Slowly but surely, artificial intelligence is creeping into everyday life. Think of all the questions Siri and Cortana can answer; or e-mail programs that schedule appointments; or Facebook’s M service that aims to do things like make purchases and book travel arrangements on your behalf.
After decades of research on A.I. technologies, companies of all sizes are looking to cash in on the promise of virtual digital assistants—software “bots” that can interact with you, automatically handle certain tasks, and even make decisions for you. That last bit might concern you—as it should—but the tradeoff is having more time to spend on things that are important (or interesting). That’s the idea, anyway.
It’s just one piece of the overall A.I. puzzle—but it’s a potentially lucrative one. Virtual assistants sit at the intersection of business and technology areas like machine learning, natural language processing, mobile interfaces, and big-data analytics. Companies that figure out how to apply advances in these fields to assist people via e-mail, social media, or other platforms stand to gain plenty of users, clout, and revenue.
One new company in the field, Cambridge, MA-based Talla, is talking about its approach after half a year in stealth mode. The startup’s experience and prospects speak to how the landscape for A.I. companies may play out.
Talla is rumored to be raising a seed round in the neighborhood of $3.5 million, but the company declined to comment on that. It previously raised a $600,000 angel round last year from investors including Converge Venture Partners.
Talla’s founder and CEO is Rob May, the founder and former CEO of Backupify, a data-protection company that was acquired by Datto in late 2014. (May left Datto last June.) His co-founder and chief data scientist is Byron Galbraith, a recent Boston University PhD in cognitive and neural systems who has worked in software development and brain-machine interfaces. The two (pictured above) started collaborating on Talla last fall, May says, after meeting through Neurala CEO Max Versace.
May says the startup began with a couple of tenets: over the next decade, he predicts, “every knowledge worker will have a virtual assistant to help them do their job.” And, he says, “text will be the primary way we communicate with them.”
The first step, then, was figuring out which type of market to go after. “If you’re doing A.I., you have to think about how you’re going to compete with Google, Facebook, and Amazon,” May says. It’s important to “not rely on public datasets where those guys can beat you,” he adds.
That led Talla to look at internal corporate data and to try to “build a knowledge graph of what’s known in a company,” May says. The team set out to build A.I.-based software that integrates with messaging apps Slack and HipChat. These communication tools have gained traction with startups, media companies, and big tech firms as alternatives to e-mail. (Speaking of e-mail, other companies such as Mimecast are also working on managing knowledge across workforces.)
The virtual assistant that first resonated with beta users, May says, is one for handling recruiting and human resources tasks. The Talla bot interacts with a user via text on a smartphone (see screenshot below). It can do things like create a dossier on a job candidate, suggest interview questions to ask, and find similar candidates on LinkedIn. It can also manage tasks such as compiling interview notes from different people in a company, or responding to basic requests for information from candidates.
Talla can also answer simple HR questions from employees, such as, “What’s my co-pay for a dental visit?” The bot knows to ask, for example, whether it’s a routine cleaning appointment or something else, and adjust its answer accordingly, May says.
Under the hood, he says, Talla uses natural language processing techniques including word vectors (modeling how a word relates to others), and some deep learning (multi-layer neural networks that find patterns in large datasets). Roughly speaking, the bot tries to understand the meaning of a question and the context around it. Then it goes into its knowledge base to find an answer and, in some cases, perform a task—such as compiling or displaying information. It also learns as it goes, trying to make sense of all the different ways people may ask the same type of question.
“The hardest problem we face right now,” May says, is “trying to automatically map tasks to questions.”
Not surprisingly, Talla is working on other types of bots as well—for example, one for assisting a company’s social media and marketing efforts. May says Talla’s software is currently being tested by about 50 companies—the sweet spot so far is customers with 100 to 1,000 employees—and its public release is slated for July.
It’s early days indeed. Talla has just under 10 employees and is looking to grow judiciously. Given how big and open the market is for messaging apps and bots, May says, he would like to “focus on profitability as opposed to pure revenue growth.” Translation: raise a modest venture round, keep burn rates low, and start selling.
Tech investors not involved with the company see upside in A.I.-based assistants from smaller players. “I do see potential for [business-to-business] applications, and I think startups can compete,” says Matt Fates, a general partner with Ascent Venture Partners. “Perhaps not on the underlying science and infrastructure, but certainly in how it is applied to various industries and verticals. By narrowing the purpose and focus of the virtual assistant, it can also increase the accuracy.”
That bodes well for Talla—and other startups blazing narrow, focused trails in A.I. and machine learning. Around Boston, they include Gamalon, led by Lyric Semiconductor co-founder Ben Vigoda, in probabilistic programming; Semantic Machines, from Voice Signal Technologies’ Dan Roth, in conversational interfaces; DataRobot, led by Jeremy Achin, in data science; Indico, led by Slater Victoroff, in software development; Nara Logics, led by Jana Eggers, in recommendations and finance; and Sentenai, led by Rohit Gupta, in machine learning for the Internet of things. (Sentenai just announced a $1.8 million seed round from Boston-area investors today.)
Talla and its peers have their work cut out for them, of course. These are hard problems to solve. Virtual assistants, specifically, may seem like relatively low-hanging A.I. fruit—see X.ai for scheduling meetings, Spiro for improving sales, and Google Now for managing search and smartphone features. But there’s a reason they haven’t fully taken off yet.
“These systems will never be perfect, due to all the pieces that have to fit together to make it ‘correct,’” Fates says, “and then taking the intended and appropriate action.” He adds that “the cost of getting it wrong on the [consumer] side is fairly minor, whereas making a mistake on my business calendar with an appointment can be far more embarrassing or difficult.”