"The Power of Analytics" - Peter

Excerpts from "The Everything Store: Jeff Bezos and the age of Amazon" by Brad Stone

Chapter 6 - Chaos Theory

Immediately upon moving to Seattle, Wilke set about filling the ranks of Amazon’s logistics division with scientists and engineers rather than retail-distribution veterans. He wrote down a list of the ten smartest people he knew and hired them all, including Russell Allgor, a supply-chain engineer at Bayer AG. Wilke had attended Princeton with Allgor and had cribbed from his engineering problem sets. Allgor and his supply-chain algorithms team would become Amazon’s secret weapon, devising mathematical answers to questions such as where and when to stock particular products within Amazon’s distribution network and how to most efficiently combine various items in a customer’s order in a single box.

Wilke recognized that Amazon had a unique problem in its distribution arm: it was extremely difficult for the company to plan ahead from one shipment to the next. The company didn’t store and ship a predictable number or type of orders. A customer might order one book, a DVD, some tools—perhaps gift-wrapped, perhaps not—and that exact combination might never again be repeated. There were an infinite number of permutations. “We were essentially assembling and fulfilling customer orders. The factory physics were a lot closer to manufacturing and assembly than they were to retail,” Wilke says. So in one of his first moves, Wilke renamed Amazon’s shipping facilities to more accurately represent what was happening there. They were no longer to be called warehouses (the original name) or distribution centers (Jimmy Wright’s name); forever after, they would be known as fulfillment centers, or FCs.

Before Wilke joined Amazon, the general managers of the fulfillment centers often improvised their strategies, talking on the telephone each morning and gauging which facility was fully operational or had excess capacity, then passing off orders to one another based on those snap judgments. Wilke’s algorithms seamlessly matched demand to the correct FC, leveling out backlogs and obviating the need for the morning phone call. He then applied the process-driven doctrine of Six Sigma that he’d learned at AlliedSignal and mixed it with Toyota’s lean manufacturing philosophy, which requires a company to rationalize every expense in terms of the value it creates for customers and allows workers (now called associates) to pull a red cord and stop all production on the floor if they find a defect (the manufacturing term for the system is andon).

*****

A year into Jeff Wilke’s tenure at Amazon, he called a former teacher of his, Stephen Graves, a professor of management science at MIT, and asked for help. Amazon operated an e-commerce distribution network of unrivaled scale but the company was still struggling to run it efficiently. Its seven fulfillment centers around the world were expensive, their output inconsistent. Bezos wanted the Amazon website to be able to tell customers precisely when their packages would be delivered. For example, a college student ordering a crucial book for a final exam should know that the book would be delivered the following Monday. But the fulfillment centers were not yet reliable enough to make that kind of specific prediction.

Wilke asked Graves if he might meet with Wilke and his colleagues later that month to take a fresh look at their problems. Bezos and Wilke were asking themselves a fundamental question that seems surprising today: Should Amazon even be in the business of storing and distributing its products? The alternative was to shift to the model used by rivals like Buy.com, which took orders online but had products drop-shipped from manufacturers and distributors like Ingram.

That St. Patrick’s Day, some of Amazon’s biggest brains descended on a drab meeting room at the Fernley, Nevada, fulfillment center. Jeff Bezos and Brewster Kahle, a supercomputer engineer and founder of Alexa Internet, a data-mining company Amazon had acquired, made the two-hour flight from Seattle on Bezos’s newly purchased private plane, a Dassault Falcon 900 EX. Stephen Graves flew from Massachusetts to Reno and then drove the dreary thirty-four miles through the desert to Fernley. A few other Amazon engineers were there, as was the facility’s senior manager at the time, Bert Wegner. In the morning, the group toured the fulfillment center and listened to a presentation by one of the company’s primary contractors, who listed the benefits of additional equipment and software that he could sell them, reflecting the same traditional thinking about distribution that wasn’t working in the first place. They then dismissed the surprised contractor for the day and spent the afternoon filling up whiteboards and tackling the question of how everything at the FC might be improved. For lunch, they brought in McDonald’s and snacked from the building’s vending machines.

For Wegner, the questions being asked that day carried personal resonance. “We had a key decision to make,” he says. “Was distribution a commodity or was it a core competency? If it’s a commodity, why invest in it? And when we grow, do we continue to do it on our own or do we outsource it?” If Amazon chose to outsource it, Wegner might be out of a job. “I basically saw my own career flash before my eyes,” he says.

Amazon’s problem boiled down to something called, in the esoteric lexicon of manufacturing, batches. The equipment in Amazon’s FCs had originally been acquired by Jimmy Wright, and, like the system in Walmart’s distribution centers, was designed by its manufacturers to operate in waves—moving from minimum capacity to maximum and then back again. At the start of a wave, a group of workers called pickers fanned out across the stacks of products, each in his or her own zone, to retrieve the items ordered by customers. At the time, Amazon used the common pick-to-light system. Various lights on the aisles and on individual shelves guided pickers to the right products, which they would then deposit into their totes—a cart of the picks from that wave. They then delivered their totes to conveyor belts that fed into the giant sorting machines, which rearranged products into customer orders and sent them off on a new set of conveyor belts to be packed and shipped.

The software required pickers to work individually, but, naturally, some took longer than others, which led to problems. For example, if ninety-nine pickers completed their batches within forty-five minutes but the one hundredth picker took an additional half an hour, those ninety-nine pickers had to sit idly and wait. Only when that final tote cleared the chute did the system come fully alive again, with a thunderous roar that rolled through the fulfillment center and indicated that it was again ready to start operating at peak capacity.

Everything in the fulfillment center happened in this episodic manner. For a company trying to maximize its capacity during the big push each holiday season, that was a huge problem. Wilke subscribed to the principles laid out in a seminal book about constraints in manufacturing, Eliyahu M. Goldratt’s The Goal, published in 1984. The book, cloaked in the guise of an entertaining novel, instructs manufacturers to focus on maximizing the efficiency of their biggest bottlenecks. For Amazon, that was the Crisplant sorting machines, where the products all ended up, but picking in batches limited how fast the sorters could be fed. As a result, the machines were operating at full capacity only during the brief few minutes at the peak of the batch. Wilke’s group had experimented with trying to run overlapping waves, but that tended to overload the Crisplant sorters and, in the dramatic terminology of the general managers, “blow up the building.” It would take hours to clean up that mess and get everything back on track.

In the meeting that day at Fernley, the executives and engineers questioned the prevailing orthodoxies of retail distribution. In the late afternoon, everyone headed back onto the facility floor and watched orders move haltingly through the facility. “I didn’t know Jeff Bezos but I just remember being blown away by the fact that he was there with his sleeves rolled up, climbing around the conveyors with all of us,” says Stephen Graves, the MIT professor. “We were thinking critically and throwing around some crazy ideas of how we can do this better.”

At the end of the day, Bezos, Wilke, and their colleagues reached a conclusion: the equipment and software from third-party vendors simply wasn’t designed for the task at hand. To escape from batches and move toward a continuous and predictable flow of orders through the facility, Amazon would have to rewrite all the software code. Instead of exiting the business of distribution, they had to reinvest in it.

Over the next few years, “one by one, we unplugged our vendors’ modems and we watched as their jaws hit the floor,” says Wegner. “They couldn’t believe we were engineering our own solutions.” When Amazon later opened small facilities in places like Seattle and Las Vegas to handle easily packable items and larger fulfillment centers in Indianapolis, Phoenix, and elsewhere, it would go even further, dispensing with the pick-to-light systems and big Crisplant sorting machines altogether and instead employing a less automated approach that favored invisible algorithms. Employees would bring their totes from the shelves right to the packing stations, their movements carefully coordinated by software. Slowly, Amazon would vanquish wave-based picking, elicit more productivity from its workers, and improve the accuracy and reliability of its fulfillment centers.

Wilke’s gradual success in making the logistics network more efficient would offer Amazon innumerable advantages in the years ahead. Tightly controlling distribution allowed the company to make specific promises to customers on when they could expect their purchases to arrive. Amazon’s operating all of its own technology, from the supply chain to the website, allowed Russell Allgor and his engineers to create algorithms that modeled countless scenarios for each order so systems could pick the one that would yield the quickest and cheapest delivery. Millions of those decisions could be made every hour, helping Amazon reduce its costs—and thus lower prices and increase volume of sales. The challenge was getting good enough to do this well.

“No matter how hard it is, the consolidation of products within fulfillment centers pays for the inventory and for pieces of the overhead,” says Jeff Wilke, who claims that he never worried that Bezos would abandon the FC model at the Fernley meeting. “The principles and math were on our side, and I realized early on that this was a company where you can carry the day when you have the principles and math on your side, and you are patient and tenacious.”

*****

As the hard-lines teams were bringing Amazon into new categories, with varying degrees of success, Jeff Wilke and his group had nearly completed their job morphing Amazon’s fulfillment process from a network of haphazardly constructed facilities into something that could more accurately be considered a system of polynomial equations. A customer might place an order for a half a dozen products, and the company’s software would quickly examine factors like the address of the customer, the location of the merchandise in the FCs, and the cutoff times for shipping at the various facilities around the country. Then it would take all those variables and calculate both the fastest and the least expensive way to ship the items.

The complete software rewrite of the logistics network was having its desired effect. Cost per unit (the overall expense of fulfilling the order of a particular item) fell, while ship times (how quickly merchandise ordered on the website was loaded onto a truck) shortened. A year after the Fernley meeting, the click-to-ship time for most items in the company’s FCs was as minimal as four hours, down from the three days it had taken when Wilke first started at the company. The standard for the rest of the e-commerce industry at the time was twelve hours.