Viable aims to quantify qualitative customer feedback with AI
source link: https://venturebeat.com/2022/05/12/viable-is-quantifying-qualitative-customer-feedback-with-ai/
Go to the source link to view the article. You can view the picture content, updated content and better typesetting reading experience. If the link is broken, please click the button below to view the snapshot at that time.
Viable aims to quantify qualitative customer feedback with AI
We are excited to bring Transform 2022 back in-person July 19 and virtually July 20 - 28. Join AI and data leaders for insightful talks and exciting networking opportunities. Register today!
There is an implicit assumption in most analytics solutions: The data analyzed and the insights derived, are almost exclusively quantitative. That is, they refer to numerical data, such as number of customers, sales and so on.
But when it comes to customer feedback, perhaps the most important data is qualitative: text contained in sources such as feedback forms and surveys, tickets, chat and email messages. The problem with that data is that, while valuable, they require domain experts and a lot of time to read through and classify. Or, at least, that was the case up to now.
This is the problem Viable is looking to address. Viable, touting itself as the only qualitative AI company to provide natural language querying of customer feedback, announced today the closing of a $5 million fundraise primarily for growth, R&D and new hires.
Viable’s CEO and cofounder, Dan Erickson, detailed the company’s origins, differentiators and its qualitative approach to customer feedback with VentureBeat.
From product-market fit to customer feedback via NLP
Erickson, an engineer by trade, cofounded Viable with his identical twin brother Jeff, who is a designer. Both have been in the tech industry for about 15 years, having skipped college to go straight into business early on.
The two have held senior roles at various startups and their career paths have intertwined off and on throughout the years, meeting in the middle at product management as Erickson put it.
Eventually, the Erickson brothers decided to start their own business, focusing on tackling the product-market fit problem which had tantalized Dan over the years. That was the start of what was initially called Viable Fit.
The team built a product to help people run what is known as the “superhuman product-market fit process.” The process is centered around a survey, followed by some analysis to help founders and product owners figure out a roadmap for their products.
In order to make that work at scale, the Viable team developed proprietary natural language processing (NLP) technology. They quickly found that this turned out to be the most valuable part of their entire approach.
Gaining traction: Push and pull
Viable gained traction among companies much larger than the traditional finding product-market fit company and the Erickson brothers decided to pivot and focus on their NLP models.
Viable stopped measuring product-market fit and began focusing on aggregating customer feedback across channels. Viable’s platform also offers a full analysis service that provides written analysis on top of the feedback. That recipe can be applied to areas such as product management, customer experience and marketing.
The analysis Viable offers can be accessed in two ways — push and pull. For the push mode, a report is sent on a weekly basis that covers what happened in your customer feedback in the last week. The report includes things such as the top complaints, compliments, questions and requests from customers. The report’s extent ranges from a dozen to a few hundred paragraphs.
Typically, when people read those reports, they have questions they need answered in order to act. Viable helps them do that by offering a natural language question-and-answer system. Users can type in a question about the data and Viable provides an answer, all in plain English.
In addition, the company offers out of the box integration with several sources such as Zendesk, Intercom, Delighted, iOS App Store, Play Store and Front. It features custom integrations via Zapier, as well as the ability to ingest data via .csv files. There are different subscription levels for the service, depending on the number of data points ingested.
Under the hood
It may sound simple and obvious, to the point of having to wonder “how come nobody else did that before.” After all, Viable uses OpenAI’s GPT-3 under the hood, so in theory, anyone — including the Zendesks of the world themselves — could have done it. The answer is twofold.
First, Viable has a head start, since it started in 2020, just when GPT-3 came out. As Erickson shared, they were among the first to work with some of GPT-3’s capabilities in a commercial setting. Second, part of Viable’s value proposition is precisely the fact that it integrates data from many different sources.
In fact, Viable is much more than a thin wrapper around GPT-3. The company uses many features of the OpenAI API, including embeddings, as well as the actual GPT-3 completion engine. But Viable also has its own models that work with GPT-3, that have been trained and fine-tuned throughout the last two years.
The company also has its own data repository, as well as its own ingestion pipeline. Whenever a new piece of content is created, it’s pulled in, along with any metadata that may be available. From there, it goes into a pipeline consisting of different models that Viable has developed, along with some GPT-3 functionality that will classify the piece of text.
The classification process figures out whether the text is a complaint, a compliment, a request or a question. It also identifies different topics within the text and performs some sentiment analysis, emotion analysis, urgency analysis and noise detection.
The platform is geared towards text analysis and can’t directly connect to sources such as databases or spreadsheets at this point. However, it can use what Erickson called “customer traits” to slice and dice the data.
Those may include job titles, locations or even numerical answers to multiple choice questions, such as “how many times a week do you use the product”. Users can then have the system perform tasks like “generate a report for my product manager enterprise customers in the Bay Area who use the product one to two times per week.”
Erickson said that Viable has developed an unsupervised system for thematic analysis based on GPT-3 embeddings plus a proprietary thematic analysis engine on top, which he characterized as state of the art. That means the system does not have to be provided with any context as to what kind of things it’s looking for other than requests, questions, compliments and complaints — so it can function in any domain.
Boundaries for avoiding bias and toxic language
GPT-3 may be one of the most impressive feats of engineering and AI, but it’s not without its flaws. Two of the most famous ones, which would render its use problematic in a commercial setting, are toxic language generation and hallucination — i.e., generating authoritative-looking answers that aren’t based on facts. As Erickson shared, Viable has managed to circumvent those via custom training.
“We’ve built out thousands and thousands of training examples for things like, what does it mean to summarize a theme? What does it mean to name a theme? How does that all work? And we’ve basically built out a fully fine-tuned version of GPT-3 that keeps it on the rails. So, it’s got sort of a more limited language set that it’s using. So, it’s not going to do any of those curse words or anything like that,” Erickson said. “Then on the hallucination side, we have done a meticulous job of building out that training data set to make sure that every example that we pipe in is only directly using facts from the feedback that is piped into it. And that way it basically tells GPT-3 — Hey, I don’t want you to be creative here. I want you to just report the facts and that’s exactly how it works.”
Beyond GPT-3 and customer feedback
The above should be valuable free advice to anyone aspiring to build a business around something like GPT-3. Not only in terms of how to circumvent its shortcomings, but also in terms of how to add value on top of it. As Erickson said, the cost of using GPT-3 is baked into Viable’s price points, as well as things such as their other processing costs and a healthy margin.
That must have worked for Viable’s investors. Streamlined Ventures led the $5 million round due to its interest in applied AI, with participation from previous investors Craft Ventures and Javelin Venture Partners. The round also included investment from Merus Capital, GTMFund, Stratminds, Tempo Ventures, Micheal Liou, Bill Butler and Samvit Ramadurgam. Viable’s total funding to date is now at $9 million.
The company has about a dozen paying customers and a total headcount of nine employees at this time. According to Erickson, the company has a few high-profile clients who are pleased with the product and Viable has made the move to expand beyond customer feedback.
“We work for any kind of experience — whether it’s employee experience, partner experience, customer experience, it’s really all about helping people analyze the qualitative nature of those experiences” said Erickson.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Learn more about membership.
Aggregate valuable and interesting links.
Joyk means Joy of geeK