The challenge
LabMart's support team was drowning. Inbound tickets had grown 4× year-over-year as the platform scaled, but the triage team — eight humans reading every ticket — was capped by hiring constraints. First-response SLAs slipped from 2 hours to 18 hours. Customer NPS dropped 14 points.
The team had tried two generic AI helpdesk SaaS tools. Both failed on LabMart's specialty: clinical lab workflows have precise terminology where small misclassifications (urgent vs. non-urgent specimen issue) create real safety incidents.
Our approach
We built a custom multi-LLM classifier trained on 14 months of historical tickets, with three tiers: a fast Llama 3.1 classifier for routine triage, GPT-4o for ambiguous cases, and Claude Sonnet 4 for clinical-language edge cases. The agent is wired into Zoho Desk for tickets and Odoo Helpdesk for internal escalations.
Architecture
- LangGraph orchestrator routes between three LLMs based on confidence score
- Custom RAG layer over LabMart's product docs + 14 months of resolved tickets
- Human-in-the-loop confidence floor: anything below 78% flags for human review
- Per-category escalation rules — specimen-safety tickets bypass automation entirely
- Live audit dashboard — every AI decision logged, reviewable, reversible
Integrations
- Zoho Desk webhook → AI agent → category + priority + suggested response
- Odoo CRM update on every customer-impacting ticket
- Slack alerts to on-call engineer for any safety-flagged ticket
- Weekly auto-generated retraining dataset for human review
Outcomes
The agent went live in 7 weeks. First response dropped from 18 hours to 4.5 minutes (240×). 92% of tickets are now auto-classified with no human touch — and the 8% that escalate get prioritized by severity within 90 seconds. NPS recovered 11 points within 60 days. The support team was redeployed to proactive customer success instead of triage.
- $340k annual savings vs. hiring equivalent FTEs
- Specimen-safety tickets: zero misclassifications in 6,200 reviewed cases
- Customer-facing self-serve docs auto-generated weekly from ticket patterns
- Foundation for full agentic resolution (in flight — replies, not just triage)
“The AI workflow they built saves our customer-success team 40+ hours every week. ROI was clear inside the first month — smoother than our last three rollouts combined.”
Sarah Adler · VP Customer Success, Acuity Labs