Conversa Health


Thoughts on Google Duplex and Chatbot Design

May '18

Earlier this month, I watched Google CEO Sundar Pichai’s reveal of Google Duplex, an AI that can make phone calls to small businesses and schedule appointments on behalf of consumers. My colleagues were awed by how human-like her voice sounded, how she navigated the cadence of a conversation with a real person so naturally. They were wowed by the following exchange during a call to a hair salon to make an appointment:

Hair Salon –    “Hello, how can I help you?”

Duplex –          “Hi, I’m calling to book a women’s haircut for a client.”

Duplex  –         “Um. I’m looking for something on May 3rd.”

Hair Salon –    “Sure, give me one second.”

Duplex –          “Mm-hmm.”


Since then, ethical questions have surfaced around Duplex’s potential to fool people into thinking it’s human. There have also been questions as to whether the demo was staged or ‘improved’.  Axios, for example, pointed out that when they called hair salons and restaurants, the person answering identified their establishment and themselves and there was noticeable ambient noise, neither of which was apparent in the Duplex demo. As of this writing, Google hasn’t responded to the questions. I hope the demo wasn’t doctored. I’m going to assume it wasn’t.

I’d like to address a different issue.

The demo was impressive. I was amused and laughed along with my colleagues at the cute human vocal cues. But after the laughter, my first thought was that we should expect this kind of return on a multi-billion dollar investment in natural language understanding, deep learning, text-to-speech, and other technologies. My second thought was, some of the design elements are not required for the use cases highlighted by the demo.

I like to think about chatbot use cases along two dimensions – Intelligence and Humanness.  Intelligence is the smarts required to solve a problem or complete an activity. Humanness is about the traits necessary to make the person interacting with the chatbot feel like they feel when they’re interacting with a human.

The Duplex chatbot requires a significant level of Intelligence.  It’s smarts needs to be able to understand natural language and choose appropriate responses and/or questions in order to fulfill the objective of making an appointment for a service on a particular day and time that works for it’s consumer. I posit, however, that it doesn’t require a high degree of Humanness.

According to Sundar, the Google Assistant is intended to make the call seamlessly in the background. It’s making a call so the consumer doesn’t have to. So, although its human vocal cues are cool, they are irrelevant to the consumer as she won’t be involved in the call.

The salon person cares about understanding the caller sufficiently to be able to book her an appointment. Does she care if the caller uses cute human vocal cues? I don’t think so. Does she care if the caller’s voice is indistinguishable from a human voice? I don’t think so. This is a transaction. If anything, the salon person needs to impress the consumer, not the other way around. The call from the consumer needs to effectively and efficiently exchange information to get the appointment scheduled.

In terms of design, this means that the natural language understanding needs to be very good to understand what the salon person is saying. The deep learning needs to be very good to respond with the appropriate questions and answers. The natural language generation and cute human cues are unnecessary.  Of course, Google is a platform company, and its AI chatbot ambitions extend beyond making hair salon appointments, so this capability will likely be helpful in other use cases.

But, it is valuable to understand that the most impressive part of the Duplex demo is not in fact necessary for the problem it is showcasing.  In fact, I think the more interesting opportunity Duplex highlighted may be addressing the 60% of small businesses that don’t have online reservation capabilities through the consumer rather than through the business.

Why is this important? Let’s look at another domain, arguably one where the stakes are higher. There are many chatbot use cases one might consider in healthcare. There are opportunities to automate transactional tasks, like making an appointment, finding a doctor, locating where to park, and checking insurance coverage. These require modest Intelligence but virtually no Humanness.  The goal for both parties is to efficiently and effectively execute a transaction. Making the chatbot appear human is a waste of time and effort and will most likely ameliorate effectiveness and/or efficiency.

By contrast, palliative care requires a very high level of Humanness. The goal is to improve quality of life for both the patient and the family by providing relief from symptoms and stress of an illness.  The solution doesn’t require much of what AI is good at – data-driven decision-making. It requires a lot of what people are good at – empathy, caring, comfort, and judgment.  Trying to build an AI that is good at natural language understanding, gesture understanding, natural language generation, avatar or robot gesture generation, empathetic human cues, and gut feel seems like a tall order with too many points of failure.

At Conversa, we’ve designed our chatbot to meet the needs of care management. Helping patients and providers manage chronic illness, recovery after a post-acute incident, or prepare for and heal after surgery.  We believe these use cases require reasonably high Intelligence and Humanness. So we designed our chatbot specifically to address these requirements: Integration with the patient record so it’s knowledgeable about the patient’s clinical circumstances, profile-driven to personalize the conversation but specifically designed not to fool the patient into believing it’s human, structured dialogs to ensure pristine patient-generated health data (PGHD) integrity and to avoid the frustrating experience of not being understood by clinical natural language processing engines.

The takeaway – chatbots are easy; conversations are hard. Artificial conversations should be purposely designed to address specific needs. If you’re looking for a chatbot for your care management needs, we’d love to chat.