Why DialpadGPT is an AI Solution Tailored for Enterprise

How do you solve a problem like AI misinformation?

One of the main criticisms levelled at models like ChatGPT is that it can answer prompts with incorrect data or even outright lies, engendering a distrust of leveraging AI for business purposes when complete accuracy is paramount. That ChatGPT effectively parses the entire internet — and by extension, the vastness of human knowledge spanning millennia — problematises that accuracy as contradictions and paradoxes inevitably present themselves.

This is why one of the subplots of the AI explosion has been the rapid development of specialist AI-powered tools designed to help businesses, notably in the customer intelligence space. An AI that could be relied upon to materially improve agent performance and productivity and enhance customer satisfaction could be a game-changer.

Enter Dialpad with DialpadGPT — a domain-specific large language model (LLM) that Dialpad described as the “first” to be designed to cater to the evolving needs of modern businesses.

“It’s actually more accurate because it’s trained on our data, relevant business data, not the entire world wide web of conversational data,” Brian Peterson, Dialpad Cofounder and CTO, told UC Today. “So, it’s interesting that a small team like ours can still make something as accurate or more accurate than the big competitors because it’s just trained on really relevant data.”

Peterson was at Google in the early years, back when its employees numbered in the three figures. He has been in tech, especially cloud, his “whole life” and helped build Google Voice, the company’s business phone system. That path ultimately led Peterson to Dialpad.

“When we started getting into business communications, we knew AI was going to be the default,” Peterson said. “The future of business communications is having the platform to run them, but also the AI to automate things. So we’ve been working towards this. We’ve been doing AI for five years now, since we bought TalkIQ, which was an AI communications company, five years ago.”

DialpadGPT was gradually built over those five years and uses real-time generative AI to automate tasks and enhance customer service, sales, and recruiting experiences for organisations of all scales.

However, the emergence of ChatGPT accelerated project timelines.

“Things really advanced quickly in the last few months, but we’ve been on top of all the technology because we’ve already been doing our own natural language processing and machine learning,” Peterson explained. “That’s one of our differentiators; we do everything in-house. Even our transcription is all our own. We don’t use a single third party for any AI. It provides a huge advantage to us because of its costs, because it’s real-time. We knew we had to take full control of the AI. We couldn’t use a third party.”

Among DialpadGPT’s other differentials is its emphasis on security.

“For us, privacy is a big deal. You’ve already seen a lot of backlash with ChatGPT with the data leakages. Well, that doesn’t work for private business conversations, so we knew we had to do something where it doesn’t go over any wall. It’s all in-house.”

“A lot of our AI is real-time,” Peterson added as another differentiator. “We do live coaching. We do live script adherence. We do sentiment analysis live. You can’t do that with something like a ChatGPT or a Google Vertex because it’s too slow. You can’t build a chatbot where it takes ten seconds to respond, right? I think that’s the one thing that’s been a huge advantage for us and is going to take it to the next level. The fact that it can detect real-time sentiment and what the person is asking about is huge.”

Another selling point is cost. By developing its AI in-house, Dialpad could make it smaller, more specialised and more efficient. This meant Dialpad could offer DialpadGPT as a default for several of its skews, or for a relatively small add-on price, such as $10 a seat.

“That’s part of our launch,” Peterson said. “The plan is, in October, we launch general availability for AI recaps to all of our skews, which we wouldn’t be able to do if you’re using ChatGPT, for example.”

Privacy, cost and real-time feedback are crucial distinguishing qualities in an increasingly crowded marketplace as more and more vendors roll out AI-powered solutions, but accuracy remains critical. DialpadGPT being trained off five billion minutes of real business conversations informs that accuracy.

“People will tell you the value is really in the data,” Peterson said. “We’ve transcribed and progressed AI through five billion minutes so far in the last few years, and it’s growing. That’s what’s making it better; that’s in our test. It has been more accurate for our needs. I’m not saying for answering anything on the web, but for, specifically, the AI that we want to release for customer intelligence, we found that it can be as accurate or more accurate.”

Using customer data to train AI models has come under scrutiny recently, with Zoom facing criticism for its Terms of Service being perceived as enabling AI training on user content without their consent. Zoom CEO Eric S. Yuan responded on LinkedIn, saying that “given Zoom’s value of care and transparency”, the company would “absolutely never train AI models with customers’ content without getting their explicit consent”.

Peterson sympathised with Zoom: “The Zoom thing was definitely just a miscommunication. I felt bad for them; that just seemed like a lot about nothing. Every single company has some form of opt-in or opt-out for this type of stuff. I mean, obviously, Google is using your emails even in work situations, I believe, for its auto-complete.”

Peterson explained that Dialpad’s customers are “already used to it” because “they’re buying us because of (the AI), and they know that to make something that’s going to be highly accurate and to be able to use AI, they’re going to want to turn it on, they’re going to want to buy Dialpad for that.”

“The other thing why a lot of people were freaked out about Zoom, in my opinion, was because they use Zoom for personal calls,” Peterson added. “When we’re dealing so much with customer communications at Dialpad, businesses are already recording that stuff. You’re recording every single call into a call centre already. So it’s not ‘Oh, my gosh, this is my personal phone call being analysed’. Naturally, we’re B2B, so people seem to understand the need for it.”

Peterson stressed that DialpadGPT follows all compliance rules, includes a full opt-out, is completely anonymised, and follows processes that customers have already been trusting them with, so no issues have arisen.

“We even do automatic opt-out for different industries because of compliance reasons, like healthcare. We have a full retention policy, so they can decide not to include the half data completely wiped from our system to meet compliance rules.”

DialpadGPT being built in-house helps protect the business from traditional AI criticisms such as data leakage or one business’s confidential data ending up in someone else’s answers.

“The way it’s built protects them from that because we’re not opening up a console where you can just ask any question for the whole worldwide web,” Peterson said.”This is just using the LLM and generative AI to do features around it. They’re all gated. If we have a chatbot, for example, we also have AI assist already. It’s like a combination of our AI NLP, but also with semantics search. So it grabs all of your knowledge base data, for example, and uses that for answers with generative AI to make it perfectly summarized and sound human.”

Peterson and Dialpad knew they had to undergo this process for the customisation and control aspects — pivotal lessons they learned from the hallucination and data leakage controversies impacting other AIs. For complete control, there could be no third-party input, and it couldn’t be exposed as a general “ask it anything” ChatGPT-style model.

“That was the control part and the customization part, and then the other thing we really realized was that these really, really big large language models, they’re cool for replacing Google search, but they’re overkill for a lot of things you want to do in specific business needs,” Peterson said.

“You’re basically getting charged for this thing that has answers for things you don’t need them for. That’s also causing the slowness. It’s a ten-second response because it’s got the entire worldwide web of information, and we don’t need that. We knew right away if we could take the generative AI tech, that’s based on what ChatGPT is based on and open source models and customize it for the specific needs, we have a huge advantage for speed, costs, and relevancy.”

This year, Dialpad announced a “Twelve months of AI” timeline, with features such as AI-powered conversation summarising — with generated action items, notifications, call purposes and intelligent call categorisation — being rolled out over time. “DialpadGPT is going to enhance that and make it even more accurate, providing insights out of those calls for you,” Peterson added.

DialpadGPT “enhanced or made possible” all twelve of these monthly features.”

AI Playbooks is another one we’re launching soon,” Peterson explained. “That’s going to be powered by a combination of generative AI, DialpadGPT. Basically, you can follow a live script of what you’re supposed to be doing. You can automatically check them off. If you’re on a sales call, you ask them this, and then you can ask them this. If it’s a support call, did you welcome with a greeting? Did you have a specific playbook? That’s going to be powered by a GPT as well.”

There are also ideas that Dialpad is experimenting with (for now), such as the ability to generate a follow-up email for a user based on their conversation with DialpadGPT. DialpadGPT could detect if it’s a sales call and could then automatically know what stage of the call it was. It could then automatically follow up with a customized email or text message and send it for the user based on the conversation and what it learned.

“That’s actually what we’re really excited about, too,” Peterson concluded. “We own the complete platform, but that does provide a big advantage when you combine it. We have all mediums, so we do all of our call contact centre, our phone system, our messaging, our social omnichannel, and email. It’s a big advantage when you combine that (with DialpadGPT).”