Services About Portfolio Products Case StudiesContact AI Profile Book a free call Start a project
Back to portfolio
Case Study  ยท  AI Engineering

How we built an AI engine that resolves 68% of support tickets automatically

A 14-week engagement to design and ship a production-grade LLM support automation layer โ€” integrated into an existing Zendesk workflow without disrupting the human team.

68%
Tickets auto-resolved by AI
4ร—
Faster first response time
40%
Reduction in support costs
14w
Discovery to production

A B2B SaaS team drowning in support tickets

NeuralDesk was a growing B2B SaaS platform with 15,000 customers and a support team handling over 3,200 tickets every month. 70% of those tickets were the same questions โ€” answered in documentation nobody could find fast enough.

Average first response time had crept to 4.2 days. CSAT was falling. The support team was burning out on repetitive work and had no bandwidth for genuinely complex issues. Leadership knew AI was the answer โ€” they just didn't want a bot that felt like a bot.

They came to FiveNodes to design and build an AI-native support layer that would integrate directly into their existing Zendesk setup, auto-resolve tier-1 tickets, and route everything else to humans with full context already surfaced.

01

Auto-resolve repetitive tickets

Password resets, plan questions, how-to queries โ€” the stuff every agent answered 20 times a day.

02

No separate chatbot interface

Everything had to live inside Zendesk. No new tools, no retraining agents, no disruption to existing workflows.

03

Confidence-based human escalation

If the AI isn't sure, it hands off to a human โ€” with context already surfaced, not a blank ticket.

04

Observable and improvable

Full analytics on what the AI resolved, what it escalated, and why โ€” so the team could tune it over time.

Architecture-first, then automation

We started with 2 weeks of discovery before writing a single line of code. We analysed 18 months of ticket data, mapped resolution patterns, and identified the exact categories that were safe to automate.

This upfront work shaped every subsequent decision โ€” which documents went into the RAG pipeline, what confidence threshold triggered human escalation, and how the admin dashboard was structured for ongoing oversight.

The result was a system that felt native to the team because it was designed around how they already worked โ€” not around what was easiest to build.

01

Discovery & data analysis

Analysed 18 months of ticket history. Categorised resolution patterns. Identified the 68% of tickets that followed predictable resolution paths.

02

RAG pipeline build

Ingested product docs, Zendesk macros, past resolved tickets, and knowledge base articles into a Pinecone vector store. Built retrieval logic that surfaces the 3 most relevant sources per query.

03

LLM integration & confidence scoring

GPT-4o drafts responses with a chain-of-thought prompt. Confidence scored on source relevance + answer specificity. Below 82% โ†’ routed to human with context pre-loaded.

04

Zendesk integration

Deployed as a Zendesk App via REST API. Agents see AI-drafted replies inline with source citations. One click to send or edit before sending.

05

Analytics dashboard

Admin view showing auto-resolution rate, escalation reasons, top unresolved categories, and CSAT correlation โ€” updated in real time.

Built on a production-grade stack

AI / ML Layer
OpenAI GPT-4oLangChainPinecone
Backend
PythonFastAPIPostgreSQL
Frontend
Next.jsTypeScriptZendesk App SDK
Infrastructure
AWS LambdaECSCloudWatch

Numbers that changed how the business runs

Within 6 weeks of going live, the support team had shifted from answering the same questions all day to handling genuinely complex issues that needed human judgement. CSAT went up. Burn-out went down.

The ROI paid for the entire engagement within the first quarter of deployment โ€” and the system continues to improve as the knowledge base grows.

68%
Auto-resolution rate
More than two-thirds of tickets never need a human to write a reply.
4ร—
Faster first response
Average response time dropped from 4.2 days to under 11 hours.
40%
Cost reduction
Support team headcount held flat while ticket volume grew 22%.
4.6/5
CSAT improvement
Customer satisfaction score up from 3.8 โ€” highest in company history.

Building something AI-powered?

Tell us what you're trying to automate. We'll respond the same day with honest architectural thoughts โ€” no sales call required.

FiveNodes AI Profile

Wondering if we can build something like this for you?

Chat with our AI — describe your idea and get honest guidance on approach, stack, and timeline.

Try the AI Profile No sign-up · Instant