"AI is so important relative to many other exciting technologies. AI can be recast as causing a drop in the cost of a first-order input into many activities in business and our lives—prediction."
|TL;DR: The meaning of AI for business is not really whatever popular culture thinks it is. Instead, the meaning of AI for business today is a massive "drop in the cost of prediction". Amazing opportunities follow from this phenomenon.|
By John Morris | January 22nd, 2019
he hype curve is supposed to have a peak, and then things calm down. But pity poor business executives that have endured sustained pressure to lever the promise of AI -- for a decade at least.
Sure, AI technologies are already delivering real benefits here and there, but one gets a sense of exhaustion. From a shortage of data scientists to AI over-promise to ever-new AI innovations, the business leader asks "where's the mainstream pay-off"?
Paradoxically, the best strategy for surfing the wave of AI is not to focus so much on AI technology, but rather to focus on one's own business savvy!
This isn't just a feel-good mantra (although it does feel good), but the implication of work by pioneering economics and business researchers. In a brilliant insight, economists Agrawal, Gans and Goldfarb of the University of Toronto's Rotman School of Management have distilled the meaning of AI into something almost shockingly simple: AI lowers the cost of predictions.
The Business Meaning Of AI: "Lowered Cost Of Prediction"
ccording to Agrawal, Gans and Goldfarb, the simple economic impact of AI is a "lowered cost of prediction" in selected business and technical domains. This single impact has enormous potential. Let's explore the meaning of a lower cost of prediction and the impact on business of such a change.
The purpose of technology is to help humans get more work done with less effort. Technology helps us be more productive. Whatever our work process, the right technology artefacts will help us reduce our cost to produce the same output.
For example, what was the simple economic impact of the advent of calculators and spreadsheets? These tools, derived from hardware and software innovation, dramatically lowered the cost of the work of arithmetic2. The follow-along effects of a lowered cost of arithmetic were equally dramatic, starting with bookkeeping and accounting and cascading forward.
Before exploring the implications for business leaders of a lowered cost of prediction, let's get rid of a lot of noise. Because AI itself is a vast field of tools and techniques. And AI can be imagined to apply to many things beyond mere prediction. Sometimes all this "hype" about AI tools and use cases can overwhelm our ability to focus on the prize.
The Deep Learning / Prediction Value Chain
f we are to focus on what's most important in AI today, we need to start with the idea that today's AI is almost a "one-trick pony"1. The vast majority of deployed AI today is basically "deep learning", which is the most advanced form of machine learning. AI-as-deep-learning is very good at finding patterns in data, under controlled circumstances. AI doesn't really do much else. Science fiction and IBM's Watson notwithstanding, there's no "general AI" that can cook eggs in the morning, compose a poem in the afternoon, or organize your supply chain overnight.
Fortunately, deep learning and pattern discovery is enough to make an enormous impact! The reason that pattern discovery is so important is that predictions are all about patterns. Use deep learning to discover a pattern in a set of data and you can make a probability statement -- in other words, a prediction. Then when the same pattern is discovered in an entirely different set of data, new probability statements or predictions can be made.
Think of different kinds of predictions: A prediction that a machine will fail. A prediction that such and such symptoms are indications of a medical condition. A prediction that the supply chain should order more widgets. Or that we are about to have a surge in traffic. All these predictions are possible because a trained deep learning AI discovered patterns in new data sets.
AI Use Case No. 1: Deep Learning-based Prediction
n the next article in this three-part series, we'll explore how prediction is married to judgment and decision-making. That's where the full value of AI in support of faster, better decision-making, is realized.
In the meantime, we can finish up our introduction to the question of AI and prediction cost with five questions. These five questions help define a lower-cost-of-prediction-based AI use case, and we are well on the way to a business case too.
1. What exactly is "prediction"?
A prediction is a probability statement about the future. Specifically a prediction is a statement about a specific object or system of interest to us. And the statement concerns a the probability that this object or system will have a certain state. Predictions apply to existing objects but could also apply to the creation or even disappearance of an object ("the object is predicted to fail"). A prediction is usually thought of as implying a changed state; equally a prediction could be made that a given state will not change.
The usual tools of management decision making apply to predictions and probability. Probability is another way of stating uncertainty. Our prediction of a state is a prediction of an outcome at the end our prediction period. Risk is the probability envelope in any direction from the expected state. The prediction value chain is that AI or analytics, including machine learning or deep learning, applied to big data, helps us detect patterns. And then this analysis can be applied to new data sets. The application of a given analysis to a new data set will give us a prediction for some future state, with associated probability.
2. How does deep learning-based AI actually cut costs?
How exactly does deep learning-based AI actually enable a significantly lower cost of predictions?
The lower cost of prediction is made possible by three enabling phenomenon, (1) cheap data, (2) cheap processing power, (3) a new generation of great data wrangling tools and by one breakthrough technology, (4) AI deep learning software.
All four are essential because together they cover every step of the prediction value chain, from data set acquisition, to data wrangling to affordable pattern discovery and to prediction artefact delivery.
We start with No. 1, "cheap data", which is enabled by inexpensive sensors and low-cost storage. And we add No. 2, "cheap processing power", which certainly enables number crunching on huge data sets.
No. 3, "data wrangling tools", refers to powerful software that is now available to manage our data sets. This is a non-trivial task -- by definition "big data" means "too big for some aspect of our systems". Don't forget, typically between 50% and 70% of AI program costs are accounted for by data acquisition and data wrangling costs. AI, ML and DL get all the glory; but being able to wrangle data sets is a critical gating factor for any AI program.
Lastly, No. 4, "deep learning software technology", is the breakthrough technology without which there'd be no AI story now. And AI itself is not standing still. The big difference today is that AI software, incorporating powerful analytical, statistical and deep learning capabilities, is now available via slick less-technical interfaces. One doesn't necessarily have to have a PhD anymore to run a deep learning or machine learning analysis.
What's the result of cheap data, cheap CPU and great software? What was previously impossible is now possible. What was previously unaffordable is now affordable. Every step of the prediction generation life-cycle from creation to analysis to artefact delivery has been dramatically reduced in cost. Deep learning-based AI cuts the cost of prediction because the cost of getting, managing and using the data we need for prediction is vastly lower.
3. How important is a lower cost of prediction?
If AI-based predictions are much less expensive, why should we care? Just because we "can" doesn't mean we "should". The wonderful thing about less expensive predictions is that business is brimming over with use cases and business cases for prediction. In fact, this is a basic premise of the Data Decisioning website. Business is all about decisioning. And this is where your business savvy and business experience come in. Because opportunities for applying AI-based prediction technology won't find themselves. And AI certainly won't find it for you. It's your opportunity!
4. How can deep learning-based AI predictions be used by business?
How does one actually use AI in business? We know AI can help us with predictions. And we know that there's a data life-cycle, whereby we acquire data, wrangle and store it, and then use our deep learning AI tools to find the patterns we seek. The "AI How To" topic is well covered by some great AI- and analytics focused websites.
Our focus on Data Decisioning is different. We call out how an AI business program is a manufacturing process!
In other words, your goal for your AI business program will be to figure out how to manufacture prediction artefacts. Find your business prediction opportunities, work with business domain experts, engage your AI specialists, wrangle your data -- and deliver prediction artefacts -- or prediction tools that can be embedded throughout your organization -- again and again and again.
And what is a prediction artefact? A prediction artefact is an executable prediction model, an model instantiated in software. The model takes data inputs and then predicts various system states, to which the organization will schedule the right responses. In terms of frequency, we build AI models "periodically" -- but we likely use prediction artefacts "every day"! We started with an AI deep learning opportunity in our work process, and then we built a predictive tool for that point -- where the predictive tool is a manufactured artefact of our AI program. As part of building this artefact, we trained our AI on data. And then once trained and fine-tuned, we are ready to deploy our new capability!
Thinking about manufacturing the artefacts of prediction is the key to making AI program wins repeatable.
5. Where does deep learning-based AI prediction fit in business?
Like any technology, a successful application of that technology is always specific to given work processes of any organization. These work processes can be formal or informal, long-standing or brand new, high volume or low volume. Put a spotlight on your processes and see where prediction plays a role -- or could play a role. Or create a brand new process from nothing. All because these new capabilities are enabled by the economics of low cost prediction. We'll explore more about how exactly AI prediction technology fits in a business setting in the next article in this series.
You've probably already done this, and both business and technical media are all working hard to discover AI use cases. But what are AI-amenable prediction use cases in your work? Does it make a difference if you think about these use cases as concerning a new lower cost of prediction?
This article is No. 1 in a Data Decisioning Three Part Series: The AIs Are Coming! The AIs Are Coming! What To Do?
Next, see No. 2 as to why the value of human judgment -- your judgment -- is likely to increase as a consequence of a drop in the cost of prediction!
Resources On AI Economics
|1.||Is AI Riding a One-Trick Pony?||James Somers||MIT Technology Review||Sep, 2017|
|2.||Prediction Machines: The Simple Economics Of Artificial Intelligence||Ajay Agrawal,
Joshua Gans, Avi Goldfarb
|Harvard Business Review Press||Apr, 2018|
|3.||Prediction, Judgment and Complexity: A Theory of Decision Making and Artificial Intelligence||National Bureau Of Economic Research||Jan, 2018|
|5.||What to Expect From Artificial Intelligence||MIT Sloan Management Review||Sep, 2017|
If you found this article useful or enjoyable please share via your Twitter, LinkedIn, and Facebook social media platforms.