NANCY POTOK IS MAKING CHANGE HAPPEN IN A VERY HIGH-stakes environment. As the chief operating officer of the US Census Bureau, she is leading a transformation of her organization that will have a major effect on how the Census Bureau conducts its highest-visibility activity, the decennial census. The General Accounting Office (GAO) listed the 2020 Census, in a November 2015 Congressional hearing, as “high risk.”

Potok’s strategy is both to transform the Census Bureau and to remake the way Census workers operate for the decennial census. These changes include reengineering scores of internal statistical processes to use the same platforms across the enterprise, using operational data and modeling to improve efficiency, and increasing the use of technology to reduce manual labor. The goal is to save $5 billion in costs for conducting the 2020 Census.

It all sounds like an impossible strategy. An example of changes in store for the decennial census is how the more than three hundred thousand people who knock on doors, the “enumerators,” will approach citizens for interviews. They won’t, as in the past, jump in their cars and rely on instinct and happenstance to guide them to people’s doorsteps. They’ll get their schedules and routes on their smartphones, travel the route prescribed, and arrive at homes at the best possible times to conduct the interview. The machine intelligence of the mathematical corporation will all the while be whirring in the background.

With so many Census workers motoring around, Uncle Sam’s tab can climb steeply. During the last census in 2010, the mileage driven by all enumerators clocked in at five times the length of all roads in the United States, in part because the enumerators were left on their own to plan when and in what order to visit households, which gave rise to inefficient routes. They routinely arrived when nobody was home, leading to many repeat visits. And because they kept their records on paper, they had to meet and hand over their paperwork every day to supervisors. This was an expensive way for the Census to conduct its business, with a lot of extra expenditure on gas, time, and energy.

This and other decades-old Census processes raised the question: How much could a transformation of the way the Census works save the US taxpayer? Not just for enumerator travel but for every part of the operation. And how innovative could Potok get her people to be? For example, was there a way an algorithm could infer when an adult was most likely to answer an enumerator’s knock on the door—and could the enumerators be scheduled and routed to optimize what the model was predicting?

The answer to that question, and many similar ones, is “yes.” Potok and the director of the Census Bureau consulted with outside experts in operations research, geography, technology, modeling, and data linkages who were assembled by the National Academy of Sciences. They also spoke with many industry experts. The experts helped the bureau to craft an approach to revolutionize census taking. That meant Potok, who has a PhD in public policy from George Washington University and who began with the bureau in 2012, could make a big promise to the US Congress and administration: with money to invest in research, testing, and development early on, the bureau will pare the budget for the 2020 census to $12.8 billion. If the bureau were to run the census in the same way as it had in 2010, it would cost $18 billion.

In other words, the new plan proposed to use machine intelligence, allied with operations research, predictive modeling, existing smartphone technology, and redesigned business processes—all enabled by her new leadership approach—to help save $5.2 billion over the twelve-year Census cycle.

Potok is undertaking this transformational challenge under the especially watchful eye of the US Congress. Recall the 2010 census, when the Census Bureau cut short the use of a handheld device to conduct the door-to-door part of the enumeration. That last-minute design change cost the US taxpayers $1.3 billion in cost overruns. The Census had to fall back on its paper-based system, following an assessment by the Secretary of Commerce and the General Accounting Office that the electronic project had a “high risk” of failure.

This time around, Potok, an award-winning public administrator, is leading the organization through a transformation of breathtaking proportions. The question for her, however, is the same as for many: How do you best lead your organization in successfully executing the new strategies of the mathematical corporation?

At the Census Bureau, a lot is riding on the answer. The decennial census produces numbers that affect the very basis of democracy. Its population count governs the apportionment of seats in Congress as well as the redrawing of Congressional districts. It also affects the distribution of billions of dollars of federal funding to states and municipalities. On top of that, the Census Bureau delivers a full slate of statistical information on the economy, including monthly economic indicators, a census of all businesses every five years, and streams of data needed by its sister organization, the Bureau of Economic Analysis, to calculate GDP. It also produces the annual poverty rate, data from the American Community Survey, and about twenty-seven surveys for other federal agencies on topics such as crime, health, education, labor, and energy.

Every organization faces challenges in pursuing its particular vision of the mathematical corporation. In her quest to transform the Census Bureau, Potok aims to remake the organization into an engine of innovation. But she faces a basic challenge right up front: an aversion to risk. No different from other federal agencies—and many corporations, for that matter—people in the Census Bureau aren’t enamored of rocking the boat, especially because members of Congress often take agency leaders to task for committing errors. But without reshaping a risk-averse culture, Potok could not begin to spur the changes that would lead to the biggest possible gains.

One of the most notable ways Potok has pursued her strategy is through putting in place an enterprise risk management program that includes a risk-rating system. Every risk the Census Bureau faces—project risk, program risk, cross-program risk, and enterprise risk—is rated. A standardized tool helps Potok’s team log risks, both big and small, posting them so everyone in the agency can see them. Potok then encourages people to work together to mitigate problems before they mushroom. That means people today are less likely to bury bad news.

Surfacing and handling risk in this fashion gives managers and front-line people the confidence and courage to move forward with the new vision. When charged with a project’s success, people can address risks as routine business practice. That eases their resistance to change, and it also avoids explosions of after-the-fact time bombs that could have been defused earlier. “If there are things we can do to mitigate [risks], there’s no shame in asking for help,” Potok says. “If people hide risks, or deal with them by themselves until there’s a crisis, that’s the problem.”

During preparations for the 2010 census, the organization’s lack of readiness remained hidden for too long. And that’s why the Census Bureau had to scrap the handheld electronic system, the Census’s centerpiece for the door-to-door household enumeration. The General Accounting Office broached the topic of risks, but by then it was too late to do much. The GAO faulted the Census Bureau for “long-standing weaknesses” in management and not testing its system under “operational conditions.”

For any organization, the foremost challenge in threading the future power of the mathematical corporation through an organization is shaping a new mind-set. Leaders have to sometimes act as evangelists to sell the change and remake policies to institutionalize it. “One thing we learned,” says Potok, “is that people take in information about change really differently, the way they internalize it.… People want to know immediately, how does this change what I do as I walk in the building each day? Some people love change; others just want to be told what to do.” And of course still others drag their feet in resistance.

Potok says that, above all, the goal in executing the Census Bureau’s new strategy requires getting people to innovate in the rapid prototyping steps: ensuring they ask the right questions and experiment quickly to get productive answers. Says Potok: “The question is always, How do we engage people? And how do we engage supervisors so they pass on positive and not negative feelings to employees” during this makeover. Organizations that don’t succeed, she says, do not sustain the effort by addressing this aspect of “organizational health.”

In our work with organizations, we suggest a number of moves you can make to realize your vision for the mathematical corporation. Though no recipe is right for everyone, we have found that three moves are critical above others: selling the vision, igniting culture change, and nurturing talent. When you attend to all three, you will have the human and behavioral foundation to drive your future strategies.

SELLING THE VISION

Many time-proven change management principles apply when you implement machine intelligence strategies. Among them is establishing a vision and achieving buy-in across the organization. In other words, the basics of leadership still matter as you prepare to help people transition to a vision that requires the mastery of new capabilities.

To implement the vision for a reengineered 2020 census, Nancy Potok has created the environment in which four major change initiatives could succeed. The first initiative is comparing aerial and satellite imagery with the Census Bureau’s maps and address file for the United States, looking for changes in roads and buildings. If the images show nothing has changed, nobody has to go out on foot to verify the status of a residence. That alone will save $450 million and eliminate up to 75 percent of the work by people who in previous censuses were paid to walk every single street in the United States before the census took place to verify addresses, a labor-intensive yet critical first step in the decennial census.

The second initiative is to make it as easy as possible for people to respond to the census on their own to reduce the number of households that require an in-person visit from an enumerator. Using the Internet, rather than mailing everyone a paper questionnaire, can make responding to the census more convenient. In one remarkable effort, the Census has been developing a means to verify the identity of people who respond without their unique Census-provided passcode, for example, using their smartphones or at a kiosk, and in that way allow them to respond to the census wherever they are. This will help the Census count certain hard-to-reach populations.

An especially frustrating demographic that is easily undercounted is footloose eighteen- to twenty-five-year-olds, who may hop from place to place, often staying just long enough to warm a friend’s couch. These young adults may be willing to respond on their smartphones instead of using a questionnaire mailed to the house. Another group that could benefit from innovative ways of counting is those in rural or low-income areas who don’t have access to the Internet at home but who would be willing to fill out the questionnaire if laptops or kiosks were set up at places of worship, community centers, or other gathering spots. In any case, people will also be able to respond over the phone through a network of call centers, and if all else fails, they can request a paper questionnaire.

The third major initiative is the enumerator effort described earlier, which incorporates smartphones, tablets, operations research, predictive modeling, and other techniques to make the field operations more efficient. And the fourth initiative at the Census is expanding the use of government data sets—from Medicare to the Veterans Administration to the Social Security Administration—to count people that enumerators would otherwise need to visit multiple times in person. Up to 6 million households that choose not to self-respond can be accurately counted without any direct contact with enumerators, says Potok—simply through reusing data that people have already provided to the government. Existing records can also help determine whether nonresponding households are actually vacant, saving yet more time and money on enumerator visits.

The remake of the Census shows the scale of the opportunity for any organization. But how do you go about implementation? In our work, we have developed a diagram to chart an organization’s progress in implementing successively more sophisticated elements of their machine intelligence strategy. See Figure 6.1. Most organizations remain in the analytics era, at the left. They retain data in functional silos within rigid databases, blocking the open, inventive exploration needed by the mathematical corporation. From there, however, organizations can proceed in steps, acquiring steadily better capabilities in discovery, prediction, and prescription of strategic actions.

As a way to understand the diagram, recall the strategy of Merck and its efforts to optimize vaccine manufacturing across its entire global supply chain. Merck collected over a billion data records covering five years but had no way to easily make sense of how one manufacturing step (e.g., sourcing) affected another (e.g., fermentation). The company could use the data for the second phase, describe, but only for each isolated phase of manufacturing. In recent years, however, it has pooled the data in a data lake and can look for new patterns, the discover phase. From the pooled data it creates models to forecast results of changes, the prediction phase. Merck is moving to where machine intelligence provides new insights to run the entire business, the goal of the Singapore plant, as mentioned in Chapter 2.

FIGURE 6.1 The Machine Intelligence Maturity Model

Source: Booz Allen Hamilton.

Whatever stage your organization is in, to progress, you need to create buy-in, getting other leaders and managers on board. This includes identifying supportive sponsors and advocates. Achieving buy-in may also include hiring a chief data scientist, chief data officer, or other C-suite executive to add a second leadership voice as a machine intelligence evangelist. This leader will also have the expertise to implement the new technology structure to get the most from the new strategy. At the Census, where the major manufactured product is data, Potok plays this leadership role. “You have to set the tone to make sure the next level of leadership is leading each one of the business lines,” she says. At the Census Bureau, this includes not only the business line leaders but also the chief information officer (CIO), the chief financial officer (CFO), the chief administrative officer (CAO), the head of communications, and the head of research and methodology.

Merck’s manufacturing chief information officer Michele D’Alessandro plays both roles. As she works toward the vision of advancing predictive capabilities to such a level of confidence that the FDA will approve manufacturing changes based on machine intelligence alone, she says advocating for technology is the easier part. “The technology piece has been fun and interesting, but it’s not the harder part,” she says. “The challenge has been more of a mind-set shift.”

D’Alessandro says that, even for technology people, the idea of pulling in raw data, without formatting or organizing it in traditional ways, has been a leap. Most of her people are accustomed to making sense of data after sorting it into a well-defined taxonomy. Using raw and unstructured data for discovery requires both a new mind-set and skill set. “The people aspect is one of the most challenging,” she says, especially because data scientists are so hard to find and hire.

The most sure way to achieve buy-in, of course, is running prototypes as you develop strategy. If you don’t know where to begin, prototypes generate momentum. A lot of people, even when they support machine intelligence, don’t know what’s possible. When they see new powers in the offing, and when they comprehend the outlines of whole new businesses, excitement about the future builds. Prototypes make that happen.

Potok at the Census Bureau has taken particular pains to advance the state of the art with prototypes. Among the biggest risks in her recent effort has been changing the enumerators’ jobs. In 2020, enumerators will get marching orders that are based on lessons learned and models created with updated 2010 census data. Algorithms, not supervisors, will issue instructions every morning and real-time course corrections during the day, all based on each enumerator’s location, time of day, workload, and progress.

Supervisors will get alerts when the enumerators stray off-course, and they will be expected to contact their charges to keep the work on track and help solve problems that arise in the field. The machine intelligence–enabled systems will comb through the reams of data so that supervisors can spend their time reacting to the alerts rather than trying to decipher where they need to invest their time supervising.

Working with the Census Bureau director, Potok set up a high-visibility team to develop prototypes. The Census Bureau has run trials of the new system in Phoenix, Los Angeles, and Houston. It will test its integrated system end to end in 2018 before rolling it out in 2020 to hundreds of thousands of temporary workers. Potok stresses the new machine intelligence–enabled mission for the Census Bureau is the only way to save that $5.2 billion. The lessons learned and platforms developed eventually will also be adapted and used for all the bureau’s programs, creating even greater efficiencies and ability to meet new demands. “There are a lot of legacy things people want to keep,” she says, “but you can’t afford things that are not priorities anymore.”

IGNITING THE MACHINE INTELLIGENCE CULTURE

After selling the vision and getting buy-in, next is reshaping culture. Your goal is to create a workplace environment—at all levels—that spurs people to innovate until they find a path to the biggest possible gains. “Working with the senior leadership team is a constant effort,” Potok says. She holds weekly operational meetings and encourages everyone to keep the big picture in view and to communicate. There are also periodic offsite sessions to assess progress, tweak the tactics, and recommit to the change. “It’s a lot of hard work, conscious and intentional, to build the leadership team,” she adds.

For the mathematical corporation to thrive, you need to build three elements at every level of the organization. The first is getting people to champion the right to be wrong. The second is working in diverse teams, or what we prefer to call “asymmetric” teams. The third is spreading influence through the machine intelligence story. When you and everyone else in the organization work and think using these elements as a guiding framework, you can engage people to make the big changes to win.

We at Booz Allen don’t just give advice on making this kind of cultural shift. We had to shift our culture ourselves as we built our five-hundred-person machine intelligence organization. And as we implemented new strategy, we learned many lessons. The most notable was that mathematical organizations have to unlearn a few pieces of conventional wisdom. One is that doing a job effectively doesn’t always mean doing it once, getting the right answer, and moving on. We learned instead—and we learn every day—to rededicate ourselves to the experimental method. Many organizations won’t take on “experiments” unless people can make a business case for them. But as a data-driven organization, we support short-duration pilots to see whether they hold promise—and then we bet big.

When it comes to culture change, sometimes you face a chicken-and-egg situation. Does the culture come first? Or does the experimental data science process? As you can see from the examples in this book, they go hand in hand because no one has the future entirely figured out yet. Like everything about the mathematical corporation, the spirit of experimentation remains paramount.

The Power of Being Wrong

Curiosity is the X factor in the data science success equation. Championing the right to be wrong encourages it. If someone is intrigued enough to experiment, you want to extract the value from the results, whether they yield a eureka moment or not. Publicizing results even of so-called flops demonstrates that errors in the search for insight won’t be punished. If you praise only the winners, you’ll obviously zero out the X factor.

At the US Census Bureau, Potok insists on tolerance of errors as key to the Census’s new strategy. When she put the innovation team together for remaking the enumerator process, she gave them specific instructions: “You’re prototyping, so you can be as creative as you need to be, so take risks,” she said. “But let us know what they are. We’ll be there every step of the way.”

Without a focus on managing discrete risks, the Census Bureau would have had a devilishly hard time changing the way enumerators operate. Changes to the enumerator role carry such risks that even the GAO noted it; yet Potok expects it alone will save $1 billion in the 2020 census. The Census Bureau is taking other risks that, although less visible, are every bit as transformative. One is changing the standard way the agency collects and provides statistical data.

For its entire existence, the Census Bureau has relied mostly on surveys. The surveys ask a sample of the population a set of questions, and then statistics are used to calculate economic and demographic numbers for the entire population with a high degree of confidence. Now, with data sets drawn from the growing universe of data, the Census can use machine intelligence to calculate numbers from a full representation of reality. The data science approach can be far less expensive because, unlike with surveys, the Census Bureau won’t have to send people to follow up with businesses and households to ask for survey responses and then spend added time following up with all the people who don’t answer their doors (or phones or letters).

“What we’re trying to do is provide a bridge for the data scientists to understand statistics,” she says, “because they’re used to working with 100 percent data sets and not statistical samples. They also are not subject matter experts in economics, demographics, sociology, and other disciplines needed to understand the nuances of the data.” Meanwhile, she says, “We have to get statisticians and survey methodologists to pick their heads up and look around and see different ways of doing their work, aside from just surveys.” Those different ways are all new territory for longtime Census employees. Both the data scientists and statisticians have to be ready to take risks and to be wrong during the research and prototyping phase.

This tolerance for experimentation and risk taking is the mind-set that an industry colleague of ours, Claudia Perlich, brought to her post as chief data scientist at Dstillery, where she and her team run hundreds of small experiments a day. Dstillery maps people’s brand affinities by analyzing browsing and location information across people’s many devices. The results of this analysis enable sophisticated advertising and targeted marketing. Perlich praises the visibility of failures: “One requirement to advance is to be willing to record failures at the same rate as successes. If you only have data on success stories, and failure falls to the wayside, you’ll never get a clear picture.” As obvious as that is, it remains a perennial challenge in most organizations. Machine intelligence “won’t help you if… people [don’t] admit when things didn’t work,” Perlich says.

At Booz Allen, we take this thinking to heart. We tell our data scientists to set sail on a voyage of discovery, to blend inductive pattern recognition with deductive hypothesis-generated discovery. If they tumble down rabbit holes of interest and end up failing, we want to learn from those failures. We found that setting such house rules creates momentum and energy that permeate the organization and drive the journey onward.

We seek to disabuse people of the notion that every project has to succeed. If you prototype a product and it crashes, the company exits the experiment having risked only a modest investment. One time we killed the development of a healthcare crisis tool. Another time we dropped a geospatial analysis tool—with very little money lost, but wisdom gained. The benefit of encouraging bootstrap efforts is that, even as we publicize successes, we also publicize the value of being wrong. And, instead of betting, say, $1 million on a big-bang effort, we bet table stakes on a trial balloon.

Early on, Booz Allen, like every institution, had the tendency to want every effort to be successful. Unsuccessful efforts weren’t tolerated, so much so that we often tried to turn losers into winners, albeit sketchy ones. That typically meant throwing good money after bad. Now we champion thinking that breaks barriers in all directions. This is elemental to executing the strategies of the mathematical corporation.

Motivating people to undertake an entrepreneurial, experimental journey is not a matter of simply paying good salaries. The journey, no matter the ups and downs, is where the fun lies—especially for tech-minded people who like to “geek out.” To attract the right people, we raised the stature of the risk-taking element of our culture, establishing reward and recognition to spur people to engage in projects they are passionate about. The payoff is in exploration, discovery, insights, and predictions based on new secrets in the data. Send the message corporation-wide that inventing new ways of doing things can create more value than the success of a single project.

Birds of Different Feathers

Succeeding in capturing the future power of the mathematical corporation requires a diverse team because no single person knows all the computer science, data science, or operations of an organization. Moreover, in many organizations, the people closest to the mission, leaders at the top, do not understand the full potential of machine intelligence. Meanwhile, people seasoned in data science do not always understand the mission well enough to spot opportunity. That’s why we like to say you need to create asymmetric teams.

Smart and productive experiments, the stepping-stones in the data science journey, need to blur the lines between different types of expertise. At Booz Allen, we often include organizational designers, strategists, operations experts, and human capital specialists on our machine intelligence teams. You never know who has the skill to see opportunity in a synthesis of talents.

One of our best data scientists is a forester by training. He previously worked for a forestry consulting company that needed him to write predictive models. We hired him as much for his creativity and inquisitiveness as for his math skills. We think it helps that he likes all things creative—furniture making, photography, cooking, guitar playing, and writing code. You never know where someone’s epiphany for solving a problem will come from.

One common misconception is that data scientists should operate as a service group. This would mean that “birds of a feather flock together,” but that is too often counterproductive because they all sing the same song. Better that people with various skill sets combine their skills in developing hypotheses, deciding on data inputs, validating assumptions, and interpreting results. There should be no “us” or “them,” in which each side believes the other doesn’t understand its realities. Leaders, managers, data scientists, and other experts have to come together to push each other out of the comfort zone to execute the new vision.

As skills related to machine intelligence become required in a host of jobs, people will need plenty of ways to learn to become vibrant players on the asymmetric team. At Booz Allen, we learned that “hackathons” could both challenge our people and offer a safe learning environment for novices to develop machine intelligence skills. Hackathons also serve as a way to accelerate the transformation of Booz Allen into a mathematical corporation. We often hold hackathons as part of the company’s pro bono work, putting scores of people in a room (or several rooms in several cities) to work on the compelling problems of our times.

In 2015, we helped Polaris, an anti–human trafficking nonprofit, figure out how to scrape the Web and identify prostitution kingpins who operate multicity businesses disguised as massage parlors. (More in Chapter 8.) We helped the Center for Prevention of Genocide, a unit of the Holocaust Museum in Washington, DC, create an algorithm to rate the probability of genocide in scores of countries. (Again, more in Chapter 8.) And we helped the US Agency for International Development (USAID) develop ways to prevent crime in Latin America. One project explored the predictive power of real estate prices and Google trends data to predict crime pattern shifts.

We also tended to the basics. One effort we launched is called Explore Data Science, an online, self-paced training program. Another is a nine-month, part-time apprenticeship program in which people continue with their day job but do data science projects in the background, work with mentors, connect with senior leadership, and go through structured training to acquire journeyman-level expertise. To date, our apprenticeship program has propelled dozens of people into the field of data science, expanding our machine intelligence team by roughly 10 percent in two years. This ensures we get a continuous influx of new thinking in the work of the mathematical corporation.

Although hard to believe, every professional who works at a desk today will, in the future, use machine intelligence. Even front-line employees will run experiments on their own as software enables drag-and-drop coding to be as common for professionals as word processing or website design is today. Booz Allen has made this future vision an operating assumption. We now use software to allow people to reuse packages of computer code to solve analytical inquiries.

Even more promising are software tools that let people of any level of expertise query data sets from inside and outside the organization using natural language processing. Individuals can pose questions and benefit from off-the-shelf algorithms to come up with answers. Our colleagues’ work in this area shows that the sooner companies democratize the culture so that everyone uses machine intelligence, the more adept the company will be at executing its strategies. Within a matter of a few years, most professionals will run machine intelligence software as routinely as they run Microsoft Excel or Word today.

Data-Driven Storytelling

No matter your organization, and whatever your vision, you need to give people a story about how to get there and how the organization will wield future power. People want to know: Why are we undertaking this new data science–driven effort? How will we use data for discovery and prediction? How do we vet and control the machine intelligence experiments? And what are we learning about the advantages of collaborating with the machine?

You may not fully understand the workings of all of the programs yourself. But the story still needs to incorporate the power and use of your data. Jim Dinkins, chief of the anti–money laundering organization introduced in Chapter 4, says the bank relies on his data scientists and other experts to take advantage of transactional data because you need experts that understand the nuances of building monitoring programs as well as governance structures to test and validate the accuracy of models. These are skills outside of his expertise. But his experience is in knowing how criminals seek to exploit financial institutions—and that drives the story. “I can tell you the behavior behind the crimes,” he says. “The data scientists can translate that into transactional activity and monitor for indicators of the type of crimes that could actually be happening.”

As the leader, you need to initiate and guide the story, to ensure that the capabilities you seek are enough to inspire and advance the organization’s mission. At one company we work with, the executive team decided to name the head of strategy and innovation—not the CIO—as the leader of the machine intelligence effort. That’s because the strategy and innovation chief could paint a vivid picture of the new customer experience for airline passengers, and he had the resources and ability to motivate people to make it happen. He was charismatic, and he made it a point to repeatedly inspire people with the company’s new vision. The CIO was an important figure, an enabler of technology changes, and he co-chaired the effort, but the innovation leader spurred action with story.

On the flip side, another company we worked with created a machine intelligence capability that had not taken well. Users could not see the connection between the strategy and the daily work. Instead, all they heard about was the whiz-bang tools they would get and how much faster and easier work would be. We then teamed with them to create an internal marketing program, shaping the story about how machine intelligence would enable specific advances in the business. People in the organization then bought right in because they understood exactly why the new work fit the new strategy and would solve their business problems.

When shaping the story, make sure to explicitly include hypotheses you seek to test. You should also include the utility of results from each experimental cycle. Especially important is chronicling the flow of questions. The story, like all good stories, follows an arc, from the first question to a final big question, followed by a climax of execution and resolution.

Remember, you’re the protagonist in this story, the one who is primed and pushing for change. You are pioneering ideas, placing bets, making decisions, and facing moments of truth in taking people on the data science journey. Without a story, you risk losing people in the flow of what can seem like geeky experiments. The story convinces people to invest in new ideas, acquire new skills, and launch new efforts, whatever the headaches, to make a great transformation happen.

The story is your prime means of influence. The first law of organizational psychology is that an organizational body at rest tends to stay at rest unless a force is applied—by the leader. So, you need to apply the force and support it with a story. If people feel like they could look like stooges in an ill-fated joy ride, they won’t go along. They want to be part of a winning team.

Without the news of a few discoveries or predictive models to hint at how the story will turn out, people will remain unconvinced. You won’t be able to galvanize them to roll up their sleeves and make the vision real. One of the best ways to keep the momentum going is to demonstrate quick wins. Often this means betting first on experiments that promise easy gains. These may be entirely tactical, not strategic.

If all goes well, you get the whole organization telling the story. The story becomes a tool for vast, self-reinforcing influence. One challenge is that sometimes it takes almost blind faith to invest in ideas people have never seen work before. How do you instill trust in machine-learned models that don’t reveal how they work but in fact get great results and make the strategy come alive? The story might be a speculative vision to start, but it still must ignite people’s interest.

Another hurdle in getting everyone in the organization to tell the same machine intelligence story is that computer models have a history of making mistakes. We’ve all laughed at the absurdity of some computer results. But with greatly enlarged technological power this has fast changed. So it’s time to rein in biases against the use of data sets, algorithms, and models by conveying the possibilities of huge advances.

A related challenge is that machine intelligence, though based on math, doesn’t actually operate without bias. Biases come from data sets that are too limited to actually make the discoveries or predictions wanted. They also come from “dirty” data sets, ones too full of errors and omissions to promise the desired reliability. They even come from the biases of the data scientists who built the models, which of course require a host of assumptions. You can be fooled, and so can your followers, if you don’t remain skeptical. So the story needs to include a chapter on how people fought the goblins of bias and kept on the road to finding truths that matter.

Despite the claim that data scientists can predict anything as we gain data, the truth is that data sets will always limit how far we can go. We cannot predict the future of climate change reliably, for example, by simply employing bigger data sets because we don’t have historical data from similar events that would enable computers to learn the patterns and, in turn, create reliable models. You have to be on the lookout for when the data just won’t perform the magic needed.

Expecting too much of the data can go along with expecting too much of the model. All else being equal, putting several models together in an “ensemble” gives you more accurate results. That’s why you do not want to let your people fall in love with one model. More complex models may be more accurate, but they may also have more limited forecasting ability. Models that are too simple may not analyze an issue from enough angles, or dimensions, to avoid oversimplification. All models also need cross validation.

A common modeling risk is overfitting: the model fits the historical data so well that it can only “forecast” the past—and it can’t even do that perfectly because of inevitable errors in the data. You need to talk with your data scientists: What is the right story to tell about the usability of the data and model?

The story you tell is a new kind of story. Around the campfire, you make sense of events by interpreting a reality full of cause and effect, one thing leading to another, a line of sequential logic. Your interpretation is based on assumptions about how you understand the world from experience. We default to seeing patterns we think we should see, and we set aside uncertainties and clean up ambiguities by glossing over particulars. Our stories get shaped, embroidered, streamlined.

We do this so we can get a firm grip on reality. As an example of our preference for neat stories, consider the myth of the “hot hand” in basketball. The story is an article of faith in some quarters: players get “hot” and sink more shots. Except that this story is not true: statistics shows it to be false. Instead, the human brain plays tricks on us. It sees patterns where there are none. It attributes skill to episodes of luck. That’s the way of Homo sapiens.

In the mathematical corporation, you need to tell a story about novel questions and data-supported experiments with results that support changes in your strategy. Again, we agree with Dstillery’s Claudia Perlich, who notes that the first step in a leader’s education in machine intelligence is this: “Step back and say, ‘There are things I know for certain, but maybe they’re not as certain as I believe.’ An honest exchange with data may teach you about the world and actually teach you to be more creative and innovative. If the things you think are true don’t predict future, maybe they aren’t reliable. If you believe in the hot hand in basketball, can you predict if the next ball will land in the basket?”

We can all be fools for stories we manufacture, so you need to create a new culture of telling the machine intelligence story with a sequence of facts driven by the data. Among the most common errors is the so-called confirmation bias. We look for evidence of “facts” that we already believe are true. Even in the face of contrary evidence, we dismiss data that disprove our view. So you need to engage in a dialogue with data scientists to keep from falling into this trap when the data lead you away from your intuition.

Another failure is mistaking correlation for causation. In data science, the mass of data, adequately “tortured,” reveals correlations of all kinds. For fun, blogger Tyler Vigen launched the Spurious Correlations project, posting on his website unlikely correlations to make this point. There is a correlation between cheese consumption in the United States and the number of people strangled in their bedsheets. These numbers align extremely well. But does one cause the other?

Alas, the human mind, looking for cause and effect, often starts down wrong roads—and keeps going to a sketchy end. Ultimately, the story of how machine intelligence works serves not just as a tool for executing your new strategy but also as a safety net to prevent you from falling into the pitfalls of data analysis. Putting the story together reveals disconnects between data quality, model assumptions, model outputs, and so on. You need to establish a rationale of why you’re undertaking the machine intelligence effort, in what way, and why it will work. If the story has problems, the approach with machine intelligence may have problems.

IGNITE THE TALENT

The third major challenge in executing your impossible strategy is finding the talent. While hard-core data scientists make up only part of a winning team, they provide the algorithmic fuel to make the engine go. In today’s world, in the rush to ramp up in a talent-starved environment, you need to make a point to be systematic in finding and managing the right people. We know this from our own experience of developing our machine intelligence team. Competing for the best people—and the right people—demands more than posting job announcements.

We learned quickly that we couldn’t just recruit “algorithm ready” talent, but that we had to look for people with that potential, even if far from realized. Fact is, it can sometimes be easier to find experts in biology, geology, or medicine and train them in machine intelligence than to do the reverse. So we went deep into an analysis of what makes good data scientists, and in turn developed a system to identify, attract, grow, and retain them.

When we began a concerted effort to build our machine intelligence capabilities back in 2014, no scientifically valid model of machine intelligence competencies existed, and we found that unstructured interviews weren’t enough to create an itemized list of the skills and abilities of competent data scientists. So we started by asking our existing data scientists to rate the importance of four hundred skills, areas of knowledge, abilities, and personality traits. Is data visualization important? Is communications? How about perseverance? Our data scientists helped us winnow the long list to a few key items.

On many counts, there were surprises, as shown in Table 6.1. Note how some of these competencies are the same cognitive skills we stressed in Chapter 2 for leaders. The survey gave us confidence in specifics. It also explains why two of our top data scientists are not what you would expect. One, who specializes in natural language processing, spent a year teaching English in Shenzhen, China, and also lived a few years studying and teaching yoga in Thailand. Another was an Army officer, bar owner, baccarat croupier, and ski lift operator. These skills are, of course, key to our asymmetric teams.

When it came to technical skills, statistical modeling and machine learning topped the list—just the skills inherent in many examples in this book. When it came to machine intelligence collaboration, key skills included teamwork, communication, and expertise in a domain (health care, oil exploration, logistics, and so on). As for cognitive skills, problem solving and reasoning came out as tops. For personality, creativity, curiosity, flexibility, and tolerance for ambiguity ranked high. This list conveys the essence of what it means to be a data scientist.

And what was the most important trait? Flexibility in overcoming setbacks and the willingness to abandon an idea and try a new approach. Experimentation to find the right approach—as in the way Thomas Edison experimented with cotton, linen, wood, and bamboo to find the right material for the carbon filament for his light bulb—requires people undeterred by interim stumbles. Even though this seemed intuitive to us because of the central role of our experimental approach, the competency profile was the first definitive study to show we were right about flexibility being so important.

Our study, of course, applies not just to Booz Allen. It applies to all companies assembling the talent for the mathematical corporation. Any company can start using this profile right away to acquire the talent it needs to acquire the future power of machine intelligence.

We took the next step and created a system, or “talent management model,” for finding, grooming, and retaining the best possible talent. It answers key questions: Whom do you need? Where do you need them? How do you keep and develop them?

An alternative to hiring people is crowdsourcing projects. We’ve covered two examples in detail—the two Data Science Bowl competitions convened by Booz Allen and Kaggle, a global matchmaker that links data scientists to machine intelligence projects. The Data Science Bowl challenges were a way to throw difficult, as-yet-unsolved problems to the world’s best data scientists. The first used a data set from the Hatfield Marine Science Center in Oregon to identify ocean organisms from millions of photos taken by a camera towed underwater behind a research ship.

The second used images supplied by the National Institutes of Health and Children’s National Medical Center to calculate the human heart’s ejection fraction, the percentage of blood pumped from the left ventricle with each heartbeat, a critical measure of heart health. In the heart competition, more than seven hundred teams of data scientists competed for ninety days. The top three teams—all of whom provided what National Institutes of Health cardiologists called “excellent” results—earned a collective $200,000.

The top team, US hedge fund analysts Qi Liu and Tencia Lee, won $125,000. The remarkable aspect about the winning team was that neither teammate knew anything about cardiology before the competition. Never before have organizations had at their disposal the global pool of talent to tackle the most complex problems of our time—including problems in fields of knowledge that data scientists know nothing about. Kaggle runs competitions as its mission, and it is joined by other organizations like Topcoder. This is an entirely new way to access talent.

Some organizations, committed to crowdsourcing, have made it a point to develop an additional skill: figuring out how to break complex problems into parts that can be farmed out piecemeal. Problems can then be solved one bite-size piece at a time. Talent acquisition becomes less about hiring people and more about divvying up and distributing problems to a worldwide talent base. The organization then assembles the solution parts to close the loop.

At the US Census Bureau, Nancy Potok has taken this idea of tapping the informal economy in-house. The Census sponsors internal competitions. Employees who compete to solve problems are awarded investment money to work on the projects of their choice—so long as they show a three-year payback. All employees vote on the proposed projects, and senior managers choose the final investments. Potok says some of the Census Bureau’s biggest advances come from these internal contests. For example, a sponsored project led to the Census being one of the first federal agencies to implement APIs to make it easy for outside developers to use Census data in their own operations.

Other activities will help you implement the strategies of the mathematical corporation, but keeping your eye on selling the vision, shaping the culture, and developing the talent remain central. In some ways, these are all about elevating people.

Even as the future power of machine intelligence disrupts industries and organizations, it provides a basis for people to rise to a new level of performance. Success depends on everyone in the organization developing new skills and capabilities that are the same as for leaders: how to think in partnership with computers, break obsolete constraints, recognize the new role of technology, see new possibilities for winning strategies, and grow into these new capabilities—jobs that everyone can find rewarding.

When people see machine intelligence as an opportunity for self-growth, not just organizational growth, the notion of the mathematical corporation generates immense organization energy. Instead of speculating that machine intelligence will take away their jobs, partnering with the machine can smarten them up. As Nancy Potok says, “The goal is to remake the way we work… and not change so rapidly that we get anxiety from people instead of excitement.” That excitement is the basic energy source for making the strategy real.

6