What this automation does
New reviews get a thoughtful reply within a day, angry customers reach a manager quickly, and the issue behind a bad review gets fixed instead of just answered.
- Collect. On a schedule, the workflow pulls new reviews for each location from Google Business Profile and any other site you connect through an official API.
- Classify. A small model labels each review: rating band, topics mentioned (wait time, staff, price, quality), whether it names an employee, and whether it raises a sensitive issue such as safety, health, discrimination or a legal threat.
- Draft. A model writes a reply in your voice that references something specific the reviewer said, thanks them by first name only if they used it, and never includes private details or offers compensation in public.
- Route. Positive, non-sensitive reviews follow the auto-post rule for that location. Everything else goes to the location manager in Slack with the review, the draft and Approve, Edit and Escalate buttons.
- Act on the problem. Negative reviews also create a ticket in Zendesk or a task in your CRM so someone contacts the customer privately and fixes what went wrong.
- Log. Every review, draft, edit, approver and posting time is written to a sheet for audit and trend reporting.
When it is worth building
It is worth building when reviews arrive faster than anyone answers them, typically several locations or more than a few dozen reviews a month, or when replies sit for a week because the owner is the only one allowed to write them. Response speed and specificity matter to the next customer reading the thread.
For a single location with a handful of reviews a month, write the replies yourself. Many reputation management platforms also include AI reply drafting; if you already pay for one, try its feature first. A custom build makes sense when you need your own approval rules per location, tickets in your own support system, and the review data feeding into customer feedback analysis.
Platform policies to respect
Keep these rules in the system design, not just in a training doc:
- No fake or incentivized reviews. Do not write, buy or reward reviews, and do not have staff or AI post them. The FTC’s rule on consumer reviews and testimonials, in effect since October 2024, prohibits fake reviews and incentives conditioned on a positive or negative review, and Google bans review incentives of any kind.
- No review gating. If you send review requests, send them to everyone, not only customers you expect to be happy.
- No suppression. Do not pressure or pay customers to remove negative reviews. Respond, fix the issue, and let them update the review if they choose.
- Respect privacy. Do not reveal order details, appointment details, or anything the customer did not say publicly.
Architecture
| Layer | What we use | Why |
|---|---|---|
| Review sources | Google Business Profile API (approved access) and other official APIs | Reading and replying through sanctioned channels keeps accounts safe |
| Orchestration | n8n on a 15 to 60 minute schedule | Built-in Google Business Profile, Slack, Sheets, HubSpot and Zendesk nodes; HTTP Request for other review APIs |
| Classification | Claude Haiku 5.5 | Fast and inexpensive for labeling short text |
| Drafting | Claude Sonnet 5.5 or OpenAI’s GPT models with a voice guide | Specific, warm replies that avoid repeating the same template |
| Approval | Slack interactive messages per location | Managers approve from their phone in seconds |
| Follow-up | Zendesk ticket or HubSpot task | The problem gets an owner, not just a public reply |
| Audit log | Google Sheets or Postgres | Who approved what and when, for every reply |
How to build it
- Request API access early. Google Business Profile API access requires an application and approval, and Google only accepts applications for profiles that have been verified and active for at least 60 days. Approval can take time. Start this on day one.
- Write reply rules per rating band. Length, tone, whether to invite offline contact, phrases to avoid, and the exact contact line to use for complaints.
- Build collection with dedupe. Store each review ID and its last-seen text. Reviewers edit reviews, so detect changes, not just new IDs.
- Classify before drafting. The classification decides the route, so test it on a few hundred past reviews and fix the sensitive-topic detection before anything else.
- Draft with variety rules. Give the model recent replies for that location and tell it not to reuse openings or closings, so replies do not read as copy and paste.
- Start with approval on everything. For the first four weeks, route every draft to a person. Track how often they edit positive-review drafts.
- Enable auto-post narrowly. Turn it on only for four and five star, non-sensitive reviews at locations where edits were rare. Keep negative reviews on human approval permanently.
The downloadable starter template below covers Google review collection, classification, drafting and Slack approval.
What breaks in production
- Same reply, every review. Even good drafts become obvious when twenty replies start the same way. Track opening phrases and enforce variety.
- Spam and fake reviews. Replying to a fake review legitimizes it. Flag suspected spam for the manager to report through the platform instead of answering.
- Edited reviews. A three-star review becomes one star after your reply. Re-check replied reviews and re-route if the rating drops.
- Named employees. Praise for a named employee is fine to acknowledge; complaints about one must never be discussed publicly and always go to a manager.
- Language mismatch. Reply in the reviewer’s language, and route languages your managers cannot read to someone who can approve them.
Timeline and cost
A Google-only setup with classification, drafting, Slack approval and a ticket for negative reviews usually takes one to three weeks, plus whatever time API approval takes. Lumeor builds this in the range of $3,500 to $9,000, depending on the number of review sources, locations and approval rules, and whether it connects to your support or CRM system. Model usage per review is a fraction of a cent to a few cents. If you also want drafted replies for support email, see AI support reply drafting.
How to measure success
- Median time from review posted to reply posted, by location
- Share of reviews with a reply
- Median time for negative reviews to reach a manager
- Manager edit rate on drafts, by rating band
- Negative reviews with a closed follow-up ticket