Fin's Perspective on Google Duplex

Sam Lessin, 27 May 2018

Over the last few weeks, everyone from investors to job candidates to customers has asked us what we think of the Google Duplex announcement and if / how it impacts how we think about our service.  Having thought about and discussed this at length, here are three thoughts on the implications of Duplex:

Excitement, But Skepticism About The Technology (It Would Be Great if It Consistently Worked, But Probably Doesn't)

Nothing would make us happier - or frankly make our business better - than if Google had indeed gotten to the point that machines could make phone calls, wait on hold, and get answers from service providers automatically.

Our fundamental belief and model is that humans should focus on doing the most 'human' work, and machines should do the work that humans either don't want to do, or are not particularly good at. If Google Duplex reliably worked as well as it did in the demo and provided an API, we would immediately implement it as a layer in our stack in order to free our team to do work that required a higher level of human skill and attention.

But, if Duplex is on par with the real world / non-demo performance of other best in class voice assistant software (Google Assistant, Alexa, Siri), we will probably have to wait quite awhile before it's reliable enough to delegate mission critical work from our customers to it. It is one thing to do a cool demo of one of these products or try them out yourself as a consumer occasionally, but it is a much greater leap to adopt them when customers are paying you to do high quality work for them.

The Real Value is in Trusted Quality Assistance (Where Human + Machine Will Beat Machine for a Long-Long Time)

Having worked on human-computer hybrid assistance now for a few years, we are fully convinced that the number one most important characteristic is quality.

Quality means that you can trust the system to work every time, not just some of the time - which means you can delegate human level intelligence decisions and questions to it and be confident that the system will get the right answer.

As a user, if you have to question whether or not the system will get the right answer each time on something important -- like scheduling, or booking appointments, or travel -- you are usually likely better off just doing the task yourself.

The difference between asking Alexa to play a song and having it get the wrong track vs. asking an assistant service to make an appointment and having it book you the wrong thing all comes down to risk-reward.

If Alexa plays you the wrong song some of the time, it isn't the end of the world - so you are willing to use it for that action even if it is pretty inaccurate some of the time.

The bar on real world assistance is so so much higher because the stakes are a lot higher - so any solution purporting to give a fully technical solution to a real world problem will lose to a hybrid human + computer system that can drive higher quality and trust.

Social Challenges Are Hard To Overcome

It has been interesting to watch the range of backlashes to the demo - from those who think that human sounding robo-callers will create huge security issues, to those that think it will make the phone system unuseable...

What we see instead - coming from our lens of working on Fin - is an economic problem.

If you get computers making calls to humans to get answers at scale, you are likely going to dramatically decrease the effectiveness of the human-to-human call protocol.

Right now, the general assumption is that to make a phone call to another person you as the calling person are taking your own time / which you value.  So, there is a natural rate-limiting to call volume (it costs you as much if not more to call me than it costs me) which keeps the calling-pipes open.

Frankly, we at Fin take advantage of this.  We make phone calls and book things on behalf of people that wouldn't choose to pick up the phone to do the same thing themself because they value their time too much.

But, Fin still costs money - the cost of the human attention to make the call - We agree from first hand experience that if Duplex makes calls at scale without identifying itself it will likely destroy phone calls as a valuable protocol for communication.

But, at the same time, if Duplex does identify itself as a computer up-front in conversations then people will likely just hang-up on it / not interact with it since it can be assumed that the person on whose behalf Duplex is calling isn't paying anything for the time & is therefore likely statistically to be a lower-value caller.

The Future of Work

We have enormous respect for Google as a team, and we thought the Duplex demo was really really cool.

We look forward to having better and better ML support from Google and other technology platforms in delivering a killer assistance service using the best of technology + human intelligence.

But, based on what we know, we are not at all worried that Google is close to rendering the hybrid human + machine approach to assistance invalid... If anything, we are betting technological advancements made by Google et al. will only make our approach stronger.