Angled aerial view of white speech bubble amidst rows and columns of blue speech bubbles
An arrow pointing leftHome

How uses machine learning to unlock the power of conversations

  • Adam Bluestein

Smart talk with founder Surbhi Rathore

Between emails, texts, meetings, chats, social media, phone calls and video conferences, we’re all swimming — some would say drowning — in a sea of language. In theory, that language holds all kinds of useful information and insights. But there’s just so much of it, that the “signals” in it all can be hard to find. A new generation of AI companies aims to help, by teaching computers to process, understand and respond to a wide variety of human language in a “natural” way.

Together, these companies are giving rise to a burgeoning subfield of AI called conversation intelligence, or CI, with a global market worth nearly $6 billion in 2020, according to Allied Market Research. That number is projected to reach $32.62 billion by 2030. So how does it all work?

Language-processing applications fall into two broad categories, explains Surbhi Rathore, the CEO and cofounder of, a Seattle-based conversation-intelligence startup that has raised nearly $24 million from investors like Great Point Ventures and the Alexa Fund, Amazon’s voice-focused investment arm, since launching in 2018.

The first big category, Rathore says, deals with documents and other text-based data that contains language of any sort; a second deals specifically with conversation data.

“Within the area of conversation itself,” she says, “there are essentially two types of data — the more structured, goal-oriented conversation that happens with a chatbot or an IVR [interactive voice response] system and then all these unstructured conversations that happen on voice and video, emails, chat.” In other words, “natural” language, with all its twists and turns, ums, ers, and likes.

From a business perspective, the conversation space includes product companies that focus on a specific vertical, domain or type of user. Big players include (which has a $7.5 billion valuation) and, which use machine learning to parse unstructured conversations and offer tips to help sales teams close deals. Other companies have products that focus on customer service or recruiting conversations. At the other end of the spectrum, there are a bunch of open-source tools, such as Hugging Face, that data scientists and engineers can use to create custom CI applications for specific problems within their organizations.

“We sit in the center,” says Rathore, “providing APIs for developers who are not as familiar with AI, and businesses that may not have enough capital to set up a data science team to solve every problem. Without paying any upfront money or subscribing to a new product, you can start to use conversations in the way that’s most applicable to you, integrating with tools you’re already using. In the way that Twilio standardized communications, or Auth0 did for identity, or Stripe for payments, we want to standardize how conversations are managed, analyzed, used, acted upon and learned from.”

Using’s AI tools, Rathore says, a non-specialized developer can implement conversation intelligence in as little as a few hours, or up to a few months, depending on the complexity of what it’s being used for. And its range of applications is continually expanding. In addition to taking notes, summarizing and capturing knowledge in sales and marketing calls, webinars and events, meetings, and customer service and support interactions, is “growing massively” in mental health and recruitment, Rathore says. “All the emotional intelligence that you can curate and identify from conversations can be used for everything from building a ‘coach’ for Gen Z for better interactions, to creating a better interview experience for job candidates.”

An interviewer, for example, might get real-time prompts about being empathetic, displaying listening skills, or not asking questions that are gender-biased. “It’s a lot of behavior analytics, as well as qualitative and quantitative aspects of the conversation,” says Rathore. This sort of AI interview “coaching” can occur in real-time, to improve the conversation on the fly; it can also provide post-interview conversation analytics to guide future learning.

“Different customers are taking completely different approaches to the problem,” she says.

The concept of CI is still new to most people, which means that Rathore and her team spend a lot of time educating prospective clients and building awareness. “We will come across people who ask, ‘Can I really do something with my conversations except for transcribing them?’” she says. “Yes! You can truly understand the context. You can attach that data to existing data sources and make better predictions. You can influence the way that conversations are happening in your business in real time.” Selling an API platform, versus a product, can be a bit more challenging, she says, because “you have to show the future without actually building it.” and other CI companies use different machine-learning models for understanding conversation based on the application. As Rathore says, “There’s GPT-3; there’s BERT; Meta just released one recently. “There are so many language models being built every day, which is more power than before.” But even with a good language model, she says, “you still have to do a lot of work on it to actually build a system that understands what it needs to do, based on the use case. I think that’s where the tricky part comes. You could solve a problem by just building transformers to do classification for audio and video and text, or you could build generative algorithms on top of GPT-3 and write a document or write a blog post.”

CI systems not only need to “understand” the conversation, they need to communicate with users about the conversation, too.

“There are multiple problems in the pipeline,” Rathore says. “One is capturing — ingesting conversation data from different channels and normalizing it to a format that can be understood. One is understanding, and then it’s analyzing, drawing inferences and generalizing something that will capture the essence in a more abstracted way. You can work on any one of the problems and provide value, but we strive to provide an end-to-end platform or experience.”

Customers now using tools include the intelligent interview and recruiting platforms and; webinar companies Hubilo, Airmeet and Remo; and “some pretty big enterprises in fintech and communications,” says Rathore. has been based in Seattle since 2019, when Rathore and her co-founder and CTO, Toshish Jowale, graduated from the Techstars accelerator program here. They now employ 27 people in and around the city — plus 40 more in Pune and Bangalore, India.

“[In Seattle] we’re plugged into a community of awesome founders, and there’s a lot of AI talent and a lot of AI focus,” Rathore says, ”including the AI2 Incubator and AI focused local VCs like Flying Fish. There are lots of dev tool companies here, too — Twilio, Stripe, Auth0, Outreach and Highspot all have a Seattle presence now.” And companies like, Rathore says, enjoy a symbiotic relationship with local Goliaths Amazon and Microsoft.

“There’s a circle of talent that keeps shifting as entrepreneurs come out of these companies or Facebook to join startups and build new companies,” she says. “That’s why I think the place is really thriving.”