Support AutomationProduction · 10-week build

Cut support costs by 60% with AI triage + RAG.

A B2B SaaS company with 4 product lines was drowning in support tickets. Response times stretched past 8 hours, CSAT was slipping, and headcount quotes were unsustainable. We shipped an AI system that deflects 74% of tickets before a human ever sees them.

PDF · 3 pages · 10-week build

−60%

Support cost

2m 30s

First response

74%

Ticket deflection

Products covered

The challenge

4 products. 8-hour response time. No unified knowledge base.

Support was fragmented across Intercom, Zendesk, and Slack channels. Agents had to context-switch between products and hunt through Confluence, Notion, and scattered PDFs. Escalation paths were inconsistent, and after-hours coverage was nonexistent.

The solution

A multi-agent system that routes, answers, and escalates with context.

We built a unified AI layer: a triage agent classifies intent, a retriever agent fetches from a live RAG knowledge base, and a response agent drafts replies in the company's brand voice. Sentiment analysis triggers human escalation before frustration peaks.

Tech stack

What we built it on.

LLM & Agents

OpenAI GPT-4o
LangChain multi-agent router
CrewAI task delegation
Guardrails validation

Knowledge & RAG

Pinecone vector store
Unstructured.io parsing
Amazon Textract (PDFs)
Auto-embedding pipeline

Channels & Ticketing

Intercom Messenger API
Zendesk API
Slack incoming webhooks
Email (SendGrid)

Backend & Infra

Python + FastAPI
Celery task queues
Redis pub/sub
PostgreSQL + SQLAlchemy

Observability

LangSmith traces
Grafana dashboards
Sentry error tracking
Custom CSAT collector

Timeline

10 weeks from audit to 24/7 auto-resolve.

Week 1–2
Ticket Audit & Intent Mapping
Analysed 12 months of tickets across 4 products. Mapped 23 intent clusters and 8 escalation triggers.
Week 3–4
Knowledge Base Ingestion
Ingested 1,400 docs, 320 PDFs, and 80 video transcripts into Pinecone. Built auto-refresh pipeline.
Week 5–6
Agent Orchestration MVP
Built triage router, RAG retriever, and sentiment guard. Hit 89% accuracy on internal test set.
Week 7
Channel Integration
Wired Intercom, Zendesk, Slack, and email into a unified routing layer with priority queues.
Week 8
Shadow Mode Pilot
Ran shadow replies on 30% of volume. Human QA reviewed AI drafts before send. 12% revision rate.
Week 9
Tuning & Guardrails
Added PII redaction, profanity filters, and brand-tone guardrails. Reduced hallucinations to 0.4%.
Week 10
Full Production Rollout
Enabled auto-send for tiers 1–2. Complex tickets routed to senior agents with full context.

Results over time

Ticket deflection ramped to 74% in 10 weeks.

Percentage of incoming tickets fully resolved by the AI layer without human involvement, measured weekly across all 4 product lines.

Results after 90 days

Hard numbers, not vibes.

60% reduction in total support operating cost (salary + tooling + outsourcing).
74% of tickets deflected at tier 1 — answered without human involvement.
First-response time cut from 8h 14m to 2m 30s, 24/7.
CSAT rose from 3.2 to 4.5 after adding sentiment-triggered escalation.
Zero after-hours backlog — Monday morning queues eliminated.

Deliverables

What shipped.

Multi-agent triage router across 4 product lines
Live RAG knowledge base with auto-refresh
Unified support layer (Intercom + Zendesk + Slack + Email)
Sentiment detection + safe human escalation
CSAT dashboard and 30-day tuning retainer

"Our Monday morning ticket queue used to be 400 deep. Now it's 30 — and those are the ones that actually need a human. The team finally has breathing room to build."

Head of Customer Success · B2B SaaS (NDA)

Explore more work