Artificial Intelligence can have a powerful impact on business, but two major challenges make implementation critical and difficult: (1) getting the target right, and (2) getting the process right.
The problem is that these challenges are rooted in pervasive, thorny management issues, and many AI initiatives focus largely on its technical viability. Managers who fail to address these critical issues from the start are in danger of simply optimizing bad processes (i.e. optimizing in the sense of improving the likelihood of achieving the process goals) and locking them in place.
The deeper issue is not whether AI works, but instead, whether managers are so enamored with AI/ML (artificial intelligence-machine learning) as a new management “wonder drug” that they fail to work through the deeper management issues needed to configure it correctly and determine where it is appropriate.
This is a pressing issue: a recent authoritative study reported that only a small minority of executives were satisfied with the returns from their AI investments.
AI: New Wonder Drug?
AI has all the hallmarks of a new wonder drug for business: It is very powerful at optimizing through pattern recognition processes characterized by massive, complex information.
To get a perspective on why AI is sweeping through business today, as well as a brief understanding of how it works and its strengths and weaknesses, here are a few observations:
• Computation keeps getting half as expensive roughly every couple of years, and this trend has persisted for over a century, cutting the cost of computing by a whopping million million million (1018) since 1900.
• Neural networks, a family of computational models loosely based on how brain neurons send signals to each other, have recently started dominating the AI subfield known as Machine Learning—the study of algorithms that improve through experience.
• In overview, MIT’s Max Tegmark notes, “Just as we don’t fully understand how our children learn, we still don’t understand how such neural networks learn, and why they occasionally fail.” Many authoritative studies confirm this view.
In essence, AI/ML is a generalized process to optimize things based purely on pattern recognition without the need for understanding either the broader context or understanding why the optimizations are effective. The essential question for managers is whether they are willing to bet their company’s prospects on a mechanistic pattern recognition process.
Getting the Targets Right
Establishing targets and goals for an AI/ML system is one of the most important challenges to successful implementation. First and foremost, the question of what is being optimized is neither clear nor uncontroversial.
What is being optimized
The most important choice is whether the system is set up to maximize net profits or to optimize a narrower target, such as maximizing advertisement conversion rates or minimizing delivery costs. This is a critical issue because mitigating these narrow problems in a way that is divorced from the big picture can decrease rather than increase net profits.
The reason for this seeming paradox is that some revenues are very costly to serve, while other revenues are ample to support higher costs. In our experience accelerating the profitability of tens of billions of dollars of client revenues, we have found that gross margin does not predict net profits.
For example, in one well-run company, 15% of the customers produce 135% of the profits, while 20% of the customers drain 55% of the profits; and 25% of the products produce 150% of the profits, while 25% of the products drain 75% of the profits. In this company, maximizing the advertising conversion rate for money-losing Profit Drain customers and minimizing the delivery costs for service-sensitive Profit Peak customers are both counterproductive. This is true in most companies.
The reason why so many managers incorrectly focus on gross margin, which leads them to the intuitively “obvious” but empirically incorrect assumption that revenue-increasing and cost-decreasing objectives are always good to optimize is that in the prior Age of Mass Markets (in most of the prior century), pricing was relatively uniform and cost to serve was uniform as well. Today, pricing varies by customer, and even within customers, while the cost to serve varies enormously, ranging from highly integrated vendor-managed inventory systems to simply dropping boxes at the receiving dock.
This means that companies need an Enterprise Profit Management (EPM) system, which is a SaaS system that can be configured in a few weeks and creates an all-in P&L on each transaction—each product sold to each customer each time. Because a well-targeted AI/ML system will maximize the net profitability of each transaction, it must “learn” from historical data at this level of granularity. In our experience, EPM enables managers to produce 10-30% year-on-year sustainable profit increases.
The bottom line is that installing an AI/ML system, for example, to maximize revenues for the sales group, while installing a separate AI/ML system to minimize costs for the supply chain group—as so many companies today are doing—will not maximize net profits, no matter how elegant the AI/ML algorithms might be.
Whose profits are optimized
A related critical question is: Whose net profits are to be maximized? The goals of a local regional sales manager, an SBU head, the CEO, shareholders, investors and others differ in fundamental ways. This is a very common problem in business.
In my graduate course at MIT, I teach a case in which a bleach purchasing manager has to respond to a supplier offer for a price cut. An alternative supplier can reduce the customer’s supply chain costs, but while this offering reduces total cost, it carries a higher price. What will the bleach buyer do? The answer depends on his or her control system: according to standard practice, the bleach buyer will be trained, evaluated, promoted and compensated on purchase price, while the supply chain cost reduction benefits another department.
It is clear what the bleach buyer will do—and this ironically is what the company’s control system is instructing him or her to do. And, just as a human purchaser incentivized to minimize cost will do so at the expense of net profit, so too will an AI/ML purchasing system whose objective function is to minimize cost. In fact, the latter could be even worse, because there is less chance of anyone noticing the problem.
If this sounds like a familiar management problem, it is. The management problem is the same, whether or not AI/ML is used. The difference is that AI/ML systems are so technically dazzling, that managers tend not to focus on the core management issues.
Complexity and sub-optimization
Similarly, AI/ML systems have to be configured to address the complexity and sub-optimization that is endemic in business. This is extremely hard to set and maintain.
For example, companies have very different strategic goals at different stages of market and company development. At times, a company’s objective may be to establish a market presence (even investing in reduced prices), later developing widespread distribution (even offering marketing support to new distributors and giving free samples to opinion/style leaders) and sometimes tying up essential channels (through package deals or special returns allowances).
Strategic paradigm shifts
Capturing strategic paradigm shifts is extremely important to a company’s long-term success. These paradigm shifts are exogenous to the historical and current experience. Think about forecasting the demand for cameras on phones for a traditional camera company. Steve Jobs famously never used market research because it reflected opinions on current products, but not opportunities to discover breakthrough paradigm-changing products. AI/ML cannot do this because the paradigm change is by definition outside the range of experiences that the system has “learned.”
Managing the Process of AI
Just as managers must take care in specifying the targets and goals of AI/ML, they must be very thoughtful and systematic about structuring the business processes in which it is embedded.
User trust and process fit
First and foremost, managers have to gain user trust in the AI/ML output recommendations, and it has to fit into current business processes. This is a surprisingly difficult set of needs.
This difficulty is rooted in the fact that most business processes are not deterministic. For example, if an AI/ML system “learns” to predict a customer’s next purchase (based on that customer’s past history in the context of the history of all customers’ purchases), the likely next purchase is really a probability distribution because human behavior is not fixed. If the system issues a “recommendation,” it will have an implicit likelihood of “errors” because the recommendation is the most likely single point, while the actual probability distribution is a range of points.
Although this is a natural artifact of the nature of the underlying business processes, it will generally seem like an “error” to unsophisticated users and will appear to discredit the “accuracy” of the system’s output. Because AI/ML is simply a mechanistic pattern recognition process, users will not have a way to develop an intuitive sense of the “rightness” of the output.
Contrast this with a very powerful system that compares each transaction, customer or product with the company’s own observed best practice (controlling for a variety of salient factors like customer industry, size and so on). Here, the user can actually look at the best practice instances to see what the differences are in order to judge the best next step and can even call the colleagues involved to talk over the situation. In short, in this situation, the user will experience a feeling of understanding and control, which is essential to acceptance.
In our experience using advanced analytics, it is necessary to root them in an intuitively clear comparison of the customer or product to the company’s own best practice. We have found that without this level of intuitive understanding, users are reluctant to use system output information.
In AI/ML systems, this problem is exacerbated by a lack of understanding of why they occasionally fail. In studies of system acceptance, users place enormous weight on occasional errors, neglecting average performance. (Think about your likelihood of returning to a restaurant in which you had nine good meals, then a tenth meal that really made you sick.)
A related issue arises in structuring the processes in which AI/ML is embedded. These systems can produce so much new information that most companies’ traditional business processes cannot accommodate it.
For example, although an AI/ML system may be able to predict when a small customer is likely to leave, most companies do not have systems and processes to act on this information—and developing such a system might be a low priority. Managers must have a way to estimate the benefits of each candidate AI/ML application and weigh its value against the change management needed to realize the practical benefits. By contrast, today many companies are implementing AI/ML systems in a one-off way based on local priorities and interests.
Intuitive understanding and paradigm shifts
A deeper problem with relying on AI/ML is that the managers involved tend to lose the stream of primary customer, product and market information that they need to develop an intuitive understanding of the subtle shifts that lead to the early discovery of new opportunities and risks, especially paradigm-changing ones. As managers simply apply AI/ML output to their business, there is a danger that they will distance themselves from their first-hand experiences of their customers, products and markets. AI/ML can only develop patterns based on the range of past experience, but humans have the unique capability to infer creative new opportunities to shape the future in ways that fundamentally move beyond past experience.
For example, Nalco sells chemicals for water treatment systems. The company had high delivery costs because customer orders were difficult to predict. In response, a manager developed a program to install wireless meters on the customers’ tanks to read inventory levels. This allowed the company to optimize its delivery routes. However, a creative manager saw that this information also enabled Nalco to compare the rate of actual chemical use to the optimum for the customer’s system—leading Nalco to develop a new initiative to call customers when they saw that the actual use indicated that the system was operating sub-optimally, and this created customer benefits that dwarfed the cost of the chemicals.
If Nalco had simply installed an AI/ML system to minimize its delivery costs, it probably would have completely missed the opportunity to become a strategic partner to its customers, a move that transformed it from a commodity supplier into a high-value-added partner.
Accommodating customer diversity
In most companies, the portfolio of customers is diverse, ranging from Profit Peaks customers (high-revenues, high-profits) to Profit Drain customers (high-revenues, money-losing) to Profit Desert customers (low-revenues, low-profits). It is very important to manage the Profit Peak customers very differently from the Profit Drain customers, and to manage the Profit Desert customers very differently from the two large-customer segments.
Moreover, using AI/ML extensively in Profit Peak customers would be problematic in many cases because these key customers tend to be price-insensitive but very service-sensitive, and they respond very well to paradigm-changing service innovations.
Profit Peak customers often want innovative extended products, like vendor-managed inventory and category management, which can transform the relationship into a strategic partnership. In these customers, sales reps and others often need to sub-optimize their pricing and other areas for a time in order to gain political advantage in a large customer’s complex buying center. Similarly, many times the decision at hand is unprofitable but critical to a larger series of following high-profit decisions.
These are crucial issues in a company’s most important high-profit customers. Simply punting product mix, pricing and related decisions to AI/ML offers the false prospect of data-based optimization, when the real issues are political and organizational.
Organizing for Successful AI
In our experience, the most productive way to manage the implementation of powerful AI/ML systems is to structure a two-level process.
• At the top level, upper management, working through a Managing Profitable Growth Committee comprised of upper-mid level managers (e.g. directors and vice presidents), determines and monitors the targets of the AI/ML systems. It plans and oversees the company’s AI/ML program, and explicitly scans for paradigm-changing strategic opportunities and risks that AI/ML cannot address. It focuses on getting the processes in which AI/ML is embedded right, along with discovering game-changing situations that must modify or supersede the AI process.
In particular, it is essential and necessary to devise a separate AI/ML strategy for each customer profit segment (Profit Peaks, Profit Drains and Profit Deserts). This enables the MPG Committee managers to custom-tailor the AI/ML-based processes to the widely differing priorities, nature and needs of each profit-based set of customers and products.
• At the grassroots management level, managers operate the profit-enhancement processes in which the AI/ML is embedded. At this level, the key to effectiveness is to present the AI/ML recommendations in parallel with non-AI/ML best practice-based, intuitively understandable recommendations. This makes all the difference in user acceptance.
By getting your business right at these two cornerstone levels, you can deploy AI/ML in a systematic way that realizes its full potential. As always, business success comes down to thoughtful management—no matter how dazzling the technology involved.