Dialpad launches generative AI trained on 5 years of proprietary conversational data

In the latest example of the trend towards smaller, domain-specific large language models (LLMs) trained on proprietary data, today Dialpad, the AI-powered customer intelligence platform, announced the launch of DialpadGPT. It is the first LLM specifically designed and fine-tuned to power features around customer service, sales and recruiting, automating tasks such as summarizing sales, customer service, and team activities.

The San Francisco-based company, which was founded in 2011 and says it is on a mission to become the “Salesforce of AI,” developed the generative AI model on over five years and five billion minutes of proprietary conversational data. DialpadGPT will be added to the company’s AI-powered platform for video meetings, calls and team collaboration.

>>Follow VentureBeat’s ongoing generative AI coverage<<

While other conversational AI providers use popular LLMs like ChatGPT, they are generic, without business-specific data. They also have scale limitations and challenges in terms of accuracy, privacy, and security.

“Unless you own that entire stack and have all those inputs, regardless of how you’re talking to a customer, it’s really hard to develop a model that’s specific to exactly these types of conversations,” Craig Walker, Dialpad’s CEO and founder, told Venturebeat in an interview.

“Dialpad’s AI infrastructure is designed for some of the most demanding enterprise uses in the world,” said Anjney Midha, general partner at a16z, whose CEO, Marc Andreessen, is on Dialpad’s board. “The launch of DialpadGPT – an LLM fine tuned for enterprise communication – will have a tremendous impact on businesses everywhere. Very few teams could pull this off.”

A 2018 AI acquisition helped put LLM expertise in place

Walker said that it is Dialpad’s longtime deep expertise in natural language processing (NLP) and AI that allowed the company to pull off DialpadGPT. In early 2018, Dialpad acquired TalkIQ,  an enterprise AI leader in conversational data.

“There’s not a lot of competitors of ours that have this in house AI capability that we’ve had for the last five and a half years since we acquired TalkIQ,” Walker said. Building an LLM is not an easy thing to do, he explained, so without a team of experts, “I don’t know where you’d start — it’s like the Black Belt ninja stuff in the AI world — you can’t quickly find an army of people willing and able to go do that.”

DialpadGPT was specifically built for the enterprise, Walker explained, and leverages the company’s transcription engine. He said that right after ChatGPT was launched last November, it was clear that Dialpad had a big opportunity.

“The first thing we did is take a call transcript, put it through ChatGPT and say ‘please summarize this call,'” he recalled. “The output was fantastic. I think everyone saw the generative AI capability to take something this complex and summarize it in a very useful way.”

The idea of building DialpadGPT was effectively born right then there: “We had to get control of this because we knew we wanted it on every single conversation,” he said.

For Dialpad, contact centers and sales are the company’s biggest current businesses, and it is about to launch an option for recruiters. “These are the areas that made the most sense, as far as having summaries of every conversation and being able to track action items,” Walker explained.

DialpadGPT is currently available to Dialpad AI platform customers in the company’s Early Adopter Program. It will be generally available to all customers by October 2023.