Last week a chief marketing officer (“CMO”) mentioned to me that he admired the CEO at a Toronto artificial intelligence company, but could not understand, what, exactly, his company did. And therein lies the black box paradox of artificial intelligence (“AI”) – everyone’s talking about it, but nobody knows exactly what it is. As the authors of the excellent book, Prediction Machines, say:
Like other game-changing, general purpose technologies (“GPTs”) like the steam engine, printing press, and internet, many of the applications for artificial intelligence are unknown. This creates a paralyzing uncertainty, which tends to “freeze” adoption by many companies. It also creates a raft of consultants offering to “help” you figure it out for hundreds of thousands of dollars. I remember similarly hiring an email marketing firm in the early 2000s at Dell, to help us send out mass emails, increase conversion rates, and build our email list/database. Today, anyone can do that reasonably well for $20 a month with a Mailchimp account. I predict that machine learning (the essence of AI) will be a core part of the 10th grade curriculum 15 years from now. Not kidding.
So what does AI do? Simply put, it makes predictions. Prediction takes data you have and uses it to generate insights you don’t have. In marketing, my field, the key predictions seem obvious. They will increase user traffic, conversion, or revenue per unit, or they will decrease customer churn (customers who leave or stop buying your product). Here are some examples:
- Flyer optimization: What combination of products should I place in my paper or digital flyer, and where, in order to increase foot traffic to my store or digital traffic to my site?
- Increasing conversion: What combination of price and product will increase my conversion rate (the percentage of people who buy a product after seeing some sort of “advertisement” for it)? Or, alternatively, which customers should I phone/email with a particular offer, who would be most likely to take that offer?
- Decreasing churn: What combination of behaviours indicate that a customer is likely to churn (stop being a customer) in the next 6 weeks? What offer could we make to them, that will likely stop them from churning?
A smart marketer could probably tell you her best 3-4 promotions over time, based on an Excel spreadsheet analysis from some financial database. So that would take care of the premium positions on the front and back of the 8-page flyer. But what should she put in the middle 4 pages? A marketer could also probably tell you that a group of prospects which visited the product overview page for a laptop computer 4 times in the past two days, and put the laptop in an online shopping cart but then abandoned it when shipping was added to the price, was most likely to respond to an offer for free shipping. But which other prospects might respond? One of the key differences between machine intelligence and human intelligence is that human logic works well at either end of the bell curve – in this case, customers with 2-3 characteristics that indicate they are either highly likely to buy or highly likely to churn. Machine learning deals better with the not so obvious – the “mushy middle” of the bell curve.
What key characteristics of machine learning make this true?
- Machines are much better than humans at factoring in complex interactions among different factors. This is simply because humans can think in, at most, 2-3 dimensions at a time, whereas a machine can handle many more. Machines can scale in a way that humans cannot. As a CMO, I know my best 2-3 promotions. I don’t know my best 48 promotions.
- Machine learning enables predictions based on unanticipated factors. A regression analysis requires human model-makers to provide the independent variables – the factors that will influence the outcome, or dependent variable. Machine learning can sift through data and discover correlations that humans might miss.
- Machines can adjust predictions on the fly – at the level of the individual user and the device itself. When someone swipes right or left on the dating app Tinder, that feeds immediately into the predictions to determine which potential date to show next.
- Humans think deterministically (“What are the features of a cat? Does this animal have those features?”). Machine learning uses past examples, correlations and probability to solve problems. “Does this animal have the same features as the cats I have seen before?”).
Of course, smart humans make predictions, too. Machines can simply do it faster, more cheaply, and using far more variables than humans can. As predictions become cheap, the value of judgment (the skill used to determine the payoff or profit from a given course of action) goes up. In addition, decisions still have to be made and actions taken. Decision-making requires applying judgment to a prediction and then acting. Human judgment and the ability to decide, to act, in conditions of probability and uncertainty, is still greatly valued.
In the coming brave new world, smart machines will not replace humans, as many fear. However, there will likely be a new division of labour, with machines providing recommendations and humans making the final decision. And there will be problems we cannot anticipate and advances we cannot yet imagine that will change the way we live and learn forever (just try explaining to your university/college student how you studied in “the time before the internet” if you don’t believe me). We’ve been blessed to see the world-changing impact of the internet, and we are fortunate to be in the early days of another general purpose technology, artificial intelligence. And we may yet live to see a third.