Key points in 30 seconds
- A dealership network manages Google reviews for 6 brands: Kia, Hyundai, MG Motor, Toyota, Nissan and BYD.
- More than 500 reviews a day are read, classified and handled by an LLM engine paired with an RPA tool.
- Result: 70% less manual work and 80% faster replies.
- Sensitive cases (legal threats, safety, aggression) are automatically escalated to a person.
Results in figures
An LLM engine paired with an RPA tool, orchestrated by Power Automate.
The context
Dealership sales staff were reading and replying by hand to hundreds of Google reviews every day, for six different brands. Each reply had to match the brand's tone, be written in the customer's language and route complaints to the right team.
A prompt that reads between the lines
- Sarcasm detection: "Thanks for making me wait 2 hours" is classed as negative: the AI responds to the real issue, never literally to the sarcasm.
- Mixed emotions: "Great service, but billing took forever": the reply acknowledges the positive and addresses the concern.
- Multilingual: The original language is detected, even under a Google translation: the reply is always in the customer's language.
- Stars or text?: 4–5 stars: positive unless there is a complaint · 3: mixed · 1–2: negative. When they conflict, the AI trusts the text.
From review to published reply
- Review arrivesNew Google reviews are collected automatically.
- Emotion analysisEmotion, intent and sarcasm identified before any reply.
- CategorisationBusiness area and 1 to 6 complaint categories assigned.
- LLM draftingReply in the brand's tone, in the customer's language.
- Human approvalWhen needed: sensitive cases go through a human check.
- Published to GoogleReply published, status stored in the database.
Orchestration: Power Automate · Access: RBAC, each dealer only sees its own data.
Every review classified, prioritised and routed
| Area | Topics covered |
|---|---|
| Service | Repair · Workshop · Maintenance · After-sales billing · Parts · Inspection |
| Purchase | Vehicle purchase · Delivery · Sales · Financing · Price negotiation |
| Prospect | Showroom visit · Test drive · Enquiry · Quote · No purchase |
| None | Positive reviews: no escalation needed |
More than 30 detailed categories (waiting time, price, warranty, appointment…) feed the CRM so the right team follows up. High priority for 1–2 stars, medium for 3 stars, low for 4–5 stars.
- Automatic escalation: legal threats or refund demands, aggressive or discriminatory tone, repeated failures for the same customer, safety concerns about the vehicle or workshop.
What makes the difference
- Not fooled by sarcasm: reads emotional intent, not just words.
- Truly multilingual: replies automatically in the review's language.
- A human tone: writes like real dealership staff, without generic AI phrases.
- Smart escalation: threats, safety issues or aggression go to a person.
- Strict JSON output: ready for direct database insertion.
- Isolated data: each dealer only sees its own reviews.
Frequently asked questions
Can AI reply to Google reviews instead of a team?
Yes, for most reviews. In this project, the LLM analyses the emotion, classifies the review and drafts a reply in the brand's tone. Sensitive cases, such as a legal threat or a safety issue, go to a person before publication.
How does the AI handle sarcasm and mixed reviews?
The prompt is designed to detect sarcasm and mixed emotions. A sarcastic review is treated as a real complaint, and a mixed review gets a reply that acknowledges the positive and addresses the concern.
Are replies written in the customer's language?
Yes. The AI detects the review's original language, even when it is shown with a Google translation, and always replies in that language.