Supply Chain Management

How Machine Learning Will Transform

Supply Chain Management
by Narendra Agrawal, Morris A. Cohen, Rohan Deshpande, and Vinayak
Deshpande
From the Magazine (March–April 2024)

Didier Engels

Summary. Businesses need better planning to make their supply chains more
agile and resilient. After explaining the shortcomings of traditional planning
systems, the authors describe their new approach, optimal machine learning
(OML), which has proved effective in a... more

The Covid-19 pandemic, the Russia-Ukraine conflict, trade wars,

and other events in recent years have disrupted supply chains and
highlighted the critical need for businesses to improve planning
in order to be more agile and resilient. Yet companies struggle
 with this challenge. One major cause is
flawed forecasting, which results in
delivery delays, inventory levels that are
woefully out of sync with demand, and
disappointing financial performance.
Those consequences are hardly surprising.
After all, how can inventory and production decisions be made
effectively when demand forecasts are widely off?

00:00 / 30:49
Listen to this article
To hear more, download the Noa app

We have developed a way to address this deficiency. Our new

paradigm uses machine learning and historical data to generate
superior recommendations for supply chain decisions. While
current machine-learning methods focus on trying to create
more-accurate forecasts, ours focuses on making actual decisions.
This new methodology, which we call optimal machine learning
(OML), involves using artificial intelligence technology to create a
mathematical model that takes key data inputs related to the
supply chain (the nodes of the network, their locations, sales and
shipment transactions, financial parameters, marketing
promotions, logistical and capacity constraints, and so on) and
links them to planning decisions (what quantities to produce, for
example, or what levels of inventory to stock at each location).
This model can take into account a company’s priorities (such as
the level of customer service it is contractually obligated or
wishes to achieve), its budget restrictions, and other resource
constraints (such as the availability of materials and labor). The
data is stored in a way that enables updating in near real time and
quick revision of the calculations that inform decision-making.
We developed OML after decades of researching supply chain
management and implementing our resulting ideas at companies
in a range of industries: semiconductor equipment
manufacturing, aerospace and defense, telecommunications, and
computing. In this article we explain why existing supply-chain-
planning methods, including other machine-learning techniques,
have failed, how our approach works, and what changes in
planning teams and metrics are needed to implement it. We
highlight our experiences putting OML to the test at two Fortune
500 companies where we served as consultants—and where the
results demonstrate the potential to increase revenue and product
availability with significantly lower investments in inventory.

One of the companies makes and sells capital-intensive

equipment for manufacturing semiconductors. To ensure that its
customers can keep the equipment up and running, the firm must
manage an inventory of thousands of spare parts. In the past,
though, it often faced shortages of key components, which led to
expensive expedited shipments or disruptive wait times that
caused costly outages at customer locations. Frustrated managers
frequently scrambled to override the inventory recommendations
offered by their enterprise resource planning (ERP) system. The
managers relied on their own experience and used data inputs not
explicitly incorporated in the existing planning software, but their
overrides were subjective, ad hoc, and time-consuming to devise,
often compounding their problems.

The other company we’ll discuss is a consumer electronics firm

that sells, through thousands of retailers’ stores, a portfolio of
advanced virtual-reality interface devices produced by contract
manufacturers in Asia. This firm was building up its process for
sales and operations planning and was challenged by the wide
range of approaches and data inputs from various stakeholders—
suppliers; teams in the company’s marketing, production,
logistics, and supply sourcing functions; and customers who
operated complex supply chains with thousands of retail store
locations. High uncertainty in consumer demand led to constant
shortages at some retail locations and excess inventory at others.

Why Agility and Resilience Remain Elusive

Companies across industries have failed to develop effective
strategies for preparing their supply chains to withstand
unforeseen disruptions. That’s because of three significant
shortcomings in existing planning methods: flawed, forecast-
driven processes; data-related challenges; and ineffective
scenario planning.
Forecast-driven planning. The most widely used approach for
supply chain planning consists of two steps. In the predict step,
demand forecasts are generated from historical data about sales
and orders, information about economic conditions and
competitors’ actions, and subjective judgments. In the optimize
step, those forecasts are fed into mathematical models of the
supply chain network in order to generate final stocking
decisions. This approach, known as predict-then-optimize (PTO),
fails for various reasons.

1. There is no single forecast that all parties agree to use for

decision-making. At the consumer electronics company, sales
managers developed monthly forecasts of aggregate sales, along
with forecasts of sales to each retail customer for the upcoming
12-month period. The inventory planning team independently
created its own forecast of customer demand, but it did so weekly
because inventory orders were typically placed once a week. To
manage orders with the consumer electronics company and
allocate inventory to its network of stores, each retail customer
also developed its own forecast. These customer forecasts often
differed significantly from those of the consumer electronics
company. Moreover, all the forecasts were made at different
points in time and were updated as new information about
demand emerged. And the estimates of end-customer demand
were often distorted because no one fully understood how
marketing promotions implemented by the consumer electronics
company or by retailers would affect sales. In situations like this,
when multiple forecasts are available, each with its own errors, it
is not clear which is the ideal one to use to optimize supply-chain-
planning decisions.

2. The objectives of the various stakeholder groups in the planning

process are not aligned, which leads to biased and suboptimal
decisions. At the consumer electronics company, the sales team
typically tended to overestimate forecasts to ensure supply, while
inventory teams tended to deflate forecasts to ensure minimal
obsolescence. The finance team cared primarily about
obsolescence costs associated with unsold inventory. Senior
management, of course, cared most about meeting quarterly
revenue targets and having minimal capital tied up in inventory.
The company’s suppliers, for their part, faced production-
capacity constraints and the competing demands of multiple
customers; as a result, the company was often at the mercy of its
suppliers’ decisions. Meanwhile, to guard against the danger of
receiving insufficient quantities of needed electronics products,
retailers often inflated their orders.
Belgian photographer Didier Engels photographs docks, cargo hulls, and shipping containers throughout Europe.
For these images he took aerial shots of colorful containers and manipulated them to create striking compositions.

Misaligned aims were an issue for the semiconductor equipment

maker too. Inventory managers often kept too few parts in stock,
hoping to minimize the overall investment in inventory and avoid
costly write-offs of unsold items. That was because senior
management provided targets for invested capital and write-offs
without a clear understanding of the resulting potential for lost
revenue.
3. The methods for deciding how to optimize inventories are flawed.
The conventional models widely used today don’t capture all the
intricacies of supply chain dynamics. They generally incorporate
overly simplistic algorithms for the optimization step in PTO. And
even if advanced “optimizer” tools are used, any errors in
forecasting, model formulation, and optimization will compound
and lead to poor results.

An additional challenge in planning is that even if perfect

forecasts were available, they would not suffice. Decisions about
supply orders and shipments to various locations must also
consider a variety of factors that will affect key financial and
operational performance indicators. Those factors include
constraints on resources (such as production, shipping, and
storage capacity), customer-specific needs, differences in
profitability by customers and regions, and both local and global
service-level requirements.

Data challenges. It is now feasible to maintain and access vast

amounts of data about store-level sales transactions in almost real
time. Unfortunately, for companies with global supply chains,
accessing and consolidating that data remains a mega challenge.

Consider the consumer electronics company. Its supply chain

data was widely dispersed among various actors both within and
outside the organization. Internally, sales teams maintained
forecast information, marketing people handled promotions-
related information, the supply planning teams managed
inventory information, the finance department had responsibility
for financial information, and so on. Externally, data on store
sales and on shipments from customers’ distribution centers to
stores was under the purview of those retailers. Making matters
worse, the data also resided in a variety of formats and
repositories—partly in Excel files, partly in Tableau tables, partly
in text form in emails, and so on. In some instances, data was
maintained at an aggregate level rather than with the specificity
needed for decision-making. Because of the scattered nature of
the information, there was simply no mechanism for leveraging
granular transaction data to support the analysis needed to drive
efficient replenishment decisions. Moreover, it was impossible to
analyze the impact of various external factors (such as business
cycles) and competitive factors (such as the introduction and
pricing of new products from rival firms) on the end-to-end
supply chain because no single representation of the entire chain
existed.
Ineffective scenario planning. A first step in developing strategies
for making a supply chain more agile (able to respond to
disruptions) and resilient (capable of bouncing back) is
articulating future scenarios corresponding to the many risks that
can affect supply chains. Black-swan events—rare, high-impact
disruptions, such as the blockage of the Suez Canal by a container
ship in 2021—are virtually impossible to predict. Other threats
that can have a major effect on supply chains—such as the Russia-
Ukraine war—can be foreseen, but the likelihood of their
occurring can be difficult to ascertain.

Although many companies have started incorporating scenario

planning into their supply-chain-planning processes, such
analysis often lacks sufficient detail to be useful. For instance, it is
not enough to know that a company’s overall sales can drop in the
event of a war that constrains a key supplier’s operations. What is
needed is a deep understanding of the magnitude and timing of
the impact on each product, customer, and region so that
appropriate strategies for ameliorating it can be developed. Such
understanding is also important for creating bespoke supply
chains—those differentiated by geography, products, and
customers—rather than using a one-size-fits-all strategy.
Unfortunately, absent a mechanism that can offer granular
insights, planning exercises often lead to subpar outcomes.
A New Paradigm
Our approach, optimal machine learning, overcomes the
significant shortcomings in existing supply-chain-planning
methods. It has three key components: a decision-support engine,
a digital twin, and an end-to-end data architecture.
Decision-support engine. OML replaces the forecast-based, two-
step planning process with a one-step process that connects input
data directly to supply chain decisions. The OML methodology
incorporates historical data about drivers of demand throughout
the supply chain (for example, actual sales transactions and
factors that influence them, such as prices, promotions, and the
size and composition of the customer base) and drivers of supply
(such as supplier inventories, lead times, capacity constraints,
and transportation delays). The engine determines the
relationship between these supply-and-demand data elements
and supply chain decisions about such things as production
quantities, replenishment orders, inventory stocking levels, and
shipments of inventory between locations in a way that optimizes
key performance indicators (KPIs). KPIs might include metrics
related to overall profit; revenue or profit by region, customer, or
product category; product availability; inventory turns; time
needed to recover from a disruption; and so on. Calculations can
be recomputed as soon as new data is available or if any changes
are made to the supplier network.
Didier Engels
Our implementations of OML at both the semiconductor
equipment maker and the consumer electronics company
leveraged the power of cloud computing to run the decision-
support engine. Both produced decision recommendations in
minutes, despite the large size of the respective supply-chain
networks. Our approach also can incorporate business constraints
specified by managers (for instance, “My budget is X” or “I can
produce no more than Y units” or “The availability of my product
has to be greater than Z%”). A lot of traditional machine-learning
approaches ignore such limitations. As a result, the solution has
to be “fixed” after the fact, which is not ideal. OML does not face
that problem.
Digital twin. A key requirement for the OML decision-support
engine to work is a detailed digital representation of the entire
supply-chain network, all material flows, and the decision-
making processes of all involved parties. By modeling the
performance of the supply chain in a highly detailed manner at
the granular level of each transaction, the digital twin allows
supply chain planners to quantify accurate KPIs for both past and
future demand-and-supply scenarios.

Such a digital representation has two advantages. First, its

calculations of the impact of historical decisions can be compared
with the actual results computed by existing business systems,
which validates its ability to measure performance accurately and
thereby gives management confidence in the system. Second, it
can be used to test the effects of alternative supply-chain
scenarios or strategies on KPIs. For example, what if the shipment
lead times double because of slowdowns in a specific shipping
lane? Or how would the disruption of a distribution center affect
revenue? Or what would be the best way to reroute shipments to
stores served by that distribution center? Together, these two
capabilities—historical analysis and predictive power—make it
possible for managers to accurately evaluate risk-mitigation
strategies, such as alternative sources of supply and
transportation, alternative stocking locations, revised production
schedules, and the use of product substitutes. As a result,
managers can make better decisions.
End-to-end data architecture. The OML decision-support engine
and the digital twin require a data storage system that works in
conjunction with all existing database-management systems
throughout the supply chain (those for the company’s operations
and those of suppliers, distributors, and customers). The storage
system should be able to pool data across teams, locations, and
products and make it possible to update and access that
information in near real time.

The architecture specifies the various data elements to be

included, their storage format and organization, linkages between
them, and the frequency with which they will be updated. The
choice of data to be incorporated should be driven by the
analytical representation of the supply chain in the digital twin
and should consider input from the managers who best
understand the potential drivers of decisions. In other words, the
OML approach entails the collection and storage of data that
reflects the supply chain structure. We recommend using a graph
database that houses the data at nodes in the supply chain (for
example, retail stores or wholesale locations) and captures
important relationships (for example, which wholesale location
supplies which retail stores). This allows meaningful visualization
of data and metrics by location, customer, product, or time. It also
enables supply chain planners to measure performance for
multiple metrics related to cost and service and to identify
instances when the observed values fall outside the expected
range and require further analysis.

It is important for senior managers to

ensure that all parties trust the
recommendations that come out of
 the planning system and don’t feel the
need to review them.

This approach is in stark contrast to one that uses all available

data, regardless of its business relevance. Some companies have
chosen to “throw everything in” while developing machine
learning models in the belief that the models will figure out what
data matters most and then weight it appropriately. The problem
with that approach is that the model is perceived as a black box,
and when decision-makers can’t understand why it produced the
results it did, they often don’t trust it.

Driving Results
At the semiconductor equipment company, the OML
methodology was used to determine specific inventory policies
that could lead to higher service levels (the fill rate) at a lower
cost. (In any supply chain, higher levels of product availability
require greater investments in inventory, which increases costs.)
Prior to the deployment of OML, the company’s legacy planning
system could maintain a fill rate of about 77% with inventory
investments of a little over $135 million. When managers used
their expertise to override the system’s recommendations, they
were able to increase the fill rate to about 81% with a slightly
higher inventory investment. The OML system gave the company
the choice of achieving this higher fill rate while spending nearly
$20 million less on inventory, or increasing the fill rate to nearly
85% while keeping the inventory investment at about $135
million. Moreover, the system freed up managers to focus on
more-strategic issues. The visualization capabilities of the digital
twin helped them understand exactly what needed to be changed
and why, which increased their willingness to accept the
recommendations. For instance, a key insight was that OML’s
decision rule considered the number of existing product
installations along with new ones and captured their effects on
inventory deployment decisions—linkages that had not been
factored in previously.

In the case of the consumer electronics company, the OML model

revealed glaring deficiencies in how inventory was being
managed. For example, the distribution center that served the
region with the highest sales volume had historically stocked the
least amount of inventory. That led to frequent shortages at stores
it supported, as well as expedited shipments to the center. The
OML model correctly identified this discrepancy and
recommended the optimal inventory-stocking policy for that
location. Our analysis also revealed that while the average level of
product availability at most retail stores was acceptable, a
significant number of stores experienced severe shortages. It is
important to note that while the OML model captured the entire
network down to retail locations, all decisions about inventory
deployment at customer distribution centers and retail stores
were made by the customers, not the consumer electronics
company. However, thanks to the model’s insights into inventory
deployment, the company was in a stronger position to influence
its customers’ inventory decisions. The key enabler in the
conversation was an easy-to-interpret visualization of the model’s
decision recommendations that showed exactly which drivers
were responsible for them. Such visualization is critical to
securing the buy-in of all stakeholders.

Succeeding with the New Approach

For an analytics-enabled tool like OML to be helpful to companies
in building agility and resilience in their supply chains, it must be
accompanied by an appropriate organizational structure,
personnel with the right skills, changes in the planning process,
and a detailed understanding of the potential and pitfalls of
machine learning.
First, senior executives should ensure that the planning team,
from the outset, includes a broad range of internal and external
stakeholders. Defining the OML objective, capturing business
constraints, and identifying KPIs and relevant data will require
functional experts and operations research analysts with
sufficient experience in modeling supply-chain-management
problems. This means that the team should include people from
marketing, sales, finance, supply chain logistics, production, and
IT; data scientists and analytics experts are also crucial. If
necessary, the internal expertise of an organization should be
supplemented by outside consultants and academic experts.
Some companies delegate machine learning projects solely to
data scientists. That is a mistake, because data scientists usually
have minimal familiarity with the supply chain domain.

Second, OML requires companies to identify, assemble, and

access the required data inputs from multiple sources and to
verify the outputs generated by the machine learning model. The
former requires computer scientists who can construct the
necessary end-to-end data architecture using modern database
systems, while the latter requires supply chain planners and
managers with sufficient domain knowledge and experience.

Third, the sales and operations planning (S&OP) process, in which

members of the planning team come together to develop sales,
production, and inventory plans, must be redesigned to leverage
the agility that OML can offer through its speed and depth of
analysis. It is common for the S&OP cycle to be one month long
because it takes that long to process information from the
previous cycle and reach consensus on actions for the next one.
But companies then struggle to respond promptly to disruptions
in supply or logistics and to the shifts in demand that occur
constantly. What they need to do is replace the typical monthlong
S&OP cycles with faster and more-responsive ones. That will
require investments in systems and processes that allow supply
chains to react with speed, such as cloud-based state-of-the-art
solver software systems that can generate solutions within a
reasonable time frame (minutes, say, for a particular scenario or
policy analysis). Companies thus need expertise in optimization
and cloud computing, which they can get from software and
cloud solutions firms, respectively.

Subscribe to our Monthly Newsletter

Analytic Services
A summary of the most recent commercial surveys and
research conducted with the HBR Advisory Council and
senior executives worldwide.
Sign Up

Finally, the planning team, in consultation with senior

management, should establish the KPIs that will ultimately drive
decisions within the supply chain. These might include fill rates
by customer, product, and region; costs and profit margins; and
inventory turns. Understanding the conflicts between internal
and external stakeholders and securing everyone’s buy-in is
important in tackling this step. By providing a detailed analysis of
the implications of various decisions for each stakeholder, OML
can help. Typically, decisions affecting overall supply-chain
performance are made by multiple stakeholders. It would be best
if the performance metrics for each stakeholder were transparent
and accepted by all parties. Our OML approach allows the model
to be run iteratively until it finds a solution that is mutually
agreeable. Ultimately, it is important for senior managers to
ensure that all parties trust the recommendations that come out
of the planning system and don’t feel the need to review them.

...
OML allows companies to base decisions on historical and current
supply-and-demand information rather than just more-accurate
forecasts. It gives them a tool that can help them reduce costs and
increase revenues, profits, and customer satisfaction. It enables
them to test strategies for mitigating risks, making it easier to
choose the best ones. By doing all these things, it offers a way to
build more-flexible, more-resilient, higher-performing supply
chains.
ABusiness
version Review.
of this article appeared in the March–April 2024 issue of Harvard

Narendra Agrawal is the Benjamin and Mae

Swig Professor of Information Systems and
Analytics at Santa Clara University’s Leavey
School of Business.

Morris A. Cohen is the Panasonic Professor

Emeritus of Manufacturing & Logistics at the
University of Pennsylvania’s Wharton School.
He is also the founder of AD3 Analytics, a start-
up that developed the OML methodology for
supply chain management.

Rohan Deshpande is a machine learning

scientist at Cerebras Systems and a former chief
technology officer at AD3 Analytics.

Vinayak Deshpande is the Mann Family

Distinguished Professor of Operations at the
University of North Carolina’s Kenan-Flagler
Business School.