ClassPass Ditches Human Agents for Decagon AI and Slashes Support Costs by 95%

By all accounts, ClassPass should have been drowning. As the global fitness platform scaled to 2,500 cities across five continents, the volume of customer support tickets exploded. Members were booking, canceling, disputing, and rescheduling classes around the clock. The company’s existing support infrastructure, fragmented and reliant on BPOs, couldn’t keep up. Costs spiraled, agents burned out, and customer satisfaction began to waver.

Instead of imploding, ClassPass implemented a solution that is rare for a consumer-facing company: it replaced every human support agent with AI. This wasn’t a pilot or a narrow use case; 100% of incoming support tickets are now handled by AI agents, autonomously, from start to finish, with no fallback to human operators. Over 2.5 million customer conversations later, satisfaction scores have remained steady, support costs have dropped by 95%, and the company runs a chat-first, 24/7 support operation at scale.

When ClassPass launched its formal RFP to overhaul its strained support system, the team didn’t look for the flashiest AI startup. Instead, they evaluated 13 vendors using four measurable criteria: QA-scored response accuracy, no-code tools for internal CX ownership, vendor responsiveness, and long-term cost reduction.

Decagon, a YC-backed startup, made a compelling case with its approach. What set Decagon apart from traditional AI support tools was its focus on Agent Operating Procedures (AOPs). Unlike many chatbot platforms that rely on rigid decision trees or complex, opaque large language models (LLMs), Decagon’s system was built on clear, programmable workflows. These workflows allowed companies like ClassPass to refine and control their support automation process over time, ensuring a solution tailored to their specific needs.

Decagon’s platform introduced Agent Operating Procedures (AOPs): structured, natural language instructions that define exactly how AI agents should respond to specific situations. These workflows are editable by customer experience teams, programmable by engineers, and transparent for any team member to audit. They reflect real company policies, rather than abstract or generic guidelines.

At Decagon, we’ve taken a radically different approach to building conversational AI agents for customer experience.

We call it Agent Operating Procedures— or AOPs. ✨ pic.twitter.com/FZKhOOLkZC

— Decagon (@DecagonAI) April 10, 2025

For ClassPass, this meant codifying complex workflows such as refunds, suspensions, and localization, into machine-executable logic, replacing human judgment with AI’s precision. The AOPs allowed the AI agents to function as full-time support staff, not just automated responders.

Decagon’s appeal was not just technical. It provided ClassPass with full operational control. Customer experience managers could modify workflows without engineering support. QA teams could test, version, and deploy updates based on real-time interaction data. Engineers could audit every decision the AI made. This level of transparency was a stark contrast to legacy systems, where a simple policy change or new ticket type could mean weeks of external consulting and fragile bot updates.

With Decagon, ClassPass could update its workflows quickly and seamlessly, making iteration a central feature of the system. As policies evolved, updates were implemented in days rather than months.

“This didn’t happen overnight,” said Jesse Zhang, Decagon’s cofounder, after the milestone deployment. “It took smart auditing of real conversations, iteration to train AI agents further on particular support policies, and a forward-thinking team at ClassPass who invested in creating a transformed customer experience.” The system is now responsible for 2.5 million conversations, operating with no human fallback, at 95 percent lower cost—and without sacrificing customer satisfaction.

Thanks to Decagon’s infrastructure, ClassPass shifted from email-heavy workflows to a chat-first model that operates around the clock. Localization vendors were eliminated, thanks to Decagon’s multilingual capabilities. New specialist roles were introduced to handle edge cases and further improve bot performance. Agent morale was preserved and reallocated to high-value tasks like quality assurance and training.

“ClassPass didn’t treat AI as a ‘set it and forget it’ tool,” said Brian Fields, Chief Revenue Officer at ClassPass. The team actively trained the system, audited real customer interactions, and fine-tuned responses to align with specific customer service policies. This iterative approach ensured AI performance continually improved, resulting in not only cost savings but also greater operational agility. ClassPass no longer needed to scale headcount or outsource support to keep pace with demand.

Unlike many AI customer support vendors that promise dazzling chatbots or human-like avatars, Decagon didn’t rely on decision trees or escalate issues to humans at the first sign of ambiguity. The AI agents became the support team. Fully integrated into ClassPass’s existing systems and working within the Zendesk framework, they automated not just responses but full resolutions.

And unlike many GenAI tools, Decagon’s agents don’t hallucinate. Every response traces back to policies defined by the customer, ensuring accuracy and consistency. This level of transparency is built into the system, providing traceability for every decision the AI makes.

Decagon’s platform is built for enterprises with real volume, complex workflows, and specific operational needs. It prioritizes reliability and structure, favoring clear, defined procedures over speculative or experimental approaches.

Implementing Decagon requires translating business operations into machine-executable procedures. While this demands more effort upfront, the result is clear: no human agents, no escalation queues, and no decline in customer experience.

As Sarah Vandenbrook, Senior Manager of CX Operations and AI at ClassPass, put it, “When we launched, it felt like we were their only client.”