In a June 2026 interview, Jonathan Barouch, GM of Contact Center at Zendesk, outlined what an AI-native contact centre actually looks like, arguing that support teams should move beyond scripted bots and measure resolution rather than deflection. His comments landed as enterprises wrestle with how to deploy AI for Customer Support without breaking existing workflows.
Barouch said the industry is shifting from reactive automation - where bots handle narrow, scripted questions - to agentic AI that can reason, understand context, and complete multi-step tasks on its own. "Agentic AI can reason, understand context, and complete multi-step tasks end to end," he said. This means an AI agent might verify a customer's identity, pull order history from a back-end system, and process a refund without handing off to a human.
What "agentic" actually means for support teams
For support leaders, the term points to a class of AI Agents & Automation that operate beyond simple decision trees. Barouch explained that agentic systems can plan a sequence of actions, access multiple tools, and adjust when something goes wrong. They do not just answer questions - they resolve the underlying issue. This shift separates older chatbots from what Barouch calls an AI-native contact centre.
He cautioned that many organisations still confuse automation with genuine agentic capability. A bot that recognises keywords and triggers a canned response is not agentic. The real test, he said, is whether the system can handle a task that normally requires a human agent to switch between three or four applications.
Why data silos and brittle infrastructure stall progress
Enterprise teams often get stuck when they try to layer AI on top of disconnected data, outdated authentication flows, and rigid legacy systems. Barouch identified data silos and brittle infrastructure as the biggest blockers. Without a unified view of the customer, AI agents cannot reason about a full history or take meaningful action.
His advice: do not bet the whole operation on day one. Instead, run a pilot that targets one high-volume, repeatable task - such as a password reset or a shipping status update - and build outward. He stressed that A/B testing is essential to prove that an agentic workflow reduces resolution time without increasing customer frustration.
Measuring resolution instead of containment
Barouch challenged a common CX metric: containment. "Leaders should measure resolution, not deflection," he said. Containment counts how many interactions stay inside a bot without reaching a human. That metric, he argued, rewards systems that deflect customers even when the issue remains unsolved. Resolution, by contrast, tracks whether the customer's problem was actually fixed.
To make this work, he recommended a learning loop. Every interaction - whether handled by AI or a human - should feed back into a knowledge base. That loop improves self-service, agent support, and the AI's own reasoning over time. Barouch said the goal is not to fire-and-forget automation but to build a system that gets smarter from real conversations.
Why this matters for customer support professionals
Support leaders who measure success by containment alone risk building a deflection machine that frustrates customers. Barouch's framework suggests starting with one high-volume task, running A/B tests to compare resolution rates, and using the results to expand agentic workflows. The practical takeaway: reframe the conversation from "how many tickets did the bot deflect?" to "how many issues did the AI agent actually resolve?" - and build the infrastructure to track that answer.
Your membership also unlocks:
