At Intersection X we have the privilege of talking to many companies, working on great products, powered by awesome new technologies. Unsurprisingly the technology conversations tend to concentrate around the hot, disruptive technologies making their way up the hype cycle.
Our job, however, is to decode the technology hype and mine for real value that can be productized from these technologies. For 2013 and 2014, the hype and conversations were dominated by Drones and UAVs. In 2015 and 2016 the winner has easily been Artificial Intelligence (AI). For our inaugural research-focused post we're going to share our insights into differentiation & defensibility of AI products. We'll leverage the patterns we’re observing across the opportunity of “AI”. As a teaser we’ll touch on everything from machine learning to bots!
Before we launch head long into this, we should warn you that it’s a long post. We’ve added a handy little table of contents below so you can read it in chunks.
To kick things off, we're going to frame our conversation to establish a common understanding. In our workshops and in our client engagements we use something we call a Product Value Model (PVM). It's a conceptual model that illustrates the components that both deliver the value to the customer, and line up with technology components. It’s a really powerful tool for linking business value and the technology's contribution to that value.
Imagine for a second that you were the Product Manager for Amazon Prime before it existed, and you wanted to explain to the execs what were the important parts of the product you wanted to build and at the same time explain the business value. We use a product value model for exactly that. Here’s a PVM for the new Amazon Prime product;
The model highlights the things Amazon prime needs to do well to have a successful product; acquiring users effectively, enticing them to quickly and easily select and purchase something, and finally delivering purchases flawlessly, consistently and rapidly to the customer.
We can use the Product Value Model to think about the AI opportunity. We’ll call it an Opportunity Value Model (OVM) instead because we’re not talking about a specific product. The OVM represents the pieces of value that combined represent a generalized AI product. We'll use that model to explore what constitutes those elements of value and, in our second post, as a lens to differentiate AI opportunities.
So, with no further-to-do, please welcome our Artificial Intelligence Opportunity Value Model;
We build products to deliver value, so before we break the model down, let’s discuss the types of value AI products generate for users today.
As humans, we generally experience two types of value from products; utility value and emotional value. For the most part, the value derived from the AI products we're seeing and working with is utility. Some simple examples of utility-enabling AI products are automated executive assistants, bots to help you book travel, or cars that drive you (instead of the other way around!). These cases are simple examples of using an AI product to deliver utility-value by saving time.
With that context, let’s break down the OVM starting with the interface;
Interface To derive value from a product, there has to be some type of interface to interact with the product and receive the value. A key driver for the growth of AI is its potential to make technology behave in a more human-like manner and in doing so make it simpler.
A nice example of simplified UI in action is the current interface trend in AI chat-bots. Most chat-bots are not being delivered as new stand-alone apps, rather existing UIs (Facebook Messenger, Slack, etc) are being used as the human interface to the bot.
Message Bots are simply extracting insights from data and models and delivering that information via lightweight interfaces. There’s no longer a need for a native app or web app to be custom created, rather the bot leverages an existing interface, e.g. SMS or Facebook Messenger.
Note If AI is enabling products to no longer need an interface, why is interface part of our AI Opportunity Value Model? Interface is relevant because there is a better interface than your thumbs. And AI is enabling it. We humans long-ago developed a cool interface called speech. Teaching a computer to interface with humans via speech (aka Automatic Speech Recognition (ASR)) is advancing very rapidly due to large investments made by Amazon, Google and Apple. Imagine what our world will look like when we just talk to our technology and it finally understands the intent as well as another person might. Plucky startups like Capio, a client of ours, are making high-performing ASR accessible by driving cost for high accuracy solutions way down. As a result high accuracy ASR is coming to the mass market now, and interfaces are changing!
Model There are many reasons the time is now for AI but perhaps the biggest reason is the advancement of the algorithms necessary to simulate human intelligence. Machine-learning, computer vision and natural language processing are all terms that are familiar to many of us and they are used to reference algorithms that process data. In simple terms, algorithms are used to process raw data into an insight. That insight is added to the other insights and a model of the world that we’re interested in is created. You can think of the model as an answer machine for a particular topic. The model represents the simulation of human intelligence and the better your model is, the better the answers you get from the machine. (P.S. If you have 20 minutes, this video is a great illustration suitable for the lay-person of a modeling technique called neural networks.)
Datasets, aka knowledge. And now comes the most important part. In fact, if it’s missing there is no AI product. It is the dataset. The value that AI products deliver is through mining some form of knowledge. That knowledge might be a dataset representing a digital map or it might be hundreds of thousands of images of skin abnormalities or purchasing responses to pricing changes. The value of our model (the answer machine) is primarily a function of that dataset’s quality. The quality of dataset can be measured in different ways, but for the sake of simplicity we can assume the larger the dataset the larger the quality. As an aside, this is the allure of big-data. Large data-sets hold the keys to many things and we’re now in an age where it is easier to mine those (big) datasets.
Infrastructure. I know I said the dataset was the most important, but coming in a close second is infrastructure. AI’s current rise is powered by the availability of the computational power needed to power the CPU-cycle hungry algorithms. Algorithms and models get complex very quickly and often they are dealing with combinatorial explosion that requires horsepower to process. Thankfully the hardware industry has been hard at work to bring ever more powerful, application specific hardware to market for us, and at a reasonable price. Anyone with an AWS account now has access to the computational power necessary to power sophisticated models.
Alright, lets summarize; Most AI-products are delivering utility value, rather than emotional value, via ever simpler interfaces. The back-end of these products is leaning on two key components; data (a source of knowledge) and a model (an answer machine). These new products need very powerful infrastructure to do their work. If AI-products were a form of transport they would be a rocket ship because they consume vast amounts of rocket fuel (computing power). Thankfully people have been working to create cheaper rocket fuel making it possible to build AI products more cost effectively.
Ok, we now have a common understanding through which to analyze our AI products. Let’s try it out with a couple of examples ahead of the second post; the cheat sheet for understanding AI opportunity quality and defensibility. We’ll try the OVM out on a messaging bot and an old-school document scanning (OCR) solution.
Messaging Bot; Product Hunt recently launched Kittybot, an AI messaging bot to uncover new products. Here’s how it work; if I want to find out about the latest products in IOT, Kittybot will surface them for me as soon as they are on Product Hunt. Product Hunt cleverly decided to use Slack as KittyBot’s interface.
Ok, so lets break it down;
· Value; The end-user value is that I get finely curated information directly to an interface that already has my attention.
· Interface; Product Hunt leveraged an existing, widely available interface; Slack.
· Model; (We’re assuming) the algorithm they are using is a simple one; search.
· Dataset; Product Hunt are leveraging their own proprietary dataset of quickly growing new products that the crowd has already characterized.
· Infrastructure; Again we’re assuming this a relatively simple hosted solution like AWS.
Document Scanning; OCR is not a new technology at all. In fact it has been around for a long time trying to solve something that most humans find easy; reading documents. It is a form of Artificial Intelligence though. Let’s test out the value model on a generic office document scanning solution.
· Value; The end-user value is that I get digitized data from a paper document that I can now act on, e.g. identify the address on a package to know where to route the mail.
· Interface; If all goes well, and the mail is routed correctly, there is no human interaction required. When there are errors the human will need to use a lightweight admin interface to solve the issue.
· Model; OCR is generally driven off a rules-based model that is built over time and is context sensitive, e.g. a text block in the top right corner is likely an address.
· Dataset; The OCR system learns by seeing lots, and lots, and lots of documents.
· Infrastructure; There is both hardware and software infrastructure required. The software and computing infrastructure is relatively simple and uses commoditized equipment. The hardware is very, very specific to get thousands of odd-shaped envelopers scanned quickly.
We can see that the model is really useful to parse both the technology and the business value of these wildly different applications.
We hope you enjoyed this first post. Stay tuned for post #2 where we’ll discuss how to identify the characteristics of AI products that drive differentiation and competitive advantage for the long run!
P.S. While you’re waiting for Post#2 you might enjoy our post on what AI means for product managers.
 Our working hypothesis for this is that a) utility is more targetable and b) emotional value doesn’t make “rational” sense so it’s harder to target.
 Utility can beget emotional value, e.g. I sent my friend in Australia a whatsapp message about a friend getting engaged (utility) and I re-connected with them about our time in school together (emotional value).