OpenAI's AI Sales Rep Hit 98% Accuracy—Then Unlocked Millions

OpenAI’s own sales team ran into a wall when ChatGPT Enterprise and Business launched. Tens of thousands of companies flooded in every month, far beyond what human reps could handle. The old playbook—routing leads through static forms—meant most prospects got a generic reply to sign up online. Harsha Chilakamarri from the Go-to-Market Innovation team described the bind: thousands of leads and only capacity to talk to a fraction. The solution was an AI-powered inbound sales assistant, built on internal connectors that pull from product docs, policy libraries, and customer stories.

The system responds to prospects in their own language within minutes. A hospital system asking about compliance gets a detailed answer in the first exchange. A company in Tokyo hears back in Japanese, not an English form letter. If a lead is enterprise-qualified, the entire conversation is handed to a human rep with full context. That alone changed the buying experience, but the real breakthrough is in how the model was trained.

Every draft response went back to sales reps for corrections, and every correction became training data. Accuracy on first emails shot from 60 percent to over 98 percent in weeks. Chilakamarri noted that a complex evaluation system built by just two engineers made that rapid iteration possible. Leadership bought in because the metrics were measurable, not just anecdotal.

The result was immediate: a small company once lost in the queue got answers within hours and signed an enterprise contract days later. Within months, the assistant unlocked multimillions in annual recurring revenue. Reps now open their inboxes to active conversations with qualified buyers instead of sifting through dead ends. OpenAI sees the same approach applying to onboarding, renewals, and support—scaling the judgment of their best reps through AI.