PayPal AI Assistant Raises Accuracy to 95% Without Model Change

What changed

PayPal says its AI assistant’s internal question-answer accuracy rose from 43% to 95%, a 52-percentage-point increase. The company says it achieved that improvement without changing the underlying large language model, instead adding proprietary business context kept separate from the model.

Why it matters

The striking part is where PayPal says the gain came from: the surrounding system, not a newer model. That suggests a practical lesson for companies building AI tools: better answers may depend as much on giving a model the right business context as on replacing the model itself.

The claim is meaningful, but it is still an internal result. The evaluation method, question set, sample size and operational error rate are not provided, so the 95% figure cannot yet be treated as a general measure of how well the assistant performs elsewhere. Still, PayPal’s account points attention toward the less glamorous layer of AI products: the information architecture that helps a model know which company-specific facts to use.