Five years ago, Brenda Dietrich started to investigate how IBM’s 40,000 salespeople could learn to rely a little more on math than on their gut instincts. In particular, Dietrich, who heads the company’s 200-person worldwide team of math researchers, was asked to see if math could help managers do a better job of setting sales quotas. She assigned three mathematicians at IBM’s Thomas J. Watson Research Center in Yorktown Heights, NY, to work on new techniques for predicting how much business the company could get from a given customer.
The mathematicians collected several years’ worth of data about every sale IBM made around the world. They compared the results with the sales quotas set at the beginning of the year, most of which were developed by district sales managers who negotiated them with sales teams on the basis of past experience. To spot opportunities the sales teams didn’t recognize, the researchers collected external data on IT spending patterns by industry and combined that information with the internal sales data. Then they used a technique called high-quantile modeling–which tries to predict, say, the 90th percentile of a distribution rather than the mean–to estimate potential spending by each customer and calculate how much of that demand IBM could fulfill.
Armed with these predictions about how much equipment IBM should be able to sell to each customer, Dietrich’s mathematicians looked at the size and makeup of the sales team on each account and compared its actual performance with the theoretical maximum. Some teams were so small they couldn’t sell enough to meet that potential demand. Other teams were unnecessarily large. So the mathematicians advised the sales department to shift its staff around, taking less productive salespeople off the big teams and putting them on teams that had been too small. Sales in the latter accounts quickly grew.
The two-year project had a tremendous payoff for IBM. The corporate controller concluded that it generated $1 billion in additional sales through 2008, the year after the team finished its work, says Dietrich, a 50-year-old PhD with a sneaking suspicion that the world would work better if it were run by mathematicians. Since then, IBM has incorporated high-quantile modeling into its workforce analytics practice, a service it offers to help clients make decisions about human-resources issues such as how best to deploy their salespeople.
And the company drew a more general lesson from the experience: it came to believe that its mathematicians’ innovations were something for which other businesses would pay handsomely. Last year, the company created a major new business analytics and optimization group within IBM Global Business Services, and the group has already trained 4,000 consultants. IBM hopes to eventually do as much business in analytics as it already does in enterprise resource planning, which helps companies coördinate their information technology across separate departments; that service is a leading source of revenue in the $17.7 billion business services unit and has been one of its fastest-growing areas over the last 10 years. The two groups already complement each other: while enterprise resource planning tracks and organizes business processes, analytics maximizes their performance.
Dietrich, whose name is on 13 patents, thinks she and her team can create models that accurately describe activities far outside what is normally considered the realm of mathematics. For example, stochastic optimization algorithms, which incorporate random elements rather than assuming that all values are exact, have been used for decades to help manufacturers and financial markets adjust to changing conditions. But IBM’s mathematicians are applying the techniques to problems in human resources and marketing. They are using mathematical models to help the company find new customers and figure out the right mix of veteran and junior programmers to assign to a big software project. They are analyzing data to determine whether it is worthwhile for IBM to advertise in specific magazines or on certain television shows, or to attend particular trade shows. “We’re able to predict the impact of certain advertising programs on revenues,” Dietrich says–though, she concedes, “not with the precision I would like.”
Even if they are imprecise, Dietrich believes, these analytic techniques can be hugely helpful to many companies, which she says often don’t fully understand their internal processes and business models. Studying all the available data about sales and manufacturing could reveal bottlenecks that might be cleared or uncover opportunities that have been missed. She and her team are increasingly getting involved directly with customers. For instance, because of her reputation as a scientist and head of a math research team, she was recently invited to talk to a big pharmaceutical company’s executives about whether mathematical modeling could improve their process for allocating funds tovarious drug development efforts.
Such activities are a big departure from what IBM mathematicians used to do. In the old days, they were an odd breed among the scientists and engineers, who worked on science and technologies that might eventually lead to new semiconductor materials, new storage devices, or parallel-processing supercomputers. The mathematicians sometimes modeled IBM production processes, but they were judged mostly by their theoretical work and their publications in academic journals.
That started to change in the early 1990s, when IBM racked up huge losses. The board
pushed out top management and brought in Louis Gerstner, then the head of RJR Nabisco, to serve as CEO. Though Gerstner took steps to break up IBM’s sclerotic bureaucracy, he chose to keep the company in one piece. He said he believed that IBM’s size, which enabled it to focus resources on big problems for large corporate and government customers, was a valuable asset that should be preserved.
A key part of Gerstner’s strategy was to unify and expand IBM’s global-services business. Paul Horn, who headed IBM Research during part of that time and is now senior vice provost for research at New York University, saw that under the circumstances, the labs could easily be viewed as a costly luxury. With services growing, he says, “if research wasn’t contributing, you could imagine someone in the future saying, ‘You don’t need to be so big.’ ” Horn, a physicist, helped convince Gerstner that IBM’s research division could play an important role in his strategy by working with customers to solve their problems. He began pushing his thousands of researchers, including the mathematicians, to start working on projects that could be useful to the services business. The motive was simple, he says: “Survival.”
For the mathematicians, the shift was a natural one. Dietrich says they had frequently worked with IBM’s own manufacturing plants on scheduling problems and logistical issues, though the results were usually considered proprietary. And they had already begun getting more involved in business operations, in part because it provided them with the large data sets that they needed for modeling. Historically, stochastic optimization had been limited by the sheer amount of computing required to deal with multiple variables. But as computer power exploded and researchers began to use massively parallel processors, they were able to manipulate much more data.
IBM Research mathematician Baruch Schieber recalls going to a Brazilian steel mill and finding that production schedules were being drawn up on whiteboards. Surely, mathematical models could do it better, he thought. He was especially interested in the issues involved in scheduling production runs for different varieties of steel. Though it’s cheaper to do long runs of one type of steel, sometimes customers need several different types immediately, so the mill has to do short runs. “Mathematical modeling is quantifying things that usually aren’t quantified,” he says–such as the tradeoff between cost and customer satisfaction. Early in a contract period, Schieber discovered, the mills wanted to optimize their schedules for maximum efficiency and minimum cost. At the end of the period, when the contract was up for renewal, they sought to focus more on improving satisfaction. Similar issues arise with airlines. Schieber says, “We ask managers: do you want to minimize crew costs or fuel, or do you want to maximize customer satisfaction?”
William Pulleyblank, who headed IBM’s math department in the 1990s, had urged the company even then to make a business out of analytics. “A lot of companies tried to do this,” he says. “It was seen as a pure product play–package it and sell it.” However, he adds, it became clear that IBM didn’t have a good way to sell the mathematicians’ skills to clients. He concluded that many companies’ needs were so specialized that designing a general-purpose software package wouldn’t be profitable–but software designed for particular businesses wouldn’t be in high enough demand. At the same time, IBM didn’t want its researchers to become consultants. The mathematicians didn’t want to do it, and they weren’t trained for relating to customers. “I realized the challenge wasn’t the math,” says Pulleyblank, who is now a vice president in the business analytics and optimization group. “It was how to make it a business.”
The path to an analytics business became clearer in 2002, when IBM paid $3.9 billion to acquire the consulting business of PricewaterhouseCoopers. Ginni Rometty, who spearheaded the deal and now heads IBM’s sales operations, recalled Pulleyblank’s idea. She thought that PWC’s consultants could expand IBM’s service offerings beyond IT; its researchers could be touted as a unique source of advice to client companies on marketing, human resources, and logistics. Each fall, when IBM’s sales teams start forecasting upcoming business, the consultants identify critical problems that are likely to affect particular industries in the coming year. If those problems look like analytics issues, the consultants contact the business analytics and optimization team and ask whether IBM has worked on anything similar before. In many cases, the problems can indeed be addressed by adapting the company’s existing software products.
When existing software can’t do the job, the consultants turn to IBM Research for help. Sanjeev Nagrath, IBM’s global leader for supply-chain management, encountered such a situation last year when clients started asking how to reduce the carbon footprint of their supply chains. So, Nagrath says, they’re working with Research to come up with industry-specific models to deal with sustainability issues. And two years ago he worked with Dietrich to create a center for supply-chain
innovation in Beijing. There, Chinese mathematicians are part of a team working with companies such as Chinese shipping giant Cosco. The innovation center’s mathematicians helped IBM consultants model Cosco’s procedures and developed a plan that cut fuel costs 25 percent and carbon dioxide emissions 15 percent. Among other things, they recommended reducing the number of distribution centers from 100 to 40.
Not all clients trust the mathematicians’ contributions, as Schieber found out when he created a model that could be used to reschedule ships if supplies were temporarily halted by bad weather. He says it was much better than human schedulers at adjusting fleet movements and speeds to minimize disruption and fuel costs. But the customer wasn’t satisfied. “It was a black box,” he recalls. “The shipper said, this is our competitive edge. They wanted to understand it.” The shipping company finally implemented the model after IBM redesigned it so that it was not a fully automated system but an aid that human dispatchers could consult.
Some businesspeople argue that many decisions are best guided by gut reactions based on years of experience. They worry that depending on analytics will make business leaders indecisive when they don’t have an abundance of data. And a math-phobic public is suspicious that analytics-driven programs cut costs at consumers’ expense. IBM researchers point to the recent backlash against recommendations that annual mammograms be delayed until women are 50 because they don’t provide statistically provable benefits for younger women.
But Dietrich is more concerned that companies will fail to analyze the petabytes of data they do collect. When she met with the pharmaceutical company about its portfolio management strategy, for instance, the executives explained how they allocated spending according to their estimates of how likely each project was to succeed. “I asked them if they ever checked to see how well the estimates matched their results,” she says. “They never had.”
Dietrich and her researchers are now working to rewrite optimization algorithms to take advantage of massively parallel computers. The older programs were written to minimize the number of operations required. But now that thousands of processors can churn through vast data sets, she says, “the issue is to reduce [run] time.” Once the team is done, those optimization programs will be available to businesses whose stores of data are too large to be analyzed with single-thread computer programs.
The most interesting problems the mathematicians envision for future projects involve situations where a model must incorporate behavioral changes that the model itself has inspired. For example, Dietrich says, a traffic congestion system might use messages sent to GPS units to direct drivers away from the site of a highway accident. But the model would also have to calculate how many people would take its advice, lest it end up creating a new traffic jam on an alternate route. She says that understanding the way systems change as humans react to incentives is one of the big challenges for mathematical modeling.
Of course, it’s never going to be easy to accurately predict what people–or businesses–will do. But thanks to their insights as mathematicians and their access to IBM’s vast computing power, Dietrich and her colleagues are getting better at it. And now, other companies are paying for that skill.
William M. Bulkeley is a former Wall Street Journal reporter who is now a freelance writer in Boston.
The new version of GPT-3 is much better behaved (and should be less toxic)
OpenAI has trained its flagship language model to follow instructions, making it spit out less unwanted text—but there's still a way to go.
A horrifying new AI app swaps women into porn videos with a click
Deepfake researchers have long feared the day this would arrive.
We can’t afford to stop solar geoengineering research
It is the wrong time to take this strategy for combating climate change off the table.
Meet Altos Labs, Silicon Valley’s latest wild bet on living forever
Funders of a deep-pocketed new "rejuvenation" startup are said to include Jeff Bezos and Yuri Milner.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.