Diztaly logo
Diztaly
Home/Blog/How We Built an AI Agent That Handles 70% of Customer Queries
Agentic AI

How We Built an AI Agent That Handles 70% of Customer Queries

A SaaS company with 80,000 customers was struggling with 6-hour response times and burning out support agents. Our AI agent system changed everything — 71% resolution without human involvement, CSAT up 35%.

10 min readMarch 3, 2026
Case StudyCustomer ServiceAI Agents
How We Built an AI Agent That Handles 70% of Customer Queries

What You'll Learn

A SaaS company with 80,000 customers was struggling with 6-hour response times and burning out support agents. Our AI agent system changed everything — 71% resolution without human involvement, CSAT up 35%.

When a SaaS company with 80,000 customers came to Diztaly, their customer support team was under serious strain: 15 agents handling 2,500 tickets per week, average first response times of 6 hours, CSAT scores of 3.4/5, and an attrition problem — support agents were burning out and leaving within 8 months. Twelve months after deploying an AI agent system, the same 15 agents were handling 4,800 tickets per week, first response times had fallen to 4 minutes, CSAT had risen to 4.6/5, and agent attrition had dropped from 65% to 18%. This is the full story of how we built that system.

The Problem in Detail

Our analysis of the client's support ticket data revealed that 71% of all incoming tickets fell into 12 categories — account access issues, billing questions, feature how-to queries, integration configuration, error troubleshooting, and similar repeatable issues that had clear, documentable resolutions. Only 29% of tickets required genuine human judgement, empathy, or escalation authority. Yet every ticket received the same level of human agent attention, creating a system that was simultaneously over-resourced for simple issues and under-resourced for complex ones.

Building the AI Agent System

We designed a tiered system. A front-line AI agent — built on Claude with a comprehensive knowledge base of product documentation, past ticket resolutions, and FAQ data — handled first contact for all incoming tickets. For tickets in the 12 high-frequency categories, the agent provided a complete resolution without human involvement. For tickets outside those categories, the agent conducted an initial classification and information gathering, then handed off to a human agent with a pre-populated context summary. The handoff included the agent's assessment of issue type, urgency, customer history, and recommended resolution approach — reducing the time a human agent spent understanding a new ticket from an average of 8 minutes to under 90 seconds.

The Technical Architecture

The system was built using a four-layer architecture. The intake layer received tickets from email, web form, and in-app chat and normalised them into a unified format. The knowledge retrieval layer used vector search to identify relevant documentation, past tickets, and known resolution patterns for each incoming query. The reasoning layer — the AI agent — analysed the retrieved context and the customer's query to determine the appropriate response or action. The action layer executed approved actions: sending responses, updating account settings, issuing refunds within defined limits, and escalating to human agents with context handoffs.

Results

MetricBaseline12 Months Post-DeploymentChange
Weekly tickets handled2,5004,800+92%
AI-resolved without human (rate)0%71%New capability
Average first response time6 hours4 minutes-98.9%
CSAT score3.4 / 54.6 / 5+35%
Human agent attrition65% / year18% / year-72%
Cost per ticket£14.20£5.80-59%

What Made It Work

Three factors drove the programme's success. First, the knowledge base investment — we spent three weeks building and structuring the knowledge base before building any agent, ensuring the agent had the information it needed to resolve issues accurately. Second, the human-in-the-loop design — the system was explicitly designed as a collaboration between AI and humans, not a replacement, which gained immediate agent buy-in and allowed rapid iteration based on agent feedback. Third, the continuous improvement loop — we reviewed every AI resolution weekly for the first three months, identifying failure patterns and improving the knowledge base and agent instructions accordingly. For other organisations exploring agentic AI deployment, our guide on how to deploy AI agents covers the methodology in detail.

Want to achieve similar results in your customer service operation? Diztaly's Agentic AI team has built AI customer service systems across SaaS, e-commerce, and financial services sectors. Start your AI customer service project →
Share this article:LinkedInTwitter / X

Turn These Insights Into Real Results

Diztaly's AI Marketing team will build a custom strategy for your business — backed by data, delivered with precision.