Zendesk's New AI Agents Plan Their Own Responses

Zendesk is making a hard pivot away from the brittle, scripted chatbots that have plagued customer service for years. The company is now piloting a new class of AI agents, built on OpenAI’s GPT-4o, that don’t just respond to messages—they reason, plan, and execute entire resolutions on their own. It’s a fundamental architectural shift from ‘message in, response out’ to autonomous task completion.

CTO Adrian McDermott didn’t mince words about the old approach. ‘The old world was message in, response out,’ he said. ‘Real customers change their minds, ask clarifying questions, and expect the AI to follow along naturally. In service, the only outcome that matters is resolution, and until now, bots have been somewhat limited in their ability to achieve it.’ The new system ditches rigid intent classification for a multi-agent architecture. Specialized agents handle task identification, conversational RAG that grounds itself in multi-turn context, procedure compilation from natural language rules, and direct API-driven execution.

One of the most tangible benefits is speed. Zendesk claims the new builder slashes setup time from days to minutes, while pushing automation rates toward 80%. Businesses define procedures in plain English, and the agent presents a preview of its planned steps before going live. Teams get full audibility through a chain-of-thought view, pulling back the curtain on the AI’s decision-making. This hybrid model lets agents fluidly move between traditional dialogue flows and generative procedures within a single conversation, which is a practical acknowledgment that not every interaction needs a fully autonomous agent.

The company is also turning its own model selection process into a product feature. Zendesk runs a rigorous internal benchmarking program that tests models like OpenAI’s o3-mini on latency, cost, and quality, allowing them to deploy new models in under 24 hours. They track performance through live metrics like resolution and edit rates. This year, they plan to productize that capability as a self-service benchmarking platform, giving any customer the same tooling Zendesk uses to evaluate its own AI. For a platform that already powers 4.6 billion resolutions annually, the shift toward truly agentic AI isn’t just a feature update—it’s a bet that the economics of customer service are about to change completely.