Erin Kelly

Most experts thought that Brexit was a long shot. But not Erin Kelly, a CPA and CEO of Advanced Symbolics, an Ottawa-based company that uses artificial intelligence to predict outcomes and help change behaviour. (Photo by Derek Shapton)

Features | From Pivot Magazine

If you could read my mind 

CPA Erin Kelly is pioneering artificial intelligence that knows what you’re thinking without even asking. Is this the future of polling?

A Facebook IconFacebook A Twitter IconTwitter A Linkedin IconLinkedin An Email IconEmail

On June 23, 2016, much of the world was glued to their televisions or obsessively checking their news feeds to see whether British voters would choose to leave the European Union. Most experts thought that Brexit was a long shot. But not Erin Kelly, the CEO of Advanced Symbolics, an Ottawa-based company that uses artificial intelligence to predict outcomes and help change behaviour. She knew that “Leave” would win. AI had told her so. 

A few months before the referendum, Kelly, a CPA, began having her AI platform—which she’s named Polly—read newspaper articles about Brexit so it could learn the major players and stay up to date on issues that could affect the vote. For most of that time Polly was predicting a win for “Remain.” That started to change on June 16, when the “Leave” and “Remain” campaigns said they would take a three-day break to honour Jo Cox, a Member of Parliament who was assassinated that day. 

It was during that break that the computer said “Leave” would win with 52 per cent of the vote. Which is exactly what happened. “We were like, ‘Oh my God, what’s going on?’ ” says Kelly. Because AI can analyze millions of data points in a near instant, including, in this case, news stories and social media sentiment, it was able to pinpoint the break as a turning point. Not only that, but by running algorithms after the vote, Kelly was able to determine why things flipped. “It was because of a drop-off in conversation,” says Kelly. Had Cox not been murdered, and had there not been a campaign break, the vote might have been “Remain,” she says.

By that point, Polly had been programmed to look back at a year’s worth of data, and it knew that when engagement declined, undecided voters were less likely to hear pro-EU messages and more likely to vote for Brexit. So, when things quieted down before the election, it knew, based on patterns in the data, that “Leave” would win.

It was a fascinating study on the predictive powers of AI, Kelly says, and along with a correct prediction on the popular vote in the U.S. election, it put her budding polling firm on the map. Advanced Symbolics is a 15-person company she founded with Kenton White, chief scientist and developer of the AI. It’s not in the business of election polling—they follow politics because it generates interest in the company—but she does work with policy makers, companies and foundations to help them better understand human behaviour. 

Companies tend to approach Advanced Symbolics for the same reason they would any polling firm: they want to know what the public thinks. Unlike traditional polling companies, however, Kelly’s business uses social media to measure public opinion. Polling companies, she says, often can’t get a truly representative sample of a population—the rise of cellphones means phone polls get less precise geographic and demographic information on those who answer, and only a certain kind of person tends to answer questionnaires in the first place. People aren’t always honest about what they think or do, either. 

Advanced Symbolics scours social media mentions—it looks at aggregate data without storing any personal information—and gathers huge amounts of intel, cross-referencing it against news stories, academic journals, websites, Statistics Canada and whatever else is relevant to come up with an answer to a company’s question. 

A more advanced machine like Polly will also give CPAs tools they’ve never had before.  

Once clients come to understand the power of AI, Kelly says, they often want Polly to look at more complicated issues, things they might never have thought to ask about. For instance, the company recently completed a major transit study for a Canadian city that found people are using transit much differently than it thought. The city (which Kelly can’t name because of a non-disclosure agreement) originally wanted to find out how people move around where they live—by car, bus or other modes of transportation. When they saw what AI could do, they asked if Polly could help plan optimal transit routes. It did, by looking at millions of social media messages, such as Twitter and Facebook posts, likes and mentions from within the city. And not just messages that said, “I’m taking a bus to work,” but also posts that referenced an activity (going to the movies) or a location (the cinema) without ever mentioning transit.

Phone polling would not have been able to capture some of the nuances in how people travel. Polly found, by parsing social media data, that single people aren’t always going to sleep at the same address—they might stay over at a boyfriend’s or girlfriend’s house—and so they’re commuting from a different area than a phone poll would have suggested. In areas with large single populations, that could affect transit planning in a big way. Thanks to the information that was gathered, the city is reworking many of its routes and adding transportation to areas that were busier than originally thought. 

What the company does is complex, but at its most basic, it’s looking for patterns in something to predict what might happen next, Kelly says: “Something might not be an issue now, but it might be six months from now. We want to know when something will go viral or be a trend.”

Kelly’s career path from accounting to AI is not as odd as it may first seem. She studied accounting, she says, because she loves analytics and because “at the end of the day everyone needs an accountant.” As she went through the CPA program, though, she realized she wanted to work for herself. Her father had an entrepreneurial background—he owned several pub-style restaurants—and she was drawn to the do-it-yourself lifestyle. The CPA designation helped, too. “Having an accounting degree and a parent who runs a company is the perfect combination,” she says. “I know the technical side of running a business, while having entrepreneurs in my family gives me the emotional background.”

After getting her CPA, Kelly joined an advertising firm, which she enjoyed, save for one thing: no one really knew if what they were doing was having an impact on their customers. She’d look at data to see, for instance, if people bought more of something after an ad campaign went live, but she could never know for sure if those campaigns were changing behaviours. She tried to create an analytics practice at the firm, but there was little interest.

In 2014, she started her own data analytics firm using existing technology. At the same time, White, a computer scientist who spent years developing Polly, was starting his own analytics firm. The two were introduced by a mutual friend and decided to join forces. Kelly had the business and marketing acumen to grow a company; White had the technical knowledge to create a powerful prediction machine. Advanced Symbolics was born. 

Kelly sees AI as her company’s critical competitive advantage. While other firms engage in “social media listening” to find out what people are saying about a topic or product, she says, that approach starts with keywords or company names. Polly’s approach starts with developing a representative sample of a population. AI can parse an almost endless number of social posts and data points and it can fact-check information against academic research and census data to ensure that what it is gathering is truly representative.

AI can tell a person’s age, how much they make, even if they were drunk-posting. 

It can also make connections between words and posts that no human would think of connecting and conclude, based on information about a user, what the person does, where they work, their interests and more. Polly knows that someone named Madison was likely born between 1995 and 2005 because census data shows that more Madisons were born in that period. It can gauge the probability that someone is a certain age or ethnicity from a photo, and can deduce from publicly available messages how much someone makes, how they commute and even, based on how something is written, if a user was drunk-posting. 

Kelly emphasizes that Polly does not collect personal information, and that nothing can be traced to an individual. It does look at names to help determine gender, but she insists it doesn’t store that data anywhere. “We never see it, we don’t even know how the name factors into what Polly is doing,” she says. They use k-anonymity and differential privacy, two methodologies used by major organizations—Statistics Canada uses the former, Apple the latter—to process data without also gathering personal information.  It can be hard for humans to understand just how this all works, as only a machine can make millions of connections between seemingly disparate data points. Polly, like other AI platforms, can also learn new information about a topic, which enables it to make even more connections and predictions. For instance, with the Brexit experiment, Kelly fed Polly five newspaper articles so it could identify, in future stories, who the main players were. She then asked it to read as many articles as it could on its own, which then helped connect Brexit-related events to the millions of social media messages it was monitoring. 

Still, Polly has its limitations. It needs at least 10,000 people within a target market to be talking about a topic—at which point messages start appearing online—before it can learn anything. So Advanced Symbolics doesn’t work on small, niche issues—Brexit is fair game, but a company wanting to poll its 200 employees is better off using an old-fashioned phone poll. More notably, it can’t account for people who don’t post to any social media at all. Kelly insists, however, that so many do that her company can still get a representative sample of the population.  

About half of Advanced Symbolics’s work is for the private sector—it counts Disney, Cadillac Fairview and Fidelity Investments among its clients—and half is with the public sector. In the recent Ontario provincial election, Polly’s analysis included 1,069 Indigenous people, many of whom are missed in traditional polls because they are a smaller subset of the population, and many don’t have landlines. “They are online much more than on landline phones,” says Kelly, and in any case, they typically refuse to take phone surveys. Kelly argues that if oft-ignored minority groups are included in polling data, even if they’re unaware they’re being listened to, their views will matter more to politicians and the result will be better policy.   

Ultimately, Kelly says, her work is meant to help clients deliver what people really want, including CPAs. Already, AI is replacing some basic accounting functions, like bookkeeping, and the more advanced it gets the more mundane tasks it will do on its own. That will allow CPAs to focus on business strategy and other added-value work, she says.  

A more advanced machine like Polly, which can help predict trends, will also give CPAs tools they’ve never had before. “CPAs want to be seen as part of the executive management team, where they can guide policy and help companies look at trends,” she says. “AI is going to be helpful for that.”

Polly’s goal is to pinpoint, in a way humans can’t, why people do what they do. 

Advanced Symbolics is also in areas that go beyond corporate concerns. The company works with a client, whom Kelly can’t name, on suicide prevention. Polly, she says, can tell through social media traffic if a community is more likely to have a rash of suicides. If, say, a plant closes in northern Quebec, it can see if there are more suicide-related messages coming out of that area. If there are, organizations could then send counsellors to the area before a crisis occurs—not after, as is usually the case. Of course, it can only know when a group is at risk, not an individual, because it doesn’t collect personal information.  

Essentially, Polly’s goal is to pinpoint, in a way humans can’t, why people do what they do—and then figure out how to change those behaviours. Currently, it does not offer recommendations on how to take action based on the information, but it would like to at some point soon. “Our hope is that we can change behaviours, but we haven’t done that yet,” says Kelly. “Right now we’re predicting them but we haven’t started the work on how to prevent them. What we are trying to do is help people get the help they need. There are different tools for different people, and we’re trying to figure out the tools that help people do what they want to do.”  

Advanced Symbolics is only three years old. New funding from former Ottawa mayor and tech entrepreneur Larry O’Brien and the Kanata-based Capital Angels Network will help Kelly and White pursue work in the U.S. They are also looking to create an interface that would allow clients to come in and use Polly. It might even make politicians obsolete one day, says Kelly. “If AI can tell me what the population wants, then I don’t need [a politician] to tell me,” she says. “We can go straight to the people.” She’s joking—mostly. But the idea is a profound one: as Polly and other AI platforms improve, they’ll know more about what people need and want better than anyone else.