Cogito

JOSHUA FEAST: So my name is Joshua Feast. I'm the CEO and founder of Cogito Corporation. And what we do is help people be more charming in conversation.

So I was lucky enough to attend MIT and encounter a stream of research that had been developed at the MIT Media Lab. And to how one can understand human behavior and interpret psychological states.

And, for me, that combines several threads. First, I was very interested in the problem of how do you take a science, turn into a technology, and then from a technology into a business?

And secondly, I was really interested in the notion of caring and what does caring mean? How to have more of it? How we can build better relationships with each other. And then what technology can do to help us with that very human element?

When the stream of research had already shown the ability from sort of a science perspective to look at voice and understand behavioral markers of distress, it was sort of where the research was at. And it also had been a whole variety of side posts that showed that you could also use voice and behavioral feedback to improve engagement in conversations.

So that, from a scientific perspective, when I was involved that had been developed. And it was developed in the human dynamics lab led by Professor Sandy Pentland. Who's a well-known professor. He's a co-founder of Cogito.

What we did subsequently was take the scientific papers and concepts and went out and got funding to do R and D for several years. And we received funding from DAPA, the high tech funding under the DOD, a number of grants from the NIMH, the National Institute of Mental Health, and a variety of other organizations.

And that funding allowed us to take the scientific concepts and develop a running technology. And after that, it was the job of commercialization. And the job of commercialization we initially focused on clinical use cases, essentially helping nurses and psychologists recognize distress in vulnerable patient populations. And we still have quite a lot of work we do in that area.

And then the technology got to a point and the product got to a point where it was available to be used for a very wide use case. And so at that point we focused on expanding our targets to help with large scale sales and service and enterprise. So how do you solve the problem of getting thousands and thousands and thousands of frontline service professionals to be motivated and stood to speak to customers how you want. That's sort of where we are now.

The software is basically a real time conversation coach. So our basic thesis is that a conversation between humans doesn't just have to be audio going back and forth anymore. It can be audio with an intelligent assistant that's sitting with you in the conversation that can help you build a better relationship, be more successful in that conversation. So that's what we do.

And in practice what that means is that our system is able to look at a conversation as it's happening. And by looking at behavior patterns, understand how well that conversation is going. So that's the first thing. Of course all conversations, as they're happening, know how well it's going.

And then because we know how well the conversation's going as it's happening we can provide guidance during that conversation that helps people end up with a more successful outcome and a better relationship.

And that guidance-- more in terms of how we apply it. So right now our big focus is on helping very, very large teams of sales and service professionals interact with customers more successfully. So I think 10,000 customer service representatives all who have a real time coach, which helps them be more emotionally intelligent and then provide a much better experience to the customer.

[MUSIC PLAYING]

SPEAKER 1: So like many things, Cogito is on the shoulders of giants. And when Cogito got started, it was a collaboration between myself and Professor Sandy Pentland, who is one of the original professors involved in the starting of the MIT Media Lab. And he had developed a theory, which he called honest signals theory, which is a way to understand and interpret human behavior and understand psychological state.

And he had developed-- over the course of eight or nine years-- a variety of scientific observations and a set of experiments that pointed out what one can look at in behavior to understand, for example, how well a conversation is going. Or understand if somebody is probably experiencing distress. Or if somebody is interested in you in a conversation.

So he had developed that from a scientific basis. And I was fortunate enough to see that science at a very, very early stage. And I got very excited about the development potential. And at a personal level, I was personally very interested in the problem of how do you take a basic science, create a technology, and then create a commercial organization out of it.

So we've been-- as I said-- not only on the shoulders of giants in terms of the technology. We have also benefited enormously from the community around us. Cogito has basically been a company in three phases. So the first phase was literally four years of applied R&D. We were sponsored by DARPA, which is the high tech funding under the DoD. We were very fortunate to receive a number of substantial grants from the National Institute of Mental Health. And that really gave us the capital required to go from science to a novel technology.

In the course of doing that, we also benefited from a whole range of mentorship. We've been enrolled the MIT venture mentoring service since the start of our existence. And that's been extremely valuable. We've benefited from the MIT entrepreneurship center. And that is the Martin Trust Center, attached to the Sloan School. And then, of course, we've also been very fortunate-- as the technology became more developed, and it was ready-- to have the benefit of our relationship with the Industrial Liaison Program at MIT, which has helped build relationships with large customers.

So today Cogito is an expansion-stage software company. And what that means is that we already have a variety of very large, well-known customers. That product is delivering substantial value to those customers in terms of generating a true win-win-win. And what I mean by that is our organizational customers get the benefit of more satisfied customers of theirs, because they have better interactions. End users-- end customers get much more pleasant conversations when they call in, which makes them much, much happier.

And then, critically, with this application-- it's also something which is a true win-win for employees. Because it's helping with the most challenging part of being a front-line service professional, which is how do you deal with difficult and emotional customers? So it's a real win-win-win. And we're getting a really tremendous amount of traction today.

We are also very, very grateful. We were recently allowed into the STEX 25, which is a program that's being set up by the Industrial Liaison Program where they've selected-- specifically-- 25 high-profile startups that are doing well in the market. They're giving us a higher profile and putting additional resources behind, allowing us to interact with--

[MUSIC PLAYING]

JOSHUA FEAST: So today in large-scale customer service, all calls are already monitored-- already recorded. And what happens is that typically, a supervisor will randomly sample, I would say, one to five calls per month per customer service agent. And they'll listen to them.

And that's really what's, from a time perspective, what's actually possible today. And then based on that random sample of a small number of calls, they'll give feedback to the agent.

So there are a number of problems with that. One, compared to the volume of calls per agent that are actually going on, it's a very, very low sample rate. And the feedback being given to an agent is happening very late. It's happening weeks after the phone call happened.

And often, the feedback is very subjective. So I don't think you listened very well. I think I listened fine. Where do we go with that? [LAUGHING] Right? And so that's sort of the state of the art right now.

So what Cogito is doing is we are providing objective measurement about how conversations are going as they are happening across every single interaction. So the information understanding of how well one service operation is going is night and day.

Secondly, the software is actually doing the job of coaching to a certain extent, because the software is actually in the conversation providing hints and tips, providing alerts, providing notifications to the customer service agent during that conversation, as if they had an incredibly rapid supervisor with them on every call.

And the customer service agents find it incredibly reassuring. And for them, it's a real breath of fresh air, because suddenly everything is objective. And in terms of what's going on, they know they've been doing a good job, but that might not have been obvious to a supervisor who randomly sampled just a few calls.

And additionally, really critically, the software provides a means to improve. Everybody wants to be good at their job. But it's not easy if you're doing 50 calls a day and half the calls are, difficult customers.

There is a lot of share in the industry. There's no doubt about it. And there are several reasons for that. So one is immeasurability So if you're in a job and you can't get any feedback into how well you're doing, you get burned out very easily.

So one of the things we address is you can immediately know the difference you're making as a customer service agent across all your interactions. You can see how well you're doing. You can see the customer's response to that. And that's, I think, something that's very, very important.

The other thing is that sometimes, people leave because they don't like their managers. That's also a reason people leave employment. And unfortunately, in a customer service environment, managers simply don't have the time to do a lot of coaching and a lot of supervision because they have a really high number of direct reports.

And so the software is coming in and really giving the manager extensibility so that the employee feels much better looked after. They have a feedback, and they have a means to improve. That's being through the software.

JOSHUA FEAST: In terms of how the technology works. So a phone-based operation has many, many, many thousands, many millions of phone calls going through it. And what we do is we stream those conversations to our cloud system.

And then as those conversations happen, we're looking at behavioral signals in those conversations, and then providing feedback to the front line customer service representative. So it's really a three-way, real-time situation.

In terms of how the technology works. So at the base level, we're taking many hundreds of measures across the voice spectrogram. And then we're synthesizing those into behavior models. These are specific actions or things that one is looking for that are psychologically relevant.

So for example, pressured speech. If somebody is suddenly very agitated and (IMITATING PRESSURED SPEECH) their speech picks up (NORMAL VOICE) like that, that's an example of a behavior that there may be interest in, and something that you don't really want your customer service representative to sound that way because you want them to be calm, and reliable, et cetera. Right?

So that's an example of a behavior. Then from a behavior, from the behaviors and the signals, we have performance measures, which is larger term concepts, like is somebody being empathic? Is this conversation going well? Is somebody under distress? So that's sort of a higher level model.

And then beyond there is the predictive component where we say, well, this conversation didn't go well. Therefore, this customer may be at risk of churning. And therefore, this is a customer that you may need to reach out to offer them something to keep them in the fold. So that's sort of the example of the analytic pyramid.

Customers are able to define particular business roles and key word patterns that are relevant to their specific operation. And when those patterns are spotted, that are creates a specific notification for an agent. So that's sort of one part of it.

The second part of it is this is a learning system. So when we're given access to ultimate core outcomes, we can use that continuously to refine the models within our system, how accurate our measures of quality are as they relate to a future business outcome.

One of the things that's really neat about having this type of system is that our engine is seeing millions of phone calls go through our system as they're happening. So millions of human conversations. And we are affecting those conversations as they're happening by giving notifications to one or potentially multiple parties in that conversation.

So we always know if we've given a notification if it made any positive difference or not. And we can continuously learn from that. So that's a very interesting system from a technical, machine-learning perspective.

And there is a lot of theory around this. And I would like to note, again, on the shoulders of giants. This is a lot of theory that Sandy Pentland developed originally. But it turns out that most of the important signals in a conversation are actually universal properties of humans, rather than local properties of language and culture.

And the reason for that is that if you want to communicate attitude or distress or these types of concepts, they're actually rooted more in ancient brain systems that were developed pre-language. And that's one of the reasons the system is so reliable in terms of its understanding of how well things are going.

So what that means in practice is that our measures of, say, for example, how well something's going, or our measures of how you should behave differently to create more rapport, are incredibly extensive. We run them in English, absolutely. We run them in English as a second language. We run them in European languages.

So already we're deployed internationally today. And there's really no algorithmic difference for those particular feature sets. The exception to that is keyword spotting, which is extremely local and requires configuration.

JOSHUA FEAST: I think, if you can do it in real time, real time's better. It just is. It's just less management over here. There's less to do. It sort of just works if you can do it in real time.

This type of system solves problems that exist with sort of voice of the customer surveys. Typically everybody's experienced this. You call into a call center. There's will you take a survey on how well we've done and it turns out that not many people respond to them.

And when they do respond, the responses are highly skewed. Often highly skewed to the positive. Which actually means that the customer service managers can be under the impression that their customers are 98% satisfied with them until they get an industry survey. And they're on the bottom of the queue, right?

So there are obviously clearly a number of problems are well-known with surveys. And this does mitigate those problems to a major extent. Because you get a comprehensive and objective understanding of how well your operations are going.

That being said, surveys absolutely have their place. There could be special, private information that a customer shared-- they would share in the right context, which is not exhibited by their behavior. So this should be viewed as a sort of a new generation but also a compliment to existing survey-based voice of the customer processes.

Now we are sometimes asked well, will chat bots replace humans in the customer service arena? And that's a question we get asked a lot. The answer is nuanced.

AI, as it's currently developed, is actually pretty good at solving, on its own, very well-formed problems. The sort of problems that probably already have a website for them that you could fill in. Say I can go to a website. I can check my balance. I can go to a chat bar and I can ask could I please check my balance. And it will respond.

And really that's kind of where the technology truly is at. The sorts of problems that our customers deal with questions of health or finance, tend to be questions about people's lives that are much, much more complex than the current state of AI can truly handle on its own today.

For example, many problems in customer service are resolved by question answer co-evolution. I ask you a question. You respond. I ask another question. And we kind of figure it out together. Computers aren't great at that yet.

Computers also add greater abstract thinking. If I want to tell you about I want to get a debit card from my son, a computer doesn't really understand the relationship I have with my son and how I feel about him. Right, so there's a lot of sort of things that computers don't do today.

So where we're are really at as a society however, is that AI can extend our intelligence in ways that we never truly thought possible. And when you think about what Cogito's doing, it's amazing. Like, we are helping you, with an AI, create better relationships, have been more emotionally intelligent, and connect with your fellow humans. Who would have thought a computer could do that, right? That's amazing.

But we're also not trying to say that the human isn't important. And we actually want to talk to each other. We want to have relationships with each other. And so no matter-- even if we do get to this theoretical nirvana of a generalized artificial intelligence, I still believe we're going to want to talk to our fellow humans myself.

[MUSIC PLAYING]