01
Ground the model in real answers
Indexed 1,400 KB articles, 3 years of resolved tickets, and live account state into a pgvector store with per-tenant filters.
Products / AI showcase
Support Assistant
Customer support copilot
A grounded assistant that answers customer queries from product docs, ticket history, and account data — with graceful human handoff.
- Client
- Mid-market SaaS (12k active accounts)
- Industry
- B2B SaaS
- Duration
- 9 weeks to production
68%
faster first-response
3×
ticket deflection
92%
answer accuracy (audited)
01 · Problem
Where the team was stuck.
- 01Support team drowning in repetitive Tier-1 tickets — password resets, billing questions, feature availability.
- 02Median first-response time of 6h 40m during business hours; weekend backlog compounded into Monday spikes.
- 03Knowledge base was accurate but under-used — customers preferred to open a ticket instead of searching.
02 · Approach
How we built it.
Stack
- OpenAI GPT-4o
- pgvector
- LangGraph
- Zendesk API
02
Route by confidence, not by keyword
A LangGraph agent grades its own answer against retrieved passages. Low-confidence responses skip the customer entirely and open a draft ticket for a human agent.
03
Instrument every reply
Every answer logs the retrieved sources, the confidence score, and the customer's follow-up. Weekly review loop flags hallucinations before they scale.
03 · Outcomes
What changed.
Support NPS moved from 41 to 58 in the first quarter. The team reallocated two agents from Tier-1 to onboarding — the ticket volume that used to require them is now handled by the copilot with a human review queue.
- 68%faster first-response
- 3×ticket deflection
- 92%answer accuracy (audited)
04 · Demo preview
Live conversation
Customer asking about a billing edge case — grounded answer with sources.
Book a live demosession preview · illustrative
User
Hey, I was charged twice for the Team plan this month. Can you refund one?
· Retrieved: Billing FAQ #14 · Account #4821 invoices · Refund policy v3
AI
I can see two Team-plan charges on Account #4821 (Nov 3 and Nov 12). The Nov 12 charge is a proration from your seat upgrade — that one is expected. Want me to walk you through the math, or open a refund ticket if it still looks wrong?
User
Walk me through the math please.
AI
You added 2 seats on Nov 12. Your plan bills monthly on the 3rd, so we charged a prorated $38.40 for the remaining 21 days on those seats. That's why it looks like a second charge — same subscription, different line item. I can share the invoice PDF if that helps.
More AI case studies
All products →Enterprise RAG
Document intelligence
Search, summarize, and cite across contracts, SOPs, and regulatory filings — grounded in a permissioned knowledge base.
Read case studyAgentic Automation
Operations agent
Multi-step agents that triage inbound email, update records across systems, and orchestrate approvals with clear audit trails.
Read case studyAnalytics
Conversational analytics
Ask business questions in plain English and get trusted charts, cohorts, and forecasts backed by your warehouse's semantic layer.
Read case study