PayPal is building an agentic AI platform that allows consumers to buy products through a chatbot. 

The payments company aims to “provide a plug-and-play experience,” Farhad Farahani, head of agentic commerce and personalization, said during a presentation at the Nvidia GPU Technology Conference on March 18.

“If a small business wants to plug and play their capabilities or inventory or services into the agentic universe, they don’t need to go and integrate with different tools and modalities,” Farahani said. 

PayPal’s pilot agentic platform is built as a consumer-facing chatbot, where users can ask questions. PayPal shows the user the products they can purchase and have shipped to them with a single click, Farahani said, adding that no launch date has been set. 

“You can tell your agent, ‘Hey, I want to buy something and once the price drops to $20 or $200 or whatever, go and execute the purchase for me or go find the best deals across the board,’” he said. 

Merchant agents 

Merchants can also benefit from the agent, which was designed to avoid friendly fraud and chargebacks. 

PayPal has trained its agent to initiate transactions with specific methodology and rules that the consumer has provided to avoid making mistakes, he said. 

Agentic transactions can happen without a consumer present when an input has been provided to the chatbot, Farahani said. This is when it is essential for the agent to already know a consumer’s preference, such as color, size and price point. 

The San Jose, Calif.-based PayPal aims to make its agentic AI platform LLM agnostic, Farahani said. The company also plans to make the platform multimodal, meaning it will be able to communicate to the chatbot through voice commands 

PayPal works with OpenAI for its LLM needs and has teamed with Mastercard to develop safe agentic AI rails, according to the company. 

The company relied on synthetic data to develop its agentic AI tool, Farahani said.   

“Synthetic data was our best friend in terms of really coming up with massive data sets that we can actually use for model training as real customer data slowly comes in.”