Reduced Agent Workload
62% of L1 tickets resolved without agent involvement, freeing the team to handle a 35% volume increase without adding headcount.
Technologies used
Python
OpenAI
Confluence
Node.js
Docker
AWS Lambda
Motivation
The client’s support team handled 8,000+ tickets per month, up 40% YoY, without insight into ticket drivers or AI opportunities. Response times were slipping, CSAT was 61% and declining, and adding headcount was becoming unsustainable. Existing tools like Zendesk and Confluence were underutilized. Azati conducted a two-week As-Is assessment to uncover the real issues, shaping the solution that followed.
Main challenges
Challenge 01
The Real Problem Was Invisible Without Assessment
What looked like a staffing problem was actually process inefficiency: 60% of incoming tickets were repetitive L1 queries answerable from existing documentation. Agents spent most of their time on non-judgment work, retrieving and formattin, rather than tasks requiring expertise. The solution wasn’t more people; it was removing the right work from their queue.
#1
Challenge 02
Existing Knowledge Base Was Unusable
The client's Confluence instance held 2,400 articles, but agents rarely referenced it. Content was inconsistently structured, partially outdated, and not organized for retrieval. Direct LLM integration without restructuring produced hallucinations at an unacceptable rate. Before any AI could work reliably, the knowledge foundation had to be rebuilt.
#2
Challenge 03
No Escalation Logic or Performance Visibility
There were no defined criteria for what constituted an L1 versus L2 ticket. Senior agents were pulled into simple queries. Critical enterprise issues sometimes sat in the general queue. Zendesk reports showed volume and CSAT, but nothing about which categories drove the most cost, which had the highest re-open rate, or where knowledge gaps were causing errors.
#3
Challenge 04
No Internal Capacity to Operate AI Ongoing
The client had no dedicated AI engineering team. Any system deployed without ongoing management would degrade within weeks as the product evolved, new features, pricing changes, updated policies. A one-time implementation was not an option. Managed operations had to be the model from day one.
#4
Our approach
As-Is Before Anything Else
Azati spent the first two weeks in assessment mode, auditing ticket data, interviewing agents, reviewing the Confluence knowledge base, and mapping the full support workflow. The output was a clear picture of where time was being lost, what could be automated, and what the AI layer needed to do to deliver measurable impact.
To-Be Built on What Already Existed
Rather than propose a new platform, Azati designed a To-Be architecture that extended the client's existing investment. Confluence content was restructured and embedded into a vector store for RAG. AI responses and routing decisions were written back to Zendesk via API, agents saw AI assistance inside the tool they already used. Zero workflow disruption, zero retraining.
Confidence-Based Routing
The To-Be system operates on three tiers: high-confidence queries trigger automated responses, medium-confidence queries surface a drafted reply for one-click agent approval, low-confidence and sensitive topics escalate immediately with an AI-generated context summary. The system was designed to know what it didn't know.
Managed Operations as the Core Deliverable
Go-live was not the end of the engagement. Azati proposed, and the client accepted, a managed operations model from the start. Azati owns the AI layer continuously: maintaining knowledge quality, monitoring hallucination rates, optimizing token costs, and expanding automation coverage as the product evolves.
Facing the same challenge?
Bring your complexity. We'll bring the plan. Select a convenient slot to start a conversation with our experts.
Schedule a callSolution
01
Knowledge Base Restructuring and Embedding
Azati audited all 2,400 Confluence articles, archived outdated content, restructured remaining articles for retrieval quality, and embedded the knowledge base into a Pinecone vector store optimized for support query patterns. An automated sync pipeline keeps the knowledge base current as new documentation is published.
Key capabilities:
- Full Confluence content audit and gap identification
- Article restructuring and metadata tagging
- Vector embedding optimized for RAG retrieval
- Automated sync pipeline for ongoing content updates
02
AI Triage, Routing, and Response Generation
Intelligent ticket classification across 40+ categories, confidence-based routing, and AI response generation, all surfaced natively inside Zendesk. Agents interact with AI suggestions in the tool they already use, with no interface change.
Key capabilities:
- Intent classification and priority scoring
- Confidence-tiered routing across three escalation levels
- RAG-based response generation with source citation
- Agent feedback loop for continuous quality improvement
03
Zendesk Integration
All AI activity is written back to Zendesk in real time, suggested replies, ticket tags, routing decisions, escalation notes. No parallel systems. No change to existing agent workflow.
Key capabilities:
- Native Zendesk API integration
- Automated ticket tagging and status updates
- AI action logging per ticket
- Reporting via existing Zendesk dashboards
04
Managed AI Operations
Azati runs the AI layer on an ongoing basis under a defined SLA. Weekly hallucination sampling, monthly CSAT and cost analysis, quarterly knowledge audits, and regular expansion of automation coverage to new ticket categories, all reported to the client's Customer Operations Director.
Key capabilities:
- Weekly hallucination sampling and correction pipeline
- Monthly performance and cost reporting
- Ongoing Confluence knowledge updates
- Quarterly automation expansion
Results & business impact
Faster Responses
First response time on automated tickets dropped to under 3 minutes. Agent handling time on assisted tickets reduced 44% through AI-generated context and suggested replies.
Higher Customer Satisfaction
CSAT improved 18 points within the first 90 days, driven by response consistency and dramatically faster resolution on common queries.
Cost Control
AI compute costs running 40% below initial projections through active model routing and caching management maintained by Azati.
Scalable Automation Growth
Automation coverage grew from 12 ticket categories at launch to 31 categories by month nine, without additional implementation budget. Every month of managed operations expanded the ROI.
