A BERT + NLG system that detects negative sentiment and auto-replies with empathetic responses across platforms, with entity detection and churn mitigation via discount logic.
Unhappy customers rarely file a ticket — they post, and then they leave. The client needed to catch that negativity in the wild and respond before it hardened into churn.
The challenge
At scale you can't read every comment, and a tone-deaf canned reply makes things worse. The system had to classify sentiment accurately, understand what the complaint was about, and respond like a human would.
Our approach
A fine-tuned BERT model classifies sentiment and detects the entities a complaint targets; an NLG layer drafts an empathetic, on-context reply; and churn-mitigation logic offers a targeted discount when a user looks likely to leave. It runs across the client's platforms automatically.
Under the hood
- Classification — Fine-tuned BERT sentiment model
- Understanding — Entity detection on complaints
- Response — NLG empathetic auto-replies
- Retention — Churn mitigation via discount logic
What we built
- Sentiment classification tuned for real customer language
- Entity detection to know what a complaint is about
- Empathetic, on-context auto-replies via NLG
- Churn-mitigation offers for at-risk users
Results
- Negative sentiment is caught and answered fast
- Flagged users churn measurably less
A Lightning Leap case study — one of the AI systems we've designed, built, and shipped.