Chatbots? AI? Virtual Assistants?Why Not Everything?
Part of the reason I love behavioral economics is that replicating human actions is really hard. You can’t just stick a bunch of inputs in an algorithm and have it generate the next action someone is going to take. Even a process as straightforward as mimicking human motion has taken many years to fine-tune to become sufficiently “human”. Disney’s animatronics are a great example of the effort it takes to create human like behavior. From the debut of human animatronics in 1964 (President Lincoln in the Hall of Presidents) to the first upgrade to the A-100 model in 1989 (the Wicked Witch of the West), the robots were too perfect. They would move body parts in isolation and stop on a dime — just perfect for an industrial robot but not great for a human mimic. The A-100 models for Disney actually overshot and added deliberate imperfections to better mimic human behavior and make them more believable.
All that effort was put in just to make robots move in a fashion that’s more lifelike, and yet counter intuitively, it required a step to add in human-like mistakes. Virtual assistants are running into much the same problem today, where the voice is too monotone and robotic (Alexa, Cortana, and Siri don’t elicit the same reactions as a human does) and it’s simply not human enough. And from personal experience, I don’t expect virtual assistants to operate in a flexible manner the way I could expect a coworker to. So, when Google launched their Duplex service back in 2018 as a function on their new Pixel phone, it changed the virtual assistant (compared to Cortana and Siri) from being table stakes to being an actual point of differentiation across these brands.
Google Duplex is Imperfectly Human
Take a look the Demo above from Google’s 2018 I/O conference. Notice how the Duplex service is able to pivot when it encounters questions, like the restaurant host saying they don’t reservations for fewer than five people. It also is able to sprinkle in filler words like “uhm” and have multiple voices with variation in speech cadence and tone. For all intents and purposes, people are having trouble distinguishing this chatbot from a real human, so it sounds like Google Duplex is a true virtual assistant. However, this isn’t quite the case as Google Duplex is simply an API that sits atop of the Google Home or Google Assistant application layer. Instead, when I ask Google Assistant to make a restaurant reservation for me, it will first look for online options via Rezy or OpenTable, and only call as a last resort.
The chart above outlines how Google Duplex works at a high level, with speech being converted into words with their Automatic Speech Recognition (ASR) system (like dictation), then converting that text into intent signals via their Natural Language Processing (NLP) algorithms. Google will then figure out what the speech actually means, and how to fulfill the original request to Google Assistant.
Duplex, However Cool, Might Not Be Driving Sales
If that seems like a LOT of effort to simply make a phone reservation at a restaurant… it absolutely is. However, that’s not to say that this isn’t on brand for Google. As with most technology firms (like Facebook), Google is highly focused on making sense of data and using that information to generate revenue by serving up ads. It’s only offered on existing Google products such as the Pixel and the Google Home products, and while Duplex is expanding to other platforms that allow Google Assistant (like iOS), it’s hard to determine if it’s actually a reason that people are picking Google phones over anything else on the market.
After all, there’s not a body of research floating out there around the percentage of sales that were being driven by this singular feature. Realistically, it’s hard to pinpoint how much incremental revenue is being driven by a new feature without a test market to act as a control. While Google Duplex is certainly an interesting study in the combination of AI and Chatbots in the handy form of a virtual assistant, there’s very little tangible value for the company to realize through sales.
But sales may not be the end goal here. Starting from the launch of Duplex, there has been speculation that Duplex may be used as an enterprise service to reduce reliance on call center agents, or that it may replace low-level sales cold callers. Google may be relying on consumer use cases to train the speech recognition models (cost of failure here is disappointing the user instead of a breach of contract) before rolling it out broadly for commercial use. At the end of the day, Duplex will need monetization to provide proper return on investment, and it’s unclear what that monetization will look like.
