Inside CRED's 98% accurate AI concierge built on GPT-5

CRED, the members-only club serving over 15 million of India’s most creditworthy consumers, is making a hard pivot to AI-native operations, and the early numbers are eyebrow-raising. In a recent conversation, Swamy Seetharaman laid out how the company is using a fleet of OpenAI-powered tools to chase a lofty goal: turning every employee into a ‘10X’ version of themselves without breaking the pristine, high-trust experience that defines the brand.

The centerpiece is Cleo, an AI conversational companion running on a stack that includes GPT-4.0, GPT-5, and o3. This isn’t a clunky chatbot deflecting queries into a void. Cleo is designed to handle informational, contextual, and transactional requests—think ‘Am I eligible for CRED Cash?’ or processing a refund to an original payment method—by diagnosing intent and mapping it to standard operating procedures. The results since launch are strikingly concrete: a 98% resolution accuracy rate, a 31% drop in session drop-offs, and an 18% bump in resolving complex, multi-intent conversations that would normally trip up less sophisticated systems.

Behind the scenes, two internal tools are reshaping workflows. Thea acts as a co-pilot for support agents, summarizing messy, multi-format conversations—including voice notes and Hinglish—and suggesting next steps. Stark is aimed at operations teams, slashing the time needed to create or update SOPs from days down to minutes. The combined effect has pushed CRED’s customer satisfaction scores up by 14 percentage points, a massive shift for a user base that Seetharaman describes as demanding ‘trust, transparency, security, reliability, and exceptional design.’

Seetharaman noted the initial skepticism that greets any new tech rollout, but that evaporated once teams interacted with the company’s internal evaluation framework. The real surprise, he said, was the speed of human adaptation once people felt the unlock. Next up is expanding Cleo across all business lines and building detection tools for ‘data dead-ends’—queries the system can’t answer—that will automatically feed back into the knowledge base to close gaps in real time. For companies still on the fence, his advice is blunt: identify whether you need efficiency, effectiveness, or both, and then align AI adoption with your actual values. For CRED, that meant betting on compounding speed with precision.