The AI Customer Support Assistant helps a growing SaaS company deflect the majority of support tickets by answering from the company's own documentation — with citations — directly inside Intercom, at scale.
Key Outcomes
60%support tickets deflected
<30 secfirst response, from 30 min
1,000+conversations every week
8 weeksfrom kickoff to production
Quick Facts
IndustrySaaS / Customer Support
PlatformWeb — embedded in Intercom messenger
UsersCustomers + the support team behind them
Core techRAG, GPT-4o, vector search, n8n workflows
Mindenza servicesAI architecture, RAG pipeline, evaluation harness, integrations
Background
A growing SaaS company was drowning in support volume. First responses took 30 minutes on a good day, and most tickets asked questions already answered in the documentation — but customers don't read docs, they open tickets.
The assistant was built to close that gap: a fully automated first line of support that answers instantly from the company's own knowledge, and hands the hard cases to humans with full context.
The Challenge
Business challenges ●30-minute first responses driving churn risk ●Most tickets repeating questions already documented ●Support headcount scaling linearly with users
Operational pain points ●No deflection layer before a human touched a ticket ●Escalations arriving with zero context ●No visibility into what customers actually ask
Technical challenges ●Documentation that contradicts and duplicates itself ●Preventing hallucinated answers in front of paying customers ●Keeping latency low at 1,000+ conversations a week
The Solution
Mindenza designed and built the assistant as a production-grade RAG system wired directly into Intercom — grounded, cited, and gated by evaluation on every deploy.
Architecture & tech stack ●Retrieval: hybrid vector + keyword search with reranking ●LLM: GPT-4o with strict grounding prompts ●Ingestion: automated pipeline with change detection over docs, changelogs, resolved tickets ●Channel: Intercom messenger with human handoff ●Orchestration: n8n for escalation and analytics workflows
Assistant design ●Every answer cites its source passage ●Confidence thresholds — low confidence escalates instead of guessing ●Handoff passes the full conversation and retrieved context to the agent
Implementation Process
Discovery & data audit ●Mapped ticket categories against existing documentation coverage ●Cleaned and de-duplicated the doc corpus before indexing
RAG architecture & prompt engineering ●Structure-aware chunking that never splits a step from its warning ●Grounding prompts with strict refusal behavior
Evaluation harness ●Fixed set of real customer questions scored for accuracy and groundedness ●Quality gate on every deploy — if scores drop, the deploy fails
Integration & rollout ●Intercom integration with staged rollout — 10% → 50% → 100% of conversations ●Deflection and resolution analytics wired into the team's dashboard
Results & Impact
Operational efficiency ●60% of tickets resolved before a human sees them ●30 min → under 30 sec first response time
Business impact ●1,000+ conversations handled weekly, support headcount flat ●Support team refocused on the queue that actually needs them
Qualitative wins ●Answers cite their sources — customers trust them ●Escalations arrive with full context, nobody starts from zero
"It answers like someone who has read every page of our docs — because it has." — Head of Support, client company
Why It Matters
This project shows that RAG assistants can be production-ready, safe, and measurable when paired with the right retrieval architecture and an evaluation harness that gates every deploy.
For Mindenza, it demonstrates deep expertise in:
●Citation-grounded RAG at production scale ●Evaluation-gated LLM deployment ●Deep third-party integrations (Intercom, n8n)
