The real problem isn't writing replies — it's triage across locations
A 12-location hotel or restaurant group gets roughly 480 reviews a month across all properties. That's not an exaggeration — at 40 reviews per location, it's just arithmetic. Each one sits in Google Business Profile waiting for a reply, and Google's algorithm quietly notices which ones don't get answered within seven days. Silence reads as indifference to guests and to the search engine.
Most groups don't have a process for this. The general manager handles it at one location, the marketing coordinator handles it at three others, and the rest get no reply at all. A shared spreadsheet or email inbox becomes unusable past three locations. Someone owns it, and then someone leaves, and suddenly nobody does.
The real problem isn't the writing. It's that you have 480 replies to sort, route, and prioritize every month, and only about 30% of them actually matter. Seventy percent are routine: five-star praise with no specifics ("Great place!"), four-star neutral feedback ("Good food, a bit slow"), or minor operational stuff that follows a script. The other 30% are operational complaints that move guests and revenue, or legal-risk reviews about illness, injury, discrimination, or misconduct. Those are buried under the routine noise because nobody's triaging at scale.
Sort every review into four buckets before you reply to anything
Before AI drafts a single reply, sort every review into one of four buckets. This taxonomy drives the entire automation decision downstream.
Bucket 1: 4–5 stars, no specifics. "Lovely hotel," "Delicious food," "Clean rooms." No names, no moments, no detail. Safe to auto-draft and one-click send. Volume: roughly 50% of all reviews.
Bucket 2: 4–5 stars with specifics. Mentions a staff member by name, a dish, a room type, or a moment ("Sarah at reception made our day," "The salmon was perfect"). AI drafts, human personalizes and sends. Volume: roughly 20% of reviews.
Bucket 3: 1–3 stars, operational complaint. Food quality, cleanliness, wait times, noise, parking, billing errors. The guest is unhappy but not alleging harm or misconduct. A human writes the reply (AI can assist with tone), the GM owns it, and it needs a 4-hour SLA. Volume: roughly 20% of reviews.
Bucket 4: Legal or safety risk. Illness, injury, allergen exposure, discrimination, staff misconduct, theft allegations, safety hazards, or sexual harassment. Human only. Loop in ops and legal. Immediate escalation. Volume: roughly 10% of reviews, but every single one matters.
The automation is the classification, not the writing. An LLM can sort a review into the right bucket in milliseconds and flag it for human action. Once it's sorted, the rest is routing and approval.
What AI should actually do: draft, route, translate, never auto-send negatives
Here's the automation scope that's safe and high-ROI for a hospitality group. Pull reviews from Google Business Profile via the official API into a queue. An orchestrator like n8n or Make ingests them daily, classifies each review into your four buckets, and tags them by location, language, and risk level. An LLM (Claude or GPT) does the classification — it's fast and reliable at this job.
For Bucket 1 reviews, auto-draft a reply in the guest's language. This is where AI pays for itself fastest in hotels: multilingual response at scale. A Spanish guest gets a reply in Spanish without a coordinator fluent in Spanish. For Bucket 2, auto-draft and route to the local GM or duty manager via Slack or WhatsApp for personalization and approval. For Bucket 3, route to the GM immediately with an AI-assisted draft they can edit. For Bucket 4, ping the GM and your ops lead via WhatsApp Business (which reaches 3 billion monthly users and where most hospitality managers already live) and flag it for legal review.
Set SLA timers: 24 hours to respond to positive reviews, 4 hours to respond to complaints, immediate escalation for Bucket 4. The critical safety net: never let the agent hit 'post' on anything below 4 stars. Assert human approval as an external state before the reply goes live. The AI drafts, routes, and flags. A person approves and sends.
What a human must still read, every time
Draw a hard line around judgment calls the model will get subtly wrong and damage your brand. A human must read and approve every reply in Bucket 3 and Bucket 4 before it posts. No exceptions.
Beyond that: any mention of illness, injury, allergen exposure, discrimination, or staff misconduct goes to Bucket 4 full stop. Any review that names a specific employee — positive or negative — needs human eyes before send, even if it's five stars. A compliment to Sarah feels great; a reply that uses Sarah's name without context can look patronizing or reveal schedule info to strangers. Look up repeat guests and loyalty members in your CRM and flag them for personalized replies. If a review is in a language no one on the local team speaks, translate it and have someone sign off before you send the reply. Anything referencing a competitor, a lawsuit, social media threats, or online escalation goes straight to the GM and marketing lead.
Ownership matters: the GM or duty manager owns Bucket 3 and Bucket 4 replies. Marketing owns tone consistency across the group. The system routes and reminds; humans decide and approve.
What this actually costs — and what it saves
A marketing coordinator managing reviews across 10+ locations typically spends 6–10 hours a week on this work. That's 300–500 hours a year, and the replies are often slow because it's reactive and fragmented. A typical build — review ingestion, classification, drafting, routing, and approval — sits in the €5,000–€10,000 range for a mid-size hospitality group. LLM costs are minimal: roughly €0.01–€0.05 per reply depending on model and length. Payback is usually 2–4 months on labour alone, assuming you free up that coordinator to do higher-value work.
The hidden ROI is faster replies. Google's ranking signals include response rate and response speed. A hotel that replies to 95% of reviews within 24 hours ranks higher than one that replies to 60% in a week. Faster replies also convert better on the listing: guests see you're responsive and engaged. Real euro ranges for this class of automation typically run €5k–€15k depending on your location mix, language complexity, and integration depth. The payback window narrows if your current response rate was below 50%.
A 30-day rollout we'd run for a hotel or restaurant group
Week 1: Audit your baseline. Pull data on how many reviews each location gets per week, current response rate (%), median time-to-reply, language mix, and who currently owns it. This is boring and essential.
Week 2: Connect Google Business Profile API access and build the four-bucket classifier. Approve a tone guide per brand (luxury hotel replies sound different from casual bistro replies). Test the classifier on 50 real reviews and adjust thresholds. Bucket 4 should be tight — no false negatives on legal risk.
Week 3: Pilot at one or two locations. Every reply — even Bucket 1 — goes through human approval. Let the GM see the drafts, feel the workflow, give feedback. Measure: % of reviews replied to, median time-to-reply, volume by bucket, how long approvals actually take.
Week 4: Expand to the full group. Flip Bucket 1 to one-click send (human spot-checks, no approval gate). Keep Bucket 2–4 under human approval. Measure the same metrics. At day 60, review what broke, what sped up, and whether you can loosen Bucket 2 approval for certain locations.
Next step: pick a pilot location and measure
Choose one hotel or restaurant location and measure your reply rate and median response time for 30 days before you build anything. Then build, and measure again for 30 days. That's your real ROI signal and your case for rolling out across the group. Once you have that baseline, reach out to us and we can scope the automation in a week and build it in a month. Faster replies, lower coordinator load, zero brand risk — that's the actual win.