78% of support tickets resolved before a human sees them
RAG-powered support across web, WhatsApp and email — fluent in 38 languages
A DTC ecommerce brand serving 40+ countries had a support team drowning in repetitive questions and translation overhead. We shipped a multilingual assistant that deflects three quarters of tickets with a 4.8/5 CSAT.
What was breaking
- Tier-1 support was 60% translation, 30% policy lookups, 10% actual reasoning — a $2.1M/year cost line growing faster than revenue.
- Peak-season CSAT dropped below 3.9 as ticket volume spiked and multilingual agents were spread thin.
- Brand voice was inconsistent across languages — European copy read cold, LATAM read overly casual.
How we built it
- Ingested the full help center, product catalog, shipping policies and 18 months of resolved tickets into a Pinecone vector store with per-locale metadata.
- Built a retrieval-augmented assistant with a locked brand-voice evaluator that scores every draft on tone, warmth and correctness before it sends.
- Integrated with Gorgias for a graceful human handoff — the agent never guesses, and always escalates below a confidence threshold.
- Deployed on-site chat, WhatsApp Business, and email intake behind a single unified conversation store.
Under the hood
- Intake: web widget + WhatsApp Business + email
- Retrieval: Pinecone vector store keyed by locale + product line
- Reasoning: GPT-class LLM with brand-voice evaluator gate
- Handoff: Gorgias with confidence-based escalation
- Observability: per-conversation trace, per-locale CSAT dashboard
- Guardrails: policy checker for refunds, returns, shipping claims
What changed for the business
How it shipped
Analyzed 18 months of tickets by locale and identified the 12 intents that made up 71% of volume.
Ingested the knowledge base, built the retrieval layer and shipped a shadow-mode assistant on the web widget.
Enabled autonomous mode with confidence-based escalation, connected WhatsApp Business, launched brand-voice evaluators.
Handed over the retrieval pipeline, evaluators and CS dashboard. Support ops team owns content freshness.
What operators said
“The evaluator was the unlock. Once every reply had to earn a warmth and correctness score before it sent, we stopped worrying about brand drift entirely.”
More case studies
What people are saying
Turn your business into an AI-powered machine
30 minutes. No pitch. We map the highest-leverage AI plays in your business — yours to keep.