The capabilities that Artificial Intelligence (AI) bring to businesses are invaluable. There is no room for doubt. Nevertheless decision makers’ skepticism over the value of AI is not unusual when it comes to applying this revolutionary technology —according to some it is also “too good to be true”— at their own company. It goes without saying (even for those who are skeptical of AI’s skills out of the lab) that AI has capacity to add value to businesses. The question is how to quantify it. Apparently thorough business case feasibility studies that provide basis for rigorous decision makers during their judgement process require a neat mathematical calculation.
In this three-part blog series, I bring a few concepts up for discussion that would help establishing a framework for translating AI’s value. These are: - Expected Value of Perfect Information - Bayesian Thinking - Value at Risk
Mentioned concepts are well-recognized in the field of Decision Science. My goal with bringing these seasoned engineering approaches, which impacted various domains from finance to healthcare, forward is to widen horizons. Perhaps that would provide a perspective for future directions of the valuation exercises for AI projects and eventually enrich the content for ROI evaluations.
In the first part of this serie, you will get familiar with the concept of Expected Value of Perfect Information (EVPI) in general terms and review a case in which EVPI is adopted to exhibit how much tangible value AI can add to a manufacturing company through a common application of AI in the manufacturing space — Predictive Maintenance.
McKinsey & Company defines Predictive Maintenance as the technology that enables smart machines to warn their operators before they break down: “Advanced Predictive Maintenance, enabled by extensive sensor integration and machine-learning techniques, is one of the most widely-heralded benefits of the fourth industrial revolution. The idea is certainly a compelling one,” they write in an 2018 article. Yet its value proposition is hard to quantify. Business benefits such as upto 25% maintenance cost reduction and upto 50% downtime reduction (Sourced from 2017 Capgemini report) are asserted in industry reports, however we should keep in mind that these numbers are generic and only indicative.
Decision makers who assesses an investment opportunity in Predictive Maintenance would like to attain solid figures that consider their own use case’s facts. But it’s quite a challenge to formalize the gains of a prediction technology. EVPI might help here. As the name suggests, EVPI is the price that one would be willing to pay in order to gain access to perfect information (Wikipedia definition).
Here is an example that would help grasping the concept of EVPI (Sourced from “Optimization Methods in Management Science” course, MIT - Spring 2013).
John is at a blackjack table. He can place a bet of $10 or do nothing. His odds of winning a bet are 49.5%. Obviously, not betting is the optimal action that maximizes the expected value of John’s return (that is $0).
Now suppose that prior to placing a bet, John will be told (with 100% accuracy) whether he will win or lose. How much is this information worth?
With a bit of quick math, we can see that (considering the fact that his odds of winning is still 49.5%) the expected value of this hypothetical scenario (retrieval of the information regarding the outcome the gamble is possible) is $4.95. In that case, John would not be willing to pay more than $4.95 for this information, meaning that EVPI equals $4.95.
In the context of our case, Predictive Maintenance, EVPI would refer to the price of the perfect information revealing machine failure times. With this information in hand, one could optimize the manufacturer’s maintenance schedule so that the total cost of maintenance be minimized, if not equal to zero (i.e., only restore machines that are going to break down, rather than periodic maintenance of all). On the other hand, unplanned downtimes as well as the cost of machine failure (lost production etc.) would be zero. These gainings in total are the EVPI of a typical Predictive Maintenance scenario.
Of course, without an oracle this scenario is completely delusive. Still it helps setting up an upper bound for the value of Predictive Maintenance. Such an exercise would provide the decision maker with the exact range of the value of Predictive Maintenance (zero being the lower bound). Eventually, the (almost) exact value can be “predicted” using the predictive power of the developed Machine Learning model (i.e., accuracy rate).
As AI advances the value of Predictive Maintenance keeps converging to its theoretical maximum that is the EVPI. Better predictions would close the gap between the attained value and the upper bound of the value range. In fact, prediction, perhaps, is the strongest muscle of AI (as manifested by notable sources; see the following quote) and therefore it is very much likely to minimize the value gap in near future.
The authors of “Prediction Machines: The Simple Economics of Artificial Intelligence” by Harvard Business Review Press highlight that “the new wave of artificial intelligence does not actually bring us intelligence but instead a critical component of intelligence—prediction.”