AUCNET built a custom image model on Google Cloud Platform using TensorFlow, there on the left-hand side of this image. But increasingly, you don't have to do that. There are a variety of domains where Google exposes machine learning services trained with our own data. For example, if you want to transcribe speech, you could use a speech API instead of having to collect audio data, train it, and predict with it. There are many such pre-trained models. And such pre-trained models are excellent ways to replace user input by machine learning. Here's an example of using a pre-trained model. Ocado is the world's largest online only grocery. It's based in the UK. And the way it used to work, was a customer sent an email. And traditionally, each email would get read and then routed to the appropriate E-department, where it would get read again. So that doesn't scale. So, Ocado turned to natural language processing. And they were able to extract the sentiment of the email and entities, the things being talked about in the email, and even the syntax of the email. And this computational technology helps Ocado parse through the body of emails and tag and route these emails to help the contact center reps determine the priority and context in a very efficient way. But increasingly, customers do not want to go to your website and click on a button. They do not want to send you an email. They want to talk to you, interactively. To get their questions and concerns answered. Manually, answering each call doesn't scale, and so Gartner estimates that in a few years, we will be spending more on conversational interfaces that even on mobile apps. So does this mean using the speech API, transcribing the speech, and then trying to make sense of it? No. What I'm showing you here is a high-level conversational agent tool called Dialogflow. So look here at this screen and we notice that the agent says, "How may I help you today?" And the customer says, "I want to get a pizza with bacon and cheese." And based on that, Dialogflow goes ahead and builds a JSON message that says that the customer wants to order a pizza, and is able to go ahead and add the toppings corresponding to bacon and cheese. And then the agent says, "Well, what size do you want?" Because that's another requirement and based on that, the size as large also gets added and the standard crust and then says, "Oh add olives." And now, into the toppings, olives gets added. Notice that is a very conversational interface and from these conversations, a JSON structured message gets built and it's this JSON message, this very structured message that goes to the rest of the application which works the same way as before, except that the user input has not come from the customer pointing and clicking their way through a webform but instead has come through a conversational interface.
