Does your data have a purpose? If not, you’re spinning your wheels.
Here’s how to discover one and then translate it into action.
by Helen Mayhew, Tamim Saleh, and Simon Williams
The data-analytics revolution now under way has the potential to transform
how companies organize, operate, manage talent, and create value. That’s
starting to happen in a few companies—typically ones that are reaping major
rewards from their data—but it’s far from the norm. There’s a simple reason:
CEOs and other top executives, the only people who can drive the broader
business changes needed to fully exploit advanced analytics, tend to avoid
getting dragged into the esoteric “weeds.” On one level, this is understandable.
The complexity of the methodologies, the increasing importance of machine
learning, and the sheer scale of the data sets make it tempting for senior
leaders to “leave it to the experts.”
But that’s also a mistake. Advanced data analytics is a quintessential business
matter. That means the CEO and other top executives must be able to clearly
articulate its purpose and then translate it into action—not just in an analytics
department, but throughout the organization where the insights will be used.
This article describes eight critical elements contributing to clarity of purpose
and an ability to act. We’re convinced that leaders with strong intuition about
both don’t just become better equipped to “kick the tires” on their analytics
efforts. They can also more capably address many of the critical and complementary
top-management challenges facing them: the need to ground
even the highest analytical aspirations in traditional business principles, the
importance of deploying a range of tools and employing the right personnel,
and the necessity of applying hard metrics and asking hard questions. (For more
on these, see “Straight talk about big data,” on page 42.1) All that, in turn,
boosts the odds of improving corporate performance through analytics.
After all, performance—not pristine data sets, interesting patterns, or killer
algorithms—is ultimately the point. Advanced data analytics is a means
to an end. It’s a discriminating tool to identify, and then implement, a valuedriving
answer. And you’re much likelier to land on a meaningful one if
you’re clear on the purpose of your data (which we address in this article’s
first four principles) and the uses you’ll be putting your data to (our focus in
the next four). That answer will of course look different in different companies,
industries, and geographies, whose relative sophistication with advanced
data analytics is all over the map. Whatever your starting point, though, the
insights unleashed by analytics should be at the core of your organization’s
approach to define and improve performance continually as competitive
dynamics evolve. Otherwise, you’re not making advanced analytics work
“Better performance” will mean different things to different companies. And
it will mean that different types of data should be isolated, aggregated,
and analyzed depending upon the specific use case. Sometimes, data points
are hard to find, and, certainly, not all data points are equal. But it’s the
data points that help meet your specific purpose that have the most value.
Ask the right questions
The precise question your organization should ask depends on your bestinformed
priorities. Clarity is essential. Examples of good questions include
“how can we reduce costs?” or “how can we increase revenues?” Even better
are questions that drill further down: “How can we improve the productivity
of each member of our team?” “How can we improve the quality of outcomes
for patients?” “How can we radically speed our time to market for product
For more on the context and challenges of harnessing insights from more data and on using new methods, tools,
and skills to do so, see “Is big data still a thing?,” blog entry by Matt Turck, February 1, 2016, mattturck.com;
David Court, “Getting big impact from big data,” McKinsey Quarterly, January 2015, McKinsey.com; and Brad
Brown, David Court, and Paul Willmott, “Mobilizing your C-suite for big-data analytics,” McKinsey Quarterly,
November 2013, McKinsey.com.
development?” Think about how you can align important functions and
domains with your most important use cases. Iterate through to actual business
examples, and probe to where the value lies. In the real world of hard
constraints on funds and time, analytic exercises rarely pay off for vaguer
questions such as “what patterns do the data points show?”
One large financial company erred by embarking on just that sort of openended
exercise: it sought to collect as much data as possible and then see what
turned up. When findings emerged that were marginally interesting but
monetarily insignificant, the team refocused. With strong C-suite support,
it first defined a clear purpose statement aimed at reducing time in product
development and then assigned a specific unit of measure to that purpose,
focused on the rate of customer adoption. A sharper focus helped the company
introduce successful products for two market segments. Similarly, another
organization we know plunged into data analytics by first creating a “data
lake.” It spent an inordinate amount of time (years, in fact) to make the data
pristine but invested hardly any thought in determining what the use cases
should be. Management has since begun to clarify its most pressing issues.
But the world is rarely patient.
Had these organizations put the question horse before the data-collection
cart, they surely would have achieved an impact sooner, even if only portions
of the data were ready to be mined. For example, a prominent automotive
company focused immediately on the foundational question of how to improve
its profits. It then bore down to recognize that the greatest opportunity
would be to decrease the development time (and with it the costs) incurred
in aligning its design and engineering functions. Once the company had
identified that key focus point, it proceeded to unlock deep insights from ten
years of R&D history—which resulted in remarkably improved development
times and, in turn, higher profits.
In the real world of hard constraints on funds and time, analytic exercises rarely pay off for vaguer questions such as “what patterns do the data points show?”
Think really small . . . and very big
The smallest edge can make the biggest difference. Consider the remarkable
photograph below from the 1896 Olympics, taken at the starting line of the
100-meter dash. Only one of the runners, Thomas Burke, crouched in the nowstandard
four-point stance. The race began in the next moment, and 12 seconds
later Burke took the gold; the time saved by his stance helped him do it.
Today, sprinters start in this way as a matter of course—a good analogy for
the business world, where rivals adopt best practices rapidly and competitive
advantages are difficult to sustain.
The good news is that intelligent players can still improve their performance
and spurt back into the lead. Easy fixes are unlikely, but companies can identify
small points of difference to amplify and exploit. The impact of “big data”
analytics is often manifested by thousands—or more—of incrementally small
improvements. If an organization can atomize a single process into its
smallest parts and implement advances where possible, the payoffs can be
profound. And if an organization can systematically combine small improvements
across bigger, multiple processes, the payoff can be exponential.
The variety of stances among runners in the 100-meter sprint at the first Olympic Games, held in Athens, Greece, is surprising to the modern viewer. Thomas Burke (second from left) is the only runner in the crouched stance—considered best practice today—an advantage that helped him win one of his two gold medals at the Games.
Just about everything businesses do can be broken down into component
parts. GE embeds sensors in its aircraft engines to track each part of their
performance in real time, allowing for quicker adjustments and greatly
reducing maintenance downtime. But if that sounds like the frontier of high
tech (and it is), consider consumer packaged goods. We know a leading CPG
company that sought to increase margins on one of its well-known breakfast
brands. It deconstructed the entire manufacturing process into sequential
increments and then, with advanced analytics, scrutinized each of them
to see where it could unlock value. In this case, the answer was found in
the oven: adjusting the baking temperature by a tiny fraction not only made
the product taste better but also made production less expensive. The
proof was in the eating—and in an improved P&L.
When a series of processes can be decoupled, analyzed, and resynched
together in a system that is more universe than atom, the results can be even
more powerful. A large steel manufacturer used various analytics techniques
to study critical stages of its business model, including demand planning
and forecasting, procurement, and inventory management. In each process,
it isolated critical value drivers and scaled back or eliminated previously
undiscovered inefficiencies, for savings of about 5 to 10 percent. Those gains,
which rested on hundreds of small improvements made possible by data
analytics, proliferated when the manufacturer was able to tie its processes
together and transmit information across each stage in near real time.
By rationalizing an end-to-end system linking demand planning all the way
through inventory management, the manufacturer realized savings
approaching 50 percent—hundreds of millions of dollars in all.
Beware the phrase “garbage in, garbage out”; the mantra has become so
embedded in business thinking that it sometimes prevents insights from
coming to light. In reality, useful data points come in different shapes
and sizes—and are often latent within the organization, in the form of freetext
maintenance reports or PowerPoint presentations, among multiple
examples. Too frequently, however, quantitative teams disregard inputs
because the quality is poor, inconsistent, or dated and dismiss imperfect
information because it doesn’t feel like “data.”
But we can achieve sharper conclusions if we make use of fuzzier stuff. In
day-to-day life—when one is not creating, reading, or responding to an Excel
model—even the most hard-core “quant” processes a great deal of qualitative
information, much of it soft and seemingly taboo for data analytics—in a
nonbinary way. We understand that there are very few sure things; we weigh
probabilities, contemplate upsides, and take subtle hints into account. Think
about approaching a supermarket queue, for example. Do you always go to
register four? Or do you notice that, today, one worker seems more efficient,
one customer seems to be holding cash instead of a credit card, one cashier
does not have an assistant to help with bagging, and one shopping cart has
items that will need to be weighed and wrapped separately? All this is soft
“intel,” to be sure, and some of the data points are stronger than others. But
you’d probably consider each of them and more when you decided where to
wheel your cart. Just because line four moved fastest the last few times doesn’t
mean it will move fastest today.
In fact, while hard and historical data points are valuable, they have their
limits. One company we know experienced them after instituting a robust
investment-approval process. Understandably mindful of squandering
capital resources, management insisted that it would finance no new products
without waiting for historical, provable information to support a projected
ROI. Unfortunately, this rigor resulted in overly long launch periods—so long
that the company kept mistiming the market. It was only after relaxing
the data constraints to include softer inputs such as industry forecasts,
predictions from product experts, and social-media commentary that the
company was able to get a more accurate feel for current market conditions
and time its product launches accordingly.
Of course, Twitter feeds are not the same as telematics. But just because
information may be incomplete, based on conjecture, or notably biased
does not mean that it should be treated as “garbage.” Soft information does
have value. Sometimes, it may even be essential, especially when people
try to “connect the dots” between more exact inputs or make a best guess for
the emerging future.
To optimize available information in an intelligent, nuanced way, companies
should strive to build a strong data provenance model that identifies the
source of every input and scores its reliability, which may improve or degrade
over time. Recording the quality of data—and the methodologies used to
determine it—is not only a matter of transparency but also a form of risk
management. All companies compete under uncertainty, and sometimes
the data underlying a key decision may be less certain than one would like. A
well-constructed provenance model can stress-test the confidence for a
go/no-go decision and help management decide when to invest in improving
a critical data set.
Connect the dots
Insights often live at the boundaries. Just as considering soft data can reveal
new insights, combining one’s sources of information can make those
insights sharper still. Too often, organizations drill down on a single data set
in isolation but fail to consider what different data sets convey in conjunction.
For example, HR may have thorough employee-performance data;
operations, comprehensive information about specific assets; and finance,
pages of backup behind a P&L. Examining each cache of information
carefully is certainly useful. But additional untapped value may be nestled in
the gullies among separate data sets.
One industrial company provides an instructive example. The core business
used a state-of-the-art machine that could undertake multiple processes.
It also cost millions of dollars per unit, and the company had bought hundreds
of them—an investment of billions. The machines provided best-in-class
performance data, and the company could, and did, measure how each unit
functioned over time. It would not be a stretch to say that keeping the
machines up and running was critical to the company’s success.
Even so, the machines required longer and more costly repairs than management
had expected, and every hour of downtime affected the bottom line.
Although a very capable analytics team embedded in operations sifted through
the asset data meticulously, it could not find a credible cause for the breakdowns.
Then, when the performance results were considered in conjunction
with information provided by HR, the reason for the subpar output became
clear: machines were missing their scheduled maintenance checks because the
personnel responsible were absent at critical times. Payment incentives,
not equipment specifications, were the real root cause. A simple fix solved
the problem, but it became apparent only when different data sets were
FROM OUTPUTS TO ACTION
One visual that comes to mind in the case of the preceding industrial company
is that of a Venn Diagram: when you look at 2 data sets side by side, a key
insight becomes clear through the overlap. And when you consider 50 data sets,
the insights are even more powerful—if the quest for diverse data doesn’t
create overwhelming complexity that actually inhibits the use of analytics.
To avoid this problem, leaders should push their organizations to take a
multifaceted approach in analyzing data. If analyses are run in silos, if the
outputs do not work under real-world conditions, or, perhaps worst of all,
if the conclusions would work but sit unused, the analytics exercise has failed.
Best-in-class organizations continually test their assumptions, processing
new information more accurately and reacting to situations more quickly.
Observe, orient, decide, and act—a strategic decision-making model developed by US Air Force colonel John R. Boyd.
Run loops, not lines
Data analytics needs a purpose and a plan. But as the saying goes, “no
battle plan ever survives contact with the enemy.” To that, we’d add another
military insight—the OODA loop, first conceived by US Air Force colonel
John Boyd: the decision cycle of observe, orient, decide, and act. Victory,
Boyd posited, often resulted from the way decisions are made; the side that
reacts to situations more quickly and processes new information more
accurately should prevail. The decision process, in other words, is a loop or—
more correctly—a dynamic series of loops (exhibit).
Best-in-class organizations adopt this approach to their competitive advantage.
Google, for one, insistently makes data-focused decisions, builds consumer
feedback into solutions, and rapidly iterates products that people not only use
but love. A loops-not-lines approach works just as well outside of Silicon
Valley. We know of a global pharmaceutical company, for instance, that tracks
and monitors its data to identify key patterns, moves rapidly to intervene
when data points suggest that a process may move off track, and refines its
feedback loop to speed new medications through trials. And a consumerelectronics
OEM moved quickly from collecting data to “doing the math”
with an iterative, hypothesis-driven modeling cycle. It first created an
interim data architecture, building three “insights factories” that could generate
actionable recommendations for its highest-priority use cases, and
then incorporated feedback in parallel. All of this enabled its early pilots to
deliver quick, largely self-funding results.
Digitized data points are now speeding up feedback cycles. By using advanced
algorithms and machine learning that improves with the analysis of every
new input, organizations can run loops that are faster and better. But while
machine learning very much has its place in any analytics tool kit, it is not
the only tool to use, nor do we expect it to supplant all other analyses. We’ve
mentioned circular Venn Diagrams; people more partial to three-sided
shapes might prefer the term “triangulate.” But the concept is essentially the
same: to arrive at a more robust answer, use a variety of analytics techniques
and combine them in different ways.
In our experience, even organizations that have built state-of-the-art machinelearning
algorithms and use automated looping will benefit from comparing
their results against a humble univariate or multivariate analysis. The
best loops, in fact, involve people and machines. A dynamic, multipronged
decision process will outperform any single algorithm—no matter how
advanced—by testing, iterating, and monitoring the way the quality of
data improves or degrades; incorporating new data points as they become
available; and making it possible to respond intelligently as events unfold.
Make your output usable—and beautiful
While the best algorithms can work wonders, they can’t speak for themselves
in boardrooms. And data scientists too often fall short in articulating what
they’ve done. That’s hardly surprising; companies hiring for technical roles
rightly prioritize quantitative expertise over presentation skills. But mind
the gap, or face the consequences. One world-class manufacturer we know
employed a team that developed a brilliant algorithm for the options pricing
of R&D projects. The data points were meticulously parsed, the analyses
were intelligent and robust, and the answers were essentially correct.
But the organization’s decision makers found the end product somewhat
complicated and didn’t use it.
We’re all human after all, and appearances matter. That’s why a beautiful
interface will get you a longer look than a detailed computation with an
uneven personality. That’s also why the elegant, intuitive usability of products
like the iPhone or the Nest thermostat is making its way into the enterprise.
Analytics should be consumable, and best-in-class organizations now include
designers on their core analytics teams. We’ve found that workers throughout
an organization will respond better to interfaces that make key findings
clear and that draw users in.
Build a multiskilled team
Drawing your users in—and tapping the capabilities of different individuals
across your organization to do so—is essential. Analytics is a team sport.
Decisions about which analyses to employ, what data sources to mine, and
how to present the findings are matters of human judgment.
Assembling a great team is a bit like creating a gourmet delight—you need
a mix of fine ingredients and a dash of passion. Key team members include
data scientists, who help develop and apply complex analytical methods;
engineers with skills in areas such as microservices, data integration, and
distributed computing; cloud and data architects to provide technical
and systemwide insights; and user-interface developers and creative designers
to ensure that products are visually beautiful and intuitively useful. You
also need “translators”—men and women who connect the disciplines of IT
and data analytics with business decisions and management.
In our experience—and, we expect, in yours as well—the demand for people
with the necessary capabilities decidedly outstrips the supply. We’ve also
seen that simply throwing money at the problem by paying a premium for a
cadre of new employees typically doesn’t work. What does is a combination:
a few strategic hires, generally more senior people to help lead an analytics
group; in some cases, strategic acquisitions or partnerships with small dataanalytics
service firms; and, especially, recruiting and reskilling current
employees with quantitative backgrounds to join in-house analytics teams.
We’re familiar with several financial institutions and a large industrial company
that pursued some version of these paths to build best-in-class advanced
data-analytics groups. A key element of each organization’s success was understanding
both the limits that any one individual can be expected to contribute
and the potential that an engaged team with complementary talents can
collectively achieve. On occasion, one can find “rainbow unicorn” employees
who embody most or all of the needed capabilities. It’s a better bet, though,
to build a collaborative team comprising people who collectively have all the
That starts, of course, with people at the “point of the spear”—those who
actively parse through the data points and conduct the hard analytics. Over
time, however, we expect that organizations will move to a model in which
people across functions use analytics as part of their daily activities. Already,
the characteristics of promising data-minded employees are not hard to see:
they are curious thinkers who can focus on detail, get energized by ambiguity,
display openness to diverse opinions and a willingness to iterate together to
produce insights that make sense, and are committed to real-world outcomes.
That last point is critical because your company is not supposed to be
running some cool science experiment (however cool the analytics may be)
in isolation. You and your employees are striving to discover practicable
insights—and to ensure that the insights are used.
Make adoption your deliverable
Culture makes adoption possible. And from the moment your organization
embarks on its analytics journey, it should be clear to everyone that math, data,
and even design are not enough: the real power comes from adoption. An
algorithm should not be a point solution—companies must embed analytics
in the operating models of real-world processes and day-to-day work flows.
Bill Klem, the legendary baseball umpire, famously said, “It ain’t nothin’ until
I call it.” Data analytics ain’t nothin’ until you use it.
We’ve seen too many unfortunate instances that serve as cautionary tales—
from detailed (and expensive) seismology forecasts that team foremen
didn’t use to brilliant (and amazingly accurate) flight-system indicators that
airplane pilots ignored. In one particularly striking case, a company we
know had seemingly pulled everything together: it had a clearly defined mission
to increase top-line growth, robust data sources intelligently weighted and
mined, stellar analytics, and insightful conclusions on cross-selling opportunities.
There was even an elegant interface in the form of pop-ups that
would appear on the screen of call-center representatives, automatically
triggered by voice-recognition software, to prompt certain products, based
on what the customer was saying in real time. Utterly brilliant—except the
representatives kept closing the pop-up windows and ignoring the prompts.
Their pay depended more on getting through calls quickly and less on the
number and type of products they sold.
When everyone pulls together, though, and incentives are aligned, the results
can be remarkable. For example, one aerospace firm needed to evaluate a
range of R&D options for its next-generation products but faced major technological,
market, and regulatory challenges that made any outcome uncertain.
Some technology choices seemed to offer safer bets in light of historical
results, and other, high-potential opportunities appeared to be emerging but
were as yet unproved. Coupled with an industry trajectory that appeared
to be shifting from a product- to service-centric model, the range of potential
paths and complex “pros” and “cons” required a series of dynamic—and, of
By framing the right questions, stress-testing the options, and, not least,
communicating the trade-offs with an elegant, interactive visual model
that design skills made beautiful and usable, the organization discovered
that increasing investment along one R&D path would actually keep three
technology options open for a longer period. This bought the company
enough time to see which way the technology would evolve and avoided the
worst-case outcome of being locked into a very expensive, and very wrong,
choice. One executive likened the resulting flexibility to “the choice of
betting on a horse at the beginning of the race or, for a premium, being able
to bet on a horse halfway through the race.”
It’s not a coincidence that this happy ending concluded as the initiative
had begun: with senior management’s engagement. In our experience, the
best day-one indicator for a successful data-analytics program is not
the quality of data at hand, or even the skill-level of personnel in house, but
the commitment of company leadership. It takes a C-suite perspective to
help identify key business questions, foster collaboration across functions,
align incentives, and insist that insights be used. Advanced data analytics
is wonderful, but your organization should not be working merely to put an
advanced-analytics initiative in place. The very point, after all, is to put
analytics to work for you.
Helen Mayhew is an associate partner in McKinsey’s London office, where Tamim Saleh is a senior partner; Simon Williams is cofounder and director of QuantumBlack, a McKinsey affiliate based in London.