sample. Consistent tags make week-on-week comparisons meaningful.

Use a cron job or scheduler that calls the flow's API every hour with the time window since the last run, storing the checkpoint so no review is missed or duplicated.

Map topics to departments (housekeeping, front desk, F&B, maintenance), send each department a digest of its reviews, and create tasks for actionable issues.

Label a sample of past reviews with topics and urgency, compare the flow's tags, measure precision and recall per topic, and review confusing cases with managers.

A fixed schema with enums (sentiment, urgency, topics) keeps tags consistent for dashboards and routing, and validation catches invalid values.

Allow multiple topics with their own sentiment, reply acknowledging both, and route each complaint topic to its department.

Post the edited version, store both for comparison, and use frequent edits to improve the prompt and examples.

Tag reviews by property, use property-specific details and sign-offs, route alerts to each property's duty manager and provide group-level dashboards.

Respect character limits and posting rules per platform, adapt length automatically, and queue posts within rate limits.

Route praise to the staff member's manager for recognition, route complaints privately to the manager, and never name staff in public replies beyond what the hotel's policy allows.

Draft a calm, factual reply that thanks the guest, corrects the key point without arguing, and invites them to contact the hotel directly. Escalate to management before posting, and use the platform's dispute process only for policy violations.

When a review reports a fault such as broken air conditioning, create a maintenance ticket with the room number if stated, link it to the review, and mark the review resolved only when the ticket closes.

Normalise for review volume and platform rating scales, compare topics rather than only average ratings, and show trends over time with confidence ranges for properties with few reviews.

Detect the language, classify in a multilingual model or translate first, route urgent issues the same way, and draft the reply in the reviewer's language for a manager who can check it.

Cleanliness, staff friendliness, check-in speed, room comfort, noise, Wi-Fi, breakfast, location, value for money and amenities such as the pool or gym.

Within hours, ideally the same shift, with an alert to the duty manager and a documented follow-up action.

Monthly summaries of rating trends, top issues fixed, response metrics and comparisons across properties.

Public replies affect reputation and may have legal implications; managers catch tone problems, factual errors and sensitive situations.

Detect the language, draft the reply in the same language using an approved style guide for that language, and show the manager a translation for approval before posting.

1. AOld reviews
2. BSafety, hygiene or discrimination issuesCorrect answer
3. CPraise for breakfast
4. DNeutral comments

Why: Urgent issues need action now.

1. ALarger prompts
2. BDeleting reviews
3. CTracking platform IDs and statusCorrect answer
4. DReplying faster

Why: Posting is idempotent by review ID.

1. AThe hotel name
2. BBooking details or personal dataCorrect answer
3. CThanks
4. DAn apology

Why: Protect guest privacy.

1. AFalse alarms over missesCorrect answer
2. BMisses over false alarms
3. CNeither matters
4. DRandom errors

Why: Missing a safety issue is worse.

1. AManual tagging only
2. BRandom labels
3. CStructured output with enumsCorrect answer
4. DFree-text tags

Why: Validation catches invalid values.

1. AFlagged for managers to report to the platformCorrect answer
2. BReplied to angrily
3. CDeleted by the bot
4. DIgnored silently

Why: Follow platform policies.

1. AServer count
2. BShare approved without editsCorrect answer
3. CNumber of reviews
4. DPrompt length

Why: Plus manager time saved.

1. ANothing
2. BTopic trends by property and departmentCorrect answer
3. CSingle reviews only
4. DStar counts only

Why: Trends spot rising issues early.

1. AStyle guide and example replies in the promptCorrect answer
2. BHigher temperature
3. CShorter replies
4. DDifferent prompts each time

Why: Review samples regularly.

1. APlatform plus review IDCorrect answer
2. BThe guest's name
3. CThe rating
4. DThe review date

Why: It drives deduplication and tracking.

1. AAverage rating only
2. BThe number of rooms
3. CComparing weekly topic counts with the property's baselineCorrect answer
4. DReading every review manually

Why: Baselines reveal unusual spikes.

1. AReviews longer than 100 words
2. BAll five-star reviews
3. CSafety, hygiene and billing issuesCorrect answer
4. DReviews mentioning breakfast

Why: Those need fast action and carry the most risk.

1. ATo translate them
2. BTo hide complaints
3. CSo trends can be counted consistentlyCorrect answer
4. DTo speed up posting

Why: A stable list makes weekly reports comparable.

1. AIt stores a checkpoint of the last windowCorrect answer
2. BIt fetches everything each time
3. CIt runs once a day
4. DIt relies on luck

Why: No gaps or duplicates.

1. AMarketing only
2. BNobody
3. CThe guest
4. DThe responsible departmentCorrect answer

Why: Create tasks for fixes.

1. ABy counting stars
2. BAgainst labelled past reviews per topicCorrect answer
3. CBy reading one review
4. DNot at all

Why: Precision and recall.

1. ANo reply
2. BA generic thank you
3. CMultiple topics with their own sentimentCorrect answer
4. DOne overall label only

Why: The reply should acknowledge both.

1. AStored and used to improve promptsCorrect answer
2. BDiscarded
3. CIgnored
4. DSent to guests separately

Why: Learn from corrections.

1. ADNS
2. BGoogle Business Profile and partner APIsCorrect answer
3. CInstagram stories
4. DPrinter drivers

Why: Always respect their terms.

1. AThe bot
2. BGuests
3. CEach property's duty managerCorrect answer
4. DOne central inbox only

Why: Local action is fastest.

1. APlatform character limits and rulesCorrect answer
2. BNo rules
3. COnly grammar
4. DOnly the logo

Why: Adapt length per platform.

1. AInto the public reply
2. BTo all staff
3. CNowhere
4. DPrivately to the managerCorrect answer

Why: Staff issues are handled privately.

1. AAn argumentative reply
2. BImmediate deletion requests
3. CNo response ever
4. DA calm factual reply after manager reviewCorrect answer

Why: Other guests read the reply.

1. AA new tag only
2. BA linked maintenance ticketCorrect answer
3. COnly a thank-you reply
4. DA marketing email

Why: Operational fixes close the loop.

1. AStar ratings from one guest
2. BNormalising for volume and rating scalesCorrect answer
3. COnly total review counts
4. DRaw averages from mixed platforms

Why: Different platforms and volumes skew raw numbers.

1. AA scheduler calls the flow's APICorrect answer
2. BThe model wakes itself
3. CGuests trigger it
4. DBy opening the Playground

Why: An external scheduler or cron job calls the flow endpoint.

1. AWithin a month
2. BNever at all
3. CAt year end
4. DWithin the same shiftCorrect answer

Why: Document the follow-up.

1. ACleanlinessCorrect answer
2. BStock prices
3. CWeather forecasts
4. DSports scores

Why: Part of the fixed topic list.

1. ARaw logs
2. BPrompt text
3. CRating trends and top issues fixedCorrect answer
4. DModel weights

Why: Leaders care about business outcomes.

1. ATo slow replies
2. BReputation and legal riskCorrect answer
3. CBots cannot type
4. DPlatforms forbid AI

Why: Managers catch tone issues.

1. AGeneric templates only
2. BBooking numbers
3. CBlame on guests
4. DSpecific points and real fixesCorrect answer

Why: Sincere and specific.

1. AWith a machine translation of the review
2. BNot at all
3. CIn French, with a translation shown to the managerCorrect answer
4. DIn English only

Why: Replying in the guest's language is courteous.