San Mateo-based Conversica Inc. has become a leader in conversational AI through its commitment to creating products that are more accurate and display more human-like responses.
But rather then train products using generic data sets, the company has released a new addition to its chat platform that instead employs customer-specific data and creates a “first of its kind” enterprise solution that ensures the best responses without going “rogue,” CEO Jim Kaskade said.
“We like to think we can make it sound better,” Kaskade told the Business Journal. “These large language models allow you to write poetry so it can take a dry answer and just make it so cool.”
Conversica was founded in 2007 in Bellingham, Washington, and brought its first AI solution tailored to help sell products to the market in 2009 as the financial markets melted down.
Conversica built much of its business around automating marketing, sales and customer success tasks while most of the AI players that started focused on customer support.
Kaskade said the company now has over 1,000 customers, making it the largest in the conversational AI space.
Some of its largest tech and telecom clients include Oracle, Hewlett Packard Enterprises, AT&T and T-Mobile. Conversica also has nearly 50 sports teams in major leagues using its products, including the National Basketball Association’s Phoenix Suns and Milwaukee Bucks.
The San Jose Earthquakes also use the company’s “win-back” campaign that generates sales from engaging lukewarm fans who have previously signaled interest in attending events.
Generative AI had a market size of about $40 billion in 2022, when consumer products like ChatGPT entered the market. Bloomberg Intelligence estimates that the market could grow to $1.3 trillion by 2032.
But as the marketplace has grown, so has the need for a more reliable product.
Testing for large language models has improved mightily in the last four years, improving its accuracy from as low as 40% to nearly 100% now, but it’s still not enough, Kaskade said.
“It’s pretty much at 100% and what you’re hearing is it’s still not accurate,” he said. “That’s why we’ve turned to individualizing models for each client.”
Conversica's latest innovation attempts to leverage the best of both retrieval augmented generation, which allows platforms to find the most relevant information, and GPT, the framework that allows platforms to create a human-like response.
The goal he said was for the company’s chat products to learn each brand, which in turn would allow it to give more specific answers and cut down on wrong or misleading information, known as “AI hallucinations.”
Chatbot satisfaction can be the difference between a sale and turning a customer off completely.
So if you asked the chatbot a question about a PepsiCo product, it would never mention anything about Coca-Cola, he said.
“It can’t respond with stupid answers,” Kaskade said. “It has to be trained to legal sensitivities and there has to be 100% confidence that it will never go rogue.”
