Zendesk built custom agents on Claude and reached 1 million agent executions in 7 weeks

The work that stayed manual

Zendesk already helps companies automate a large share of their customer requests with AI. The harder part is the work that happens between a customer’s first message and the final outcome, like exceptions to a return policy, complicated warranty claims, or detailed billing questions.

Smaller companies handled that work with people, one request at a time. Larger ones with AI teams built one-off automations, which took a lot of engineering time and only covered the easiest cases. “Every customer is so unique to what they do,” said Abhinay Kathuria, Director of AI/ML at Zendesk. “There’s a lot that happens behind-the-scenes, and a lot of this was being handled manually.”

Powering custom agents with Claude Sonnet

Kathuria’s team tested models from Anthropic and other providers on the testing system they use across Zendesk products. It scores each model on 15 metrics using agents like the ones customers actually write, and the team runs it again whenever a new model comes out. Because these agents act on real customer requests, getting the answer right matters a lot.

Zendesk chose Claude Sonnet 4.6. Sonnet did “exceedingly well” on pilot tasks like reading PDFs, and in Zendesk’s testing, it needed fewer steps to reach a decision. “A competitor model might come to the same conclusion, but it might take four or five planning turns to get there,” Kathuria said. “Sonnet could handle a lot of those in one shot.”

Kathuria now makes the same argument to Zendesk customers. “You can’t just look at token price in isolation,” he said. “You have to look at the cost per task.”

Where Claude runs came down to security and data locality. Zendesk already runs its infrastructure on Amazon Web Services (AWS), so it serves Claude through Amazon Bedrock. Zendesk keeps model processing in the AWS hosting regions it already uses, which helps it meet customers’ data-residency and governance requirements. Zendesk says this also means its customers don’t have to approve or manage an additional cloud subprocessor. “Our most important security benefit is that Claude runs through the AWS environment we already operate and govern,” Kathuria shares.

Agents written in English, connected to the tools teams already have

To set up a custom agent, a Zendesk customer writes instructions in plain English, then connects the agent to Zendesk actions and outside tools like Jira and Google Drive in a few clicks. It takes about 30 minutes to set up something that used to take months of engineering work.

Take incident triage. When a lot of similar tickets come in at once, the agent notices, files a Jira issue, posts in Slack to let engineers know, and tells the affected customers the problem is being worked on. Once the fix ships, it closes all the related tickets and adds a note explaining what went wrong.

Zendesk is also building a router to send each task to the right place. Simple requests, like working out what a customer is asking for, take one call to Sonnet. Harder ones go through a longer process where the agent plans its steps before acting. After each run, the agent reviews its own work and saves what it learned as a skill it can use next time.

Customers decide which actions each agent is allowed to take. Zendesk is also shipping ways to watch and test agents, including a log of every tool call, simulated conversations, and A/B testing. “Our customers are asking for exactly the same things we need internally,” Kathuria said. “A lot more agent observability, and how to test these agents in production.”