AI Is Already Resolving 90% of Customer Service Tickets - and It's Getting Smarter | Shashi Upadhyay, Zendesk

What It Covers

Zendesk's Head of Product and Engineering Shashi Upadhyay explains how agentic AI now resolves 70–90% of customer service tickets, how the company's resolution learning loop enables self-improving agents, and why adoption barriers—not technology limits—determine how fast this transformation reaches consumers globally.

Key Questions Answered

• •Automation rate benchmarks: Zendesk's best-performing customers achieve 70–90% AI resolution rates today. High-volume transactional businesses like e-commerce reach into the 90s quickly, while complex B2B scenarios still land at 30–40%. The baseline jumped from 10–20% to ~50% the moment reasoning models capable of multi-step action replaced simple chatbot search systems.
• •Resolution learning loop mechanics: When an AI agent solves a ticket, that successful trace gets converted into a reusable deterministic automation, bypassing LLM reasoning entirely on repeat queries. When it fails, the system studies the human agent's subsequent actions, runs simulations, and updates procedures via A/B testing—creating compounding improvement without manual retraining intervention.
• •Deterministic guardrails for enterprise reliability: LLM creativity must be constrained once customer intent is identified. For a return or refund workflow, Zendesk routes to deterministic code the moment the issue type is confirmed—e.g., refunds under $50 auto-approve, over $50 escalate to human. This architecture prevents the class of AI errors seen in publicized enterprise failures.
• •Forethought's auto-procedure generation accelerates go-live: Rather than requiring customers to manually author prompts and workflows, Forethought's system reads the prior year of support tickets, classifies them, and auto-generates step-by-step resolution procedures. This compresses implementation timelines from months to days and surfaces institutional knowledge that organizations have never formally documented anywhere.
• •Pricing model shift signals market direction: Zendesk has moved from seat-based SaaS pricing to resolution-based pricing—charging only when a customer issue is actually resolved. This model, adopted faster than larger incumbents, aligns vendor incentives directly with outcomes and serves as a practical signal for enterprises evaluating AI service vendors on accountability rather than access fees.