Zendesk has helped businesses deliver great customer experiences for over a decade. Its platform powers more than 4.6 billion resolutions each year.
In early 2023, Zendesk began working closely with OpenAI to explore how AI could reshape service and product development. Today Zendesk is piloting a new class of AI agents(opens in a new window), powered by OpenAI models, that not only manage entire conversations but plan and execute responses autonomously:
- Reducing setup time from days to minutes
- Increasing automation rates toward 80%
- Giving teams full control over how the AI behaves
Even the most sophisticated service platforms face limitations when it comes to traditional automation. The standard model relied on intent classification: predict an intent, trigger a predefined dialogue or workflow, and hope the customer followed the script.
This setup worked for structured interactions, but broke down quickly with nuance, follow-ups, or edge cases.
“The old world was message in, response out,” says Adrian McDermott, CTO at Zendesk. “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.”
Zendesk began working with OpenAI to adopt a generative approach using Retrieval-Augmented generation (RAG) for basic FAQ interactions. Today, their focus has shifted to generative reasoning that enables AI agents to plan and execute tasks independently.
Zendesk’s new class of agentic AI agents is purpose-built for service. Powered by OpenAI models like GPT‑4o, the agents don’t just answer questions—they lead conversations, reason through context, and drive toward resolution.
The platform leverages a multi-agent architecture comprising of specialized agents such as:
- Task identification agent: Instead of relying on manual training this AI agent has a real conversation to understand what the user needs, asking clarifying questions and disambiguating similar issues.
- Conversational RAG agent: Extends traditional RAG by grounding in multi-turn conversation. For example, when a user asks about payment options, the agent can follow up to ask where the user is located before retrieving region-specific policies.
- Procedure compilation agent: Balancing agency with control, Zendesk’s procedure compliance agent converts business rules from natural language into a structured flow, ensuring the AI understands and visually reflects how to execute the company’s procedures.
- Procedure execution agent: Carries out actions by calling APIs, triggering workflows, and updating systems, all within the logic defined by the business.
By combining RAG with reasoning, Zendesk’s AI agents can now engage in multi-step conversations, ask follow-up questions, and adapt responses based on user input. This allows the platform to resolve complex issues autonomously, without relying on rigid dialogue flows.
“We’ve given the bot more agency in guiding the conversation while operating within Zendesk’s guardrails for quality and accuracy,” says McDermott. “The process started by understanding the customer’s issue with a high focus on driving towards resolution.”
One of the biggest shifts in Zendesk’s AI agent development has been their evolution to a hybrid development model, where agents can seamlessly move between dialogue flows and generative procedures within a single conversation.
With the new AI agent builder, businesses can define procedures in natural language. The AI agent then plans a course of action using adaptive reasoning and presents a preview of its proposed steps before going live.
AI reasoning controls provide real-time visibility into how AI agents think, ensuring teams can audit every conversation by reviewing the agent’s chain of thought (CoT) to understand how decisions were made.
This shift reduces setup time from days to minutes, and makes generative automation accessible to a far broader set of Zendesk customers.
“We’ve broken down the biggest barriers to AI adoption. Customers can now use these new agentic AI agents out of the box.”
Under the hood, Zendesk runs a rigorous internal benchmarking program to select and deploy the best models and tune prompts for each use case. The team considers latency, cost, and quality, testing new models like OpenAI’s o3‑mini across use cases ranging from RAG to background reasoning tasks.
This process allows Zendesk to evaluate, test, and deploy new models in under 24 hours.
Zendesk tracks performance both before and after deployment, using offline evals and live metrics like resolution rate, edit rate, and latency. Each model decision is documented and auditable, ensuring transparency and reliability as the system evolves.
This year, Zendesk plans to go a step further: rolling out a self-service benchmarking platform so any Zendesk engineering team can test and deploy models without needing hands-on support from machine learning experts.
Zendesk is currently piloting the new agentic AI platform with early adopter customers. The platform is designed to integrate easily with existing setups, accelerating customers’ path to 80% automation without requiring them to rebuild from scratch.
While broader metrics will follow later in 2025, early feedback has been strong: faster setup, more accurate responses, and smoother user journeys across every channel.
