Viable aims to quantify qualitative customer feedback with AI

There’s an implicit assumption in most analytics options: The information analyzed and the insights derived, are virtually solely quantitative. That’s, they discuss with numerical knowledge, comparable to variety of clients, gross sales and so forth.

However on the subject of customer feedback, maybe a very powerful knowledge is qualitative: textual content contained in sources comparable to feedback kinds and surveys, tickets, chat and electronic mail messages. The issue with that knowledge is that, whereas beneficial, they require area consultants and plenty of time to learn by means of and classify. Or, not less than, that was the case to this point.

That is the issue Viable is seeking to deal with. Viable, touting itself as the one qualitative AI firm to offer pure language querying of buyer suggestions, introduced at the moment the closing of a $5 million fundraise primarily for progress, R&D and new hires. 

Viable’s CEO and cofounder, Dan Erickson, detailed  the corporate’s origins, differentiators and its qualitative approach to buyer suggestions with VentureBeat.

From product-market match to buyer suggestions through NLP

Erickson, an engineer by commerce, cofounded Viable along with his similar twin brother Jeff, who’s a designer. Each have been within the tech business for about 15 years, having skipped faculty to go straight into enterprise early on. 

The 2 have held senior roles at varied startups and their profession paths have intertwined on and off all through the years, assembly within the center at product administration as Erickson put it.

Finally, the Erickson brothers determined to start out their very own enterprise, specializing in tackling the product-market match drawback which had tantalized Dan over time. That was the beginning of what was initially referred to as Viable Match.

The group constructed a product to assist individuals run what is called the “superhuman product-market fit process.” The method is centered round a survey, adopted by some evaluation to assist founders and product house owners determine a roadmap for his or her merchandise. 

So as to make that work at scale, the Viable group developed proprietary natural language processing (NLP) know-how. They rapidly discovered that this turned out to be probably the most beneficial a part of their total strategy. 

Gaining traction: Push and pull

Viable gained traction amongst firms a lot bigger than the standard discovering product-market match firm and the Erickson brothers determined to pivot and deal with their NLP fashions. 

Viable stopped measuring product-market match and started specializing in aggregating customer feedback throughout channels. Viable’s platform additionally affords a full evaluation service that gives written evaluation on high of the suggestions. That recipe will be utilized to areas comparable to product administration, buyer expertise and advertising.

The evaluation Viable affords will be accessed in two methods — push and pull. For the push mode, a report is distributed on a weekly foundation that covers what occurred in your buyer suggestions within the final week. The report contains issues comparable to the highest complaints, compliments, questions and requests from clients. The report’s extent ranges from a dozen to some hundred paragraphs.

Sometimes, when individuals learn these experiences, they’ve questions they want answered with the intention to act. Viable helps them try this by providing a pure language question-and-answer system. Customers can sort in a query concerning the knowledge and Viable gives a solution, all in plain English.

As well as, the corporate  affords out of the field integration with a number of sources comparable to Zendesk, Intercom, Delighted, iOS App Retailer, Play Retailer and Entrance. It options customized integrations through Zapier, in addition to the flexibility to ingest knowledge through .csv recordsdata. There are completely different subscription ranges for the service, relying on the variety of knowledge factors ingested.

Beneath the hood

It might sound easy and apparent, to the purpose of getting to marvel “how come no person else did that earlier than.” In any case, Viable makes use of OpenAI’s GPT-3 under the hood, so in idea, anybody — together with the Zendesks of the world themselves — may have achieved it. The reply is twofold. 

First, Viable has a head begin, because it began in 2020, simply when GPT-3 got here out. As Erickson shared, they have been among the many first to work with a few of GPT-3’s capabilities in a industrial setting. Second, a part of Viable’s worth proposition is exactly the truth that it integrates knowledge from many various sources.

In actual fact, Viable is way more than a skinny wrapper round GPT-3. The corporate makes use of many options of the OpenAI API, together with embeddings, in addition to the precise GPT-3 completion engine. However Viable additionally has its personal fashions that work with GPT-3, which were skilled and fine-tuned all through the final two years.

The corporate additionally has its personal knowledge repository, in addition to its personal ingestion pipeline. At any time when a brand new piece of content material is created, it’s pulled in, together with any metadata which may be accessible. From there, it goes right into a pipeline consisting of various fashions that Viable has developed, together with some GPT-3 performance that may classify the piece of textual content.

The classification course of figures out whether or not the textual content is a criticism, a praise, a request or a query. It additionally identifies completely different matters inside the textual content and performs some sentiment evaluation, emotion evaluation, urgency evaluation and noise detection.

The platform is geared in the direction of textual content evaluation and may’t immediately connect with sources comparable to databases or spreadsheets at this level. Nevertheless, it might use what Erickson referred to as “buyer traits” to slice and cube the information. 

These might embody job titles, areas and even numerical solutions to a number of alternative questions, comparable to “what number of occasions every week do you employ the product”. Customers can then have the system carry out duties like “generate a report for my product supervisor enterprise clients within the Bay Space who use the product one to 2 occasions per week.”

Erickson mentioned that Viable has developed an unsupervised system for thematic evaluation primarily based on GPT-3 embeddings plus a proprietary thematic evaluation engine on high, which he characterised as state-of-the-art. Meaning the system doesn’t must be supplied with any context as to what sort of issues it’s on the lookout for aside from requests, questions, compliments and complaints — so it might operate in any area.

Boundaries for avoiding bias and poisonous language

GPT-3 could also be probably the most spectacular feats of engineering and AI, however it’s not with out its flaws. Two of probably the most well-known ones, which might render its use problematic in a industrial setting, are poisonous language technology and hallucination — i.e., producing authoritative-looking solutions that aren’t primarily based on information. As Erickson shared, Viable has managed to avoid these through customized coaching.

“We’ve constructed out hundreds and hundreds of coaching examples for issues like, what does it imply to summarize a theme? What does it imply to call a theme? How does that every one work? And we’ve principally constructed out a totally fine-tuned version of GPT-3 that retains it on the rails. So, it’s acquired type of a extra restricted language set that it’s utilizing. So, it’s not going to do any of these curse phrases or something like that,” Erickson mentioned. “Then on the hallucination aspect, now we have achieved a meticulous job of constructing out that coaching knowledge set to make it possible for each instance that we pipe in is barely immediately utilizing information from the suggestions that’s piped into it. And that means it principally tells GPT-3 — Hey, I don’t need you to be artistic right here. I need you to only report the information and that’s precisely the way it works.”

Past GPT-3 and buyer suggestions

The above ought to be beneficial free recommendation to anybody aspiring to construct a enterprise round something like GPT-3. Not solely when it comes to learn how to circumvent its shortcomings, but additionally when it comes to learn how to add worth on high of it. As Erickson mentioned, the price of utilizing GPT-3 is baked into Viable’s worth factors, in addition to issues comparable to their different processing prices and a wholesome margin.

That should have labored for Viable’s traders. Streamlined Ventures led the $5 million spherical attributable to its curiosity in utilized AI, with participation from earlier traders Craft Ventures and Javelin Venture Partners. The spherical additionally included funding from Merus Capital, GTMFund, Stratminds, Tempo Ventures, Micheal Liou, Invoice Butler and Samvit Ramadurgam. Viable’s complete funding thus far is now at $9 million.

The corporate has a couple of dozen paying clients and a complete headcount of 9 staff right now. In keeping with Erickson, the corporate has just a few high-profile purchasers who’re happy with the product and Viable has made the transfer to broaden past buyer suggestions.

“We work for any form of expertise — whether or not it’s worker expertise, associate expertise, buyer expertise, it’s actually all about serving to individuals analyze the qualitative nature of these experiences” mentioned Erickson.