The Return of Artificial Intelligence: Is AI Finally Ready for Business?
Booth, Carey, Buluswar, Shashi, The McKinsey Quarterly
Artificial intelligence has come in and out of vogue more times than Madonna in the past 20 years: it has been hyped and then, having failed to live up to the hype, been discredited until being revived again. In the late 1990s, an observer at a World Wide Web technology conference reported that most of the proposals there had been floated, several years earlier, under the AI moniker and were now being recycled--good technology solutions looking for real business problems to solve.
AI's biggest enemy may be the promises made by its proponents: ambitious entrepreneurs looking for venture capital and academics who underestimate the challenge of meeting the needs of business users. The technology seems like a good way to automate everything from entrapping hackers to following money trails, but we are still a long way from Stanley Kubrick's menacing computer, HAL, or Steven Spielberg's AI.
Nonetheless, the AI-development community has generated techniques that are beginning to show promise for real business applications. Like any information system, AI becomes interesting to businesses only when it can perform necessary tasks more efficiently or more accurately than they have been performed before or can exploit untapped opportunities. What makes AI much more likely to succeed now is the fact that the underlying Web-enabled infrastructure creates unprecedented scope for collecting massive amounts of information and for using it to automate business functions.
The exhibits in this article introduce three types of AI, along with real business applications for each. In every case, the company involved has derived real economic benefit.
Working with complex data in dynamic environments
All AI systems share a few characteristics: they are designed to work with data sets that are large, complex, or both; to search through them and find relevant data; and to look for patterns. (1) AI systems are useful for dealing with any dynamic environment in which they have to make intelligent choices, depending upon the input, among a number of possible answers. They also use samples to form generalizations about an entire data set and, in some cases, make or help make intelligent decisions. Potentially, they could even execute tasks. These systems can be categorized into three types (Exhibit 1).
1. Numerical analytics systems find patterns and rules in big numerical data sets. They are most useful for problems (such as the detection of fraud) that require heavy number crunching to distinguish among different sets of items.
2. Rule-based decision systems use predetermined rules or logic, with or without numerical data, to make decisions and determine outcomes. They are useful for automating work flows.
3. Autonomous execution systems (also known as agents, or bots), which run continuously, monitor information as it arrives, typically from several distributed sites, and execute specific tasks in response to what they find. They are most useful for automating tasks across organizations by using data shared over the Internet, especially when the underlying data are structured according to prevailing standards such as the Extensible Markup Language (XML).
Numerical analytics systems
The biggest risk in the credit card business is default. Capital One Financial uses AI tools to identify potential customers with different risk profiles; the company then tailors its offerings and adjusts its interest rates to match them (Exhibit 2, on the previous page). This information-based strategy helps Capital One to process about five times as much data as do typical competitors, to create two or three times as many product offerings, and to write off only 60 to 75 percent as much bad debt as the industry average.
While this strategy is the cornerstone of Capital One's broader competitive approach, other companies have made smaller, targeted investments in AI-based tools to achieve similar efficiencies. …