See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220249734

Inventory Optimization at Procter & Gamble: Achieving Real Benefits Through

User Adoption of Inventory Tools

Article  in  Interfaces · February 2011

DOI: 10.1287/inte.1100.0546 · Source: DBLP

CITATIONS READS

41 28,511

10 authors, including:

Ingrid Farasyn John J. Neale

Farasyn Analytics Consulting Boston College
7 PUBLICATIONS   257 CITATIONS    10 PUBLICATIONS   419 CITATIONS   

SEE PROFILE SEE PROFILE

Oscar Rosen William Tarlton

Procter & Gamble Procter & Gamble
8 PUBLICATIONS   351 CITATIONS    1 PUBLICATION   41 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

bullwhip View project

All content following this page was uploaded by Oscar Rosen on 26 August 2014.

The user has requested enhancement of the downloaded file.

 informs ®

Vol. 41, No. 1, January–February 2011, pp. 66–78 doi 10.1287/inte.1100.0546

issn 0092-2102
eissn 1526-551X
11
4101
0066 © 2011 INFORMS

THE FRANZ EDELMAN AWARD

Achievement in Operations Research

Inventory Optimization at Procter & Gamble:

Achieving Real Benefits Through User Adoption of
Inventory Tools
Ingrid Farasyn
Procter & Gamble Services Company NV, 1853 Strombeek-Bever, Belgium, [email protected]

Salal Humair
Harvard School of Public Health, Boston, Massachusetts 02115, [email protected]

Joel I. Kahn
PS Analytics, The Procter & Gamble Company, Cincinnati, Ohio 45202, [email protected]

John J. Neale
Boston University, Boston, Massachusetts 02215, [email protected]

Oscar Rosen
The Procter & Gamble Company, Cincinnati, Ohio 45202, [email protected]

John Ruark
Logility, Inc., Burlington, Massachusetts 01803, [email protected]

William Tarlton
The Procter & Gamble Company, Hunt Valley, Maryland 21030, [email protected]

Wim Van de Velde

Procter & Gamble Services Company NV, 1853 Strombeek-Bever, Belgium, [email protected]

Glenn Wegryn
The Procter & Gamble Company, Cincinnati, Ohio 45202, [email protected]

Sean P. Willems
Boston University, Boston, Massachusetts 02215, [email protected]

Over the past 10 years, Procter & Gamble has leveraged its cross-functional organizational structure with
operations research to reduce its inventory investment. Savings were achieved in a two-step process. First,
spreadsheet-based inventory models locally optimized each stage in the supply chain. Because these were the
first inventory tools installed, they achieved significant savings and established P&G’s scientific inventory prac-
tices. Second, P&G’s more complex supply chains implemented multiechelon inventory optimization software
to minimize inventory costs across the end-to-end supply chain. In 2009, a tightly coordinated planner-led
effort, supported by these tools, drove $1.5 billion in cash savings. Although case studies show the mathematics
employed, of equal importance is the presentation of the planning process that facilitates inventory management
and the decision tree that matches a business to the optimal inventory tool depending on the requirements of the
business. Today, more than 90 percent of P&G’s business units (about $70 billion in revenues) use either single-
stage (70 percent) or multiechelon (30 percent) inventory management tools. Plans are underway to increase the
use of multiechelon tools to manage 65 percent of P&G’s supply chains in the next three years.
Key words: industries: consumer packaged goods; inventory/production: applications, multi-item/echelon/
stage; dynamic programming: applications.

66
Farasyn et al.: Inventory Optimization at P&G
Interfaces 41(1), pp. 66–78, © 2011 INFORMS 67

P rocter & Gamble (P&G), founded in 1837, is a

leading global consumer products company with
$76.7 billion in sales in 180 countries. Four billion
comprises 145 P&G-owned manufacturing facilities
and 300 contract manufacturers, resulting in over
6,900 unique product-category market combinations
times a day, P&G touches the lives of people around served. Each supply chain requires well-coordinated
the world through its familiar brand names—Tide, efforts based on best-available information, commu-
Crest, Gillette, Pantene, and more than 200 others. nication, and planning tools. Production batch sizes,
It competes in 26 distinct product-category markets planning reaction times, order policies, replenishment
such as hair care, paper towels, cosmetics, skin care, timing, physical facility restrictions, customer order
oral care, blades and razors, diapers, and fabric care. patterns, modes of transportation, new-product intro-
Inventory management in a company of P&G’s ductions, promotions on existing brands, assortment
size, scale, and complexity requires leveraging the management, and regulatory requirements combine
right people, organization, and tools. By marrying the to present complex challenges and trade-offs.
appropriate operations research (OR) techniques with Each of P&G’s three global business units (GBUs),
a unique planning-organization structure, P&G has Beauty & Grooming ($26.3 billion), Household Care
achieved two step-change improvements in inventory ($37.3 billion), and Health & Well-Being ($14.4 bil-
levels. The first improvement came from the broad lion), would be large enough to be placed on the 2009
application of spreadsheet-based inventory models; Fortune 200 list. As such, each has an internal sup-
this work produced four tools that locally optimize port and line organization responsible for planning
different portions of the supply chain. The second that takes into account the unique constraints of the
improvement integrated multiechelon inventory soft- business.
ware in P&G’s more-complex supply chains. In 2009, To drive scale and efficiency through P&G’s sup-
these tools, implemented through a well-coordinated ply network operations, horizontal process networks
planning community, were instrumental in driving (HPNs) (see Figure 1) span across all of P&G. HPNs
$1.5 billion in cash savings, while maintaining or define, manage, and execute standard work processes
increasing service levels. across P&G’s business operating units.
This paper details how P&G implemented advanced For example, the category supply planning HPN
inventory optimization technology in more than
is responsible for developing plans that optimize the
90 percent of its businesses to reduce its inven-
supply network’s capability to deliver products as
tory investment significantly over the past 10 years.
required by the customer at lowest delivered cost.
Although case studies show the mathematics emp-
This involves balancing numerous constraints against
loyed, of equal importance is the presentation of the
objectives that can conflict (e.g., defining capacity
planning process that facilitates inventory manage-
requirements to ensure that targeted utilization rates
ment and the decision tree that matches a business
are met, while ensuring that sufficient inventories are
to the optimal inventory tool depending on business
maintained in the right quantities and at the right
requirements.
place). Planners also must ensure that manufactur-
ing resources are planned in a manner that supports
How P&G Manages Inventory strategic objectives and delivers against customer
The goal of P&G’s logistics planning effort is to requirements.
delight our consumers and shoppers with world-class HPN members include GBU representatives, infor-
products that are designed and packaged to suit their mation technology (IT) service managers, and global
needs, available wherever they shop, and priced to business and technical experts such as the supply
represent the best value for them. chain OR group. The teams are responsible for deter-
P&G’s logistics planning workforce consists of mining the scope of a globally standard solution set
over 5,000 individuals who plan material supply, consisting of work processes and tools. The teams also
capacity, inventory, and logistics for the company’s govern the use of the solutions—maintaining stan-
500 supply chains. The total supply chain network dards and determining innovation requirements for
 Farasyn et al.: Inventory Optimization at P&G
68 Interfaces 41(1), pp. 66–78, © 2011 INFORMS

GBS MDO
Global Market
GBU business development
Global business unit services organization

Beauty & Household Health & SAP ERP N America

Grooming care Well-Being IT Services L America
Beauty Fabric care Health care Analytics Europe
Grooming OR Group Asia
Home care Snacks
Baby care Pet care
Family care

Material supply planning

Site-integrated planning
Category supply planning
Demand planning
Business planning
Artwork planning
Supply network planning
Retail supply
Customer logistics
Distributor logistics
Customer service

Figure 1: P&G’s 11 HPNs enable cross-business unit best practices and scalability for P&G.

improving the current solution and supporting the correct inventory levels. In recognition of this need
company’s growth plans. for different inventory models, P&G has deployed a
These global networks provide the backbone to framework (see Figure 2) that assigns the appropri-
drive scalability and efficiency for the company and ate inventory technology to any business based on its
are a primary reason why P&G is ranked as one operating characteristics and complexity. In the next
of the best-managed supply chains in the world. paragraph, we describe the framework and its appli-
It has won numerous best-practice awards, includ- cable terms.
ing multiple consecutive top-five placements in AMR The primary selection criterion is network com-
Research’s list. Through the HPNs, planners are con- plexity—whether or not the planner is setting inven-
tinually trained to use planning tools and systems tory targets at several echelons in the supply chain.
to help meet business objectives regarding cost, cash, As we will show, the majority of supply chains in
and service management. the Beauty & Grooming GBU are in this category.
For a company with P&G’s product depth and These supply chains span multiple echelons and ben-
breadth, a one-size-fits-all inventory strategy is not efit tremendously from multiechelon inventory opti-
optimal. Capacity utilization of high-speed papermak- mization. Depending on the workflow defined by
ing operations, in which capital and logistics costs the planner, the technology employed is a third-
demand high utilization and turnover for profitabil- party software solution from Logility, Inc.—either
ity, differs dramatically from hair-care batch-making Voyager Inventory Strategy Optimization (VISO) or
operations in which many stock-keeping units (SKUs) Voyager Inventory Target Optimization (VITO). Alter-
are made and distributed through a multiple-tier dis- natively, many of the supply chains in the House-
tribution network. Different businesses require differ- hold Care GBU are planned at a single echelon.
ent modeling methods and tools to determine the A single-stage inventory tool has been customized
Farasyn et al.: Inventory Optimization at P&G
Interfaces 41(1), pp. 66–78, © 2011 INFORMS 69

One DC FIM
# of P&G DCs
Single Stage (Finished Goods)
P&G

Two-Stage Distribution XIM

Customer 1 Central + n Remote
or P&G? Locations (Finished Goods)
Finished
Goods
Customer RIM
FG or Single Stage (Retailer)
Set safety
raw/pack?
stock target
Raw/Pack MIM
Single Site
or Single Stage (Materials)
1 Plant to Many DCs
Exception Based
Review Monthly/Weekly VITO
Complexity?
Multiechelon Frequency Logility (Multiechelon)
Multiple-site
Network or
Large # of Strategic and Tactical Analysis VISO
SKUs Logility (Multiechelon)

Figure 2: P&G’s OR group has developed a decision tree for its businesses to follow to match the correct inventory
tool to the appropriate business conditions. Complex networks, defined by multiple echelons or many SKUs,
employ multiechelon inventory optimization models, whereas single-stage settings employ customized single-
stage inventory management solutions tailored to the operating environment.

to each echelon’s most important focus—materials about 50 percent in the Americas to over 80 per-
(MIM), finished goods held at a P&G plant location cent in Europe, the Middle East, and Africa (EMEA)
(FIM), or at a distribution point with several remote and Asia.
stocking locations (XIM). Custom variations of the The single-stage models were developed by P&G’s
model have been developed for retail customer needs internal OR group and implemented as end-user
(RIM) and to better predict prebuild inventory levels spreadsheets for use by planners. We refer to Farasyn
for new products. To demonstrate the breadth of these et al. (2008) for a deeper discussion of our single-stage
tools, we focus on two implementations that corre- inventory models, including a review of the design
spond to the nodes in Figure 2: single-stage inventory choices, benefits, and weaknesses of spreadsheets as
models in Western Europe and multiechelon models the technology platform for this type of application.
in North America.
Business Benefits
Statistical inventory models avoid supply chain plan-
Single-Stage Inventory Models ner bias toward excessively high levels of safety
P&G has a long history of applying inventory man- stock. Indeed, safety stock at too low a level would
agement techniques. Its use of scientific inventory quickly result in customer service incidents; the plan-
planning dates to the 1970s (Murphy 1975); how- ner would fix this issue by choosing higher safety
ever, the adoption of distribution requirements plan- stock targets. However, when a planner manually sets
ning (DRP) systems in the late 1980s triggered P&G’s a high safety stock target, no automatic correction
development of simple but robust models for set- mechanism is available to adjust the safety stock lev-
ting inventory targets across a distribution network, els downward.
one echelon at a time. P&G has successfully adopted Across all regions and business units, we see a dra-
single-stage inventory models in over 60 percent of matic inventory reduction the first time the models
its supply chains. Coverage by region varies from are used for a specific supply chain (see Figure 3). For
 Farasyn et al.: Inventory Optimization at P&G
70 Interfaces 41(1), pp. 66–78, © 2011 INFORMS

Global Total inventory

business unit Region Model scope Inventory model reduction (%)

Beauty & Grooming North America Skin-care plants FIM −50

Household Care Western Europe Amiens, France plant to XIM −12
Germany warehouse
Household Care Western Europe Mechelen, Belgium plant to XIM −30
UK warehouse
Multiple Central and Eastern Europe Regional plants and FIM/XIM −10
warehouses
Household Care Central and Eastern Europe Moscow central XIM −10
warehouse
Household Care Asia Pan-Asia Pringles FIM −12
Household Care Asia China Pringles FIM −10

Figure 3: Examples illustrate applying single-stage inventory models to various businesses at P&G.

example, the initial implementation of the extended (FY) 1999–2000 through FY2009–2010. Inventory is
single-stage model, XIM, on the Lenor fabric softener normalized on a 100-unit scale, and service is mea-
inventory produced at the plant in Amiens, France, sured as case fill rate. Implementing the initial single-
yielded a realized inventory reduction of 12 percent. stage model (FIM) in the F&HC Western Europe
Figure 3 provides additional examples of the model’s businesses resulted in a 29 percent inventory decrease
application. Although not exhaustive, it provides evi- in the first five years. In subsequent years, in which
dence of the model’s robustness across a number of the next-generation, eXtended Inventory Model (XIM)
different business categories and regions. was introduced, an additional 10 percent inventory
In addition to the quantifiable results, there are reduction was realized.
intangible benefits. The single-stage models have cre- Following the initial dramatic improvement as
ated a common language for inventory discussions a result of introducing the single-stage tool in
in the supply chain function. Planners have become FY1999–2000, service has remained stable and the
more aware of the role and influence of the various F&HC business in Western Europe has achieved its
inventory building blocks (e.g., safety, cycle, transit 99.5 percent case fill-rate objective.
stocks). Inventory concepts are now an important part F&HC Western Europe realized this improve-
of the supply chain function’s curriculum, helping ment in spite of growing complexity. Over the past
planners to manage the complex trade-offs between
capacity, inventory, and service.
100 100.0

Case Study: Fabric & Home Care, Western Europe 90 99.5

Customer Service (%)

Despite growing business complexity, our single-

Inventory (index)

80 99.0
stage models have enabled the Fabric & Home Care
(F&HC) business in Western Europe to reduce its 70 98.5
inventory and maintain high customer-service levels.
60 98.0
This business unit comprises approximately 20 manu-
facturing and customization centers. Fabric care prod- 50 Total Inventory 97.5
Customer Service
ucts include laundry granules, liquids and tablets, 40 97.0
liquid tablets, bleach, and fabric enhancers. Home 99/00 00/01 01/02 02/03 03/04 04/05 05/06 06/07 07/08 08/09 09/10
Fiscal Year
Care includes automatic and manual dish care, sur-
face cleaners, and air care products. Figure 4: By implementing single-stage inventory models in F&HC
Figure 4 tracks inventory and customer service Western Europe, total inventory decreased 36 percent; service rose
measures over the 10-year period from fiscal year above 98.5 percent.
Farasyn et al.: Inventory Optimization at P&G
Interfaces 41(1), pp. 66–78, © 2011 INFORMS 71

10 years, it has had to deal with significant prod- information is necessary. Bossert and Willems (2007)
uct proliferation—partly because of acquisitions and discuss the complexity of review periods.
partly driven by new-product introductions and inno- The costs and demand uncertainties differ signif-
vations. Although its number of finished goods has icantly at various echelons of the supply chain. On
doubled (i.e., from under 2,500 to over 5,000) during a relative basis, packaged finished goods are sig-
these 10 years, inventory levels have held steady. nificantly more expensive than semifinished goods,
which are more expensive than raw materials. Fur-
Multiechelon Inventory Optimization ther, because demand is aggregated at fewer locations
in P&G’s Beauty & Grooming GBU as it is passed up the supply chain, its relative uncer-
Of P&G’s three GBUs, Beauty & Grooming’s prod- tainty decreases. Combined, these present a tremen-
ucts have the most complex supply chains. Several dous postponement opportunity.
factors drive this complexity, all of which increase the The supply chain in Figure 5 is a relatively sim-
benefits derived from multiechelon inventory opti- ple representation; an average network modeled by
mization. In comparison to the other GBUs, it has Beauty & Grooming has 4,000 to 5,000 stages and
more finished goods SKUs, both for everyday items 6,000 to 10,000 arcs. Models of this size involve
and promotional items. It also commonly customizes approximately 500 finished goods across the supply
packaging to match a customer’s unique require- chain. Although the supply chain in Figure 5 is sig-
ments. In general, customized packaging has a lower nificantly more modest in size, it presents a complex
unit volume per item, a higher sales price per item, acyclic network. The use of common components that
and more echelons in the supply chain. Raw materi- range from bottles to chemicals to cardboard, and
als are often turned into intermediate products using multiple sources of supply to downstream stages, cre-
a batch manufacturing process; these products then ate a network without any predictable structure. Fig-
become many different finished goods, which then ure 6(a), which reduces the network from 45 nodes
receive specialized packaging and pass through mul- and 52 arcs to 26 nodes and 33 arcs, simplifies Fig-
tiple tiers of distribution, depending on the region of ure 5 to show its network complexity.
the world. Figure 5 shows an example of a Beauty & Prior to the research we present in the appendix
Grooming supply chain modeled in Logility’s VISO and document in Humair and Willems (2011), the
software. supply chain in Figure 6(a) could not be solved to
To properly set inventory targets across the supply
optimality. Existing exact solution techniques, such
chain, the supply chain must be modeled at the level
as Humair and Willems (2006), require partitioning
of complexity consistent with the production plan-
the network into disjoint bipartite subgraphs; Fig-
ning system. Therefore, it is modeled as a network in
ure 6(b) shows that this is not possible. Other tech-
which a stage represents a specific SKU at a location
niques that rely on linear programming (LP) relax-
and arcs denote precedence between the stages; this
ations require concave stage costs. Because planners
corresponds to SAP’s classification for raw materials
(ROH), semifinished products (HALB), and finished require the problem to be solved exactly, these kinds
products (FERT). The mapping must be one-to-one to of relaxations were not acceptable to the planning
implement the optimization results. community. In effect, the planners were already using
From a mathematical perspective, the need for mul- single-echelon tools; they would only adopt new tools
tiechelon inventory optimization is driven by review that were demonstrably superior.
periods, demand pooling, cost accrual, and net- To demonstrate the value of multiechelon inventory
work complexity. Although production runs each day, optimization at P&G, we present the implementation
many SKUs are reviewed and replenished weekly process that began with the North America Cosmetics
or monthly, thus introducing important dependencies unit and now spans the entire Beauty business unit,
between stages. The demand at upstream stages is a which consists of cosmetics, deodorants, hair care,
complex function of downstream demand, and prop- personal cleansing, prestige fragrances, and skin-care
erly defining how adjacent stages transmit demand products.
 Farasyn et al.: Inventory Optimization at P&G
72 Interfaces 41(1), pp. 66–78, © 2011 INFORMS

Figure 5: Real-world supply chains at P&G Beauty & Grooming span multiple echelons and have significant
component commonality as modeled in this view from Logility’s VISO software. Note: The text associated with
each stage is intentionally blurred to protect confidential information.

Leveraging North America Cosmetics’ Success The supply chain uses over 4,000 unique materials
Across the Beauty & Grooming GBU from more than 100 vendors. Despite this complexity,
The Cosmetics supply chain in North America poses Beauty & Grooming was able to leverage its inter-
a significant challenge to inventory optimization nal supply chain mastery, coupled with single-stage
because it supplies the market with over 1,300 stan- inventory tools, to reduce its cosmetics inventory by
dard SKUs and 400–500 promotional SKUs per year. half from 1999 to 2004. Improved capabilities and
Farasyn et al.: Inventory Optimization at P&G
Interfaces 41(1), pp. 66–78, © 2011 INFORMS 73

(a) 5 (b)

6 18 18
24 24
7 19 19

1 9

2 10 20 20
C1
25 25
3 11 21 21

4 12

13

14 22 22
26 26
15 23 23

16

17 17

Figure 6: Figure 6(a) simplifies Figure 5 to its underlying acyclic network. Figure 6(b) further reduces 6(a) by
representing its largest cluster as a single cluster node, C1.

responsiveness in manufacturing and distribution, aging materials, 18 intermediate subassemblies (i.e.,

real-time information and planning systems, and partially assembled finished goods), and 75 finished
end-to-end integration from suppliers to retailers all goods that move from finished packaging to US and
contributed to this success. However, continuing to Canada distribution centers, and ultimately to retail
reduce the leaner inventory became increasingly dif- customers. Intermediate subassemblies also must sat-
ficult. Although results reached a plateau in 2005, the isfy demand for promotional items.
business need to continue to free up cash and reduce The model incorporates existing service policies
the risk of obsolescence did not subside. Supply chain
(e.g., service level target of 99.5 percent case fill rate),
leaders decided to replace the single-stage inventory
material lead times (generally ranging from 7 days
optimization toolset with multiechelon optimization
to 8 weeks), production times (1 to 2 days), review
technology.
Figure 7 is a simplified Logility VISO multiech- periods (7 to 28 days), transportation and movement
elon model of one product family (liquid makeup) times (1 to 7 days), quality assurance durations (1 to
within the total North America cosmetics supply 5 days), and costs added at each location. Demand
chain. The chain consists of eight unique raw materi- characterization (mean and standard deviation) for
als (plus another 50 shared with other products that each finished-goods SKU was based on the previous
are not shown), 10 blank uncolored work-in-process 13 weeks of actual shipments and forecast, and the
(WIP) materials, 24 colored WIP materials, 150 pack- future 13 weeks of forecast.
 Farasyn et al.: Inventory Optimization at P&G
74 Interfaces 41(1), pp. 66–78, © 2011 INFORMS

Raw materials
Blank WIP

Colored WIP Packaging

materials
Intermediate
subassemblies

Promotion
Finished good demand
packing Ship to US DC

US
Ship to Canada DC customer
demand

Figure 7: In this screenshot of the multiechelon inventory model for liquid makeup, all 75 finished goods are
modeled in a network that has 500 stages and 700 arcs. A stage represents an SKU at a location; an arc captures
the bill-of-materials goes-into relationship between SKU locations.

Applying Logility’s multiechelon optimization algo- that the target 99.5 percent service level was protected.
rithms to the cosmetics liquid-makeup portion of Figure 8 provides an overview of the change in
the supply chain yielded a change in level and safety-stock days on hand across the four major
placement of inventory safety stocks, while ensuring safety-stock inventory echelons—materials, WIP, fin-
ished goods in the US distribution center, and finished
30
Previous single-stage process goods in the Canada distribution center. Safety-stock
25 New multiechelon process days decreased in materials and finished goods, and
increased in WIP.
Safety stock days

20
Most importantly, Figure 9 provides the safety
15 stocks in dollars. The total investment in safety stock
10
for this supply chain was reduced by 17 percent
because the dollar reduction in finished goods and
5
materials far outweighed the dollar increase in work
0 in process.
Materials Work in Finished goods Finished goods Application of the multiechelon technology to the
process (US distribution (Canadian
centers) distribution entire North America cosmetics supply chain yielded
centers) even better results, with the total inventory reduction
reaching 7 percent. During the implementation, the
Figure 8: Multiechelon inventory optimization shifted safety stock to the cosmetics supply chain team also identified a poten-
work-in-process echelon, reducing inventory at downstream finished
goods locations in the United States and Canada. Raw materials inventory
tial four percentage points of incremental reduction by
was also reduced. doing “what-if” analyses of current operating policies.
Farasyn et al.: Inventory Optimization at P&G
Interfaces 41(1), pp. 66–78, © 2011 INFORMS 75

350 of these changes because the supply chain echelons

Previous single-stage process
300 New multiechelon process were linked.
It is worth noting that these financial savings were
Safety stock dollars

250
achieved on top of an already strong base of inventory
200 reductions that had been achieved by implementing
150 the single-stage inventory models. These savings were
100
at a site that was a sophisticated consumer of single-
stage inventory models—not at a site new to inven-
50
tory optimization.
0 As the Logility technology was rolled out beyond
Materials Work in
process
Finished goods Finished goods
(US distribution (Canadian
P&G Cosmetics to the other Beauty & Grooming sup-
centers) distribution ply chains worldwide, a standard multiechelon work
centers)
process using the P&G standard SAP platform (Fig-
ure 11) was developed. The process ensures consis-
Figure 9: Multiechelon inventory optimization reduced safety stock cost by tent data visibility and accuracy. Data are extracted
more than 17 percent, equating to a 5 percent reduction in total inventory
cost. from the SAP system modules, passed through an
internally developed database for formatting and
manipulation, and then passed into the Logility
Implementation of the multiechelon model in North
multiechelon software, which automatically maps
America Cosmetics provided 7 percentage points of
and ultimately optimizes the model. As part of a
the 9 percent inventory investment reduction that
monthly workflow, planners review the resulting rec-
was achieved from FY 2005–2006 to 2006–2007 (see
ommended safety stocks and upload them back into
Figure 10), while maintaining its customer service
the appropriate SAP systems.
levels above target; that is, multiechelon inventory P&G Beauty & Grooming also set up a global
technology was credited with 77.8 percent of the support structure (Figure 12) for its multiechelon
inventory savings at North America Cosmetics. In technology to enable maintenance of current imple-
addition, the technology has been a key enabler of mentations, user skill development, expansion to new
subsequent annual reductions of 2–3 percent since the supply chains (e.g., acquisitions), expansion of model
initial implementation. The most significant change scope (e.g., into suppliers), and further exploitation
was reducing some review periods in finished-goods of the technology into more strategic applications
inventory. The planners could understand the impact (e.g., sensitivity and what-if analysis). Key users
at each business site own the supply chain model
NA cosmetics care supply chain results and interface with planners in their supply chain
105 99.6
to maintain, apply, and leverage their multiechelon
99.4
100 model. Regional leaders in Asia, Latin America, North
99.2
America, and Europe, who are part of P&G’s Product
Service (%)

95
99.0
Inventory

90 Supply organization, provide training and support

98.8
85 for the site key users, lead new implementations in
98.6
80
their regions, identify unique regional requirements,
98.4
Single-stage inventory Multiechelon Multiechelon closed-
and implement work-process enhancements. A global
75 98.2
modeling implementation loop planning Product Supply leader supports the regional lead-
70 98.0
02/03 03/04 04/05 05/06 06/07 07/08 08/09 ers, identifies global enhancements, maintains work-
Fiscal year process standardization across regions, and provides
overall strategic direction. These leaders receive IT
Figure 10: Total inventory in North America Cosmetics decreased support from the Supply Network Solutions (SNS)
9 percent during the multiechelon implementation process. Multiechelon
inventory optimization was credited with 78 percent of this reduction. With organization, which is within P&G Global Busi-
the technology established as a backbone of the inventory planning pro- ness Services, to maintain hardware and software
cess, sustained reductions of 2–3 percent have been achieved. capabilities.
 Farasyn et al.: Inventory Optimization at P&G
76 Interfaces 41(1), pp. 66–78, © 2011 INFORMS

Parameter,
SAP BOM data
R/3

Logility multiechelon application

Financial data

Data manipulation database

MRP

Individual planner analysis

SAP R/3 and APO/DRP

Recommended Implemented
SAP safety stocks safety stocks
APO/ Parameter data
DRP

Ship from
SAP locations
CSP
Demand data
(BW)

Figure 11: As part of the multiechelon inventory process, P&G implemented a standardized interface between its
supply chain related systems and the Logility software. This allows the automatic creation of inventory models
and targets for planner approval.

Applying multiechelon technology to the North 2010, over 85 percent of all SKUs in Beauty & Groom-
American Cosmetics supply chain, although one of ing are being modeled using the multiechelon inven-
the most complex in P&G, is only one of 31 successful tory process. Similar to the North America Cosmetics
implementations in P&G’s Beauty & Grooming (see example described above, these 30 supply chains are
Figure 13). Multiechelon models are in place in every achieving total inventory reductions in the 3–7 per-
geographic region; associated work processes are cent range and are maintaining service levels at or
executed as frequently as biweekly. As of February above target.+

Site-
Regional Global Beauty
Integrated
IT Leader IT Leader
Planners

Market Site Key Regional Global

Planners User PS Leader PS Leader

Materials
Planners

Figure 12: A global workflow structure has been implemented to support multiechelon decision making across
Beauty & Grooming.
Farasyn et al.: Inventory Optimization at P&G
Interfaces 41(1), pp. 66–78, © 2011 INFORMS 77

Average Average Average Key area Guaranteed Service Model Formulation

Number of number of number of number of of model
Region models echelons model nodes model arcs complexity In the GS framework, the supply chain is defined as
a network with node set N and arc set A. Each node
North 10 5 5
000 7
500 SKUs and corresponds to a stage in the supply chain. A stage
America materials
Asia 8 8 3
500 10
000 Distribution
is a processing activity for a given SKU and location
network in the supply chain, such as manufacturing an item,
Europe 10 6 4
000 6
000 SKUs and
materials transporting an item from one location to another,
Latin 3 6 2
000 6
500 Distribution or packaging an item. Arcs denote the precedence
America network
relationship between stages. To put this in a system
context, if one is using a system (e.g., SAP), stages
Figure 13: As of February 2010, 31 multiechelon inventory models are in
correspond to raw materials (ROHs), semifinished
production in Beauty Care; 85 percent of Beauty & Grooming’s SKUs are
modeled in the multiechelon inventory process. products (HALBs), and finished products (FERTs).
The arcs correspond to the bill-of-materials (BOM)
Conclusion routes that convert ROHs to FERTs.
In its simplest form, the GS model requires very lit-
By marrying the tools of inventory optimization with
tle data. Tj is the lead time at stage j; this is the time
the people in P&G’s horizontal planning networks,
required to complete the processing requirements of
P&G has defined an inventory management pro-
stage j, assuming that all its raw materials are avail-
cess that has significantly reduced its total inventory
able. sj is the maximum outgoing service time at
investment. The work began with single-stage inven-
stage j; this is the longest time that stage j can take
tory models designed by P&G’s OR group and imple-
to fulfill demand after it receives an order. There are
mented in spreadsheets that planners can use directly.
two decision variables at each stage: incoming service
To this day, these spreadsheet tools drive 60 percent of
time and outgoing service time. The outgoing service
P&G’s business. For more complex supply chain net-
time, Sj , is the delivery time stage j quotes its adjacent
works, multiechelon inventory models have replaced
downstream stages. The incoming service time, SIj ,
the single-stage models, producing additional average
is the longest outgoing service time from upstream
inventory reductions of 7 percent. These multiechelon
adjacent stages quoted to stage j.
models now drive 30 percent of P&G’s business.
Each stage has a cost function, cj SIj
Sj , which is a
P&G’s work on inventory optimization is not stand-
function of its incoming and outgoing service times.
ing idle. We assume that another 40 percent of
The cost function represents the cost of holding inven-
the single-stage models will migrate to multiechelon
tory at the stage in all its various forms.
models in the next three years. Beauty Care has also
The GS model can be formulated as the following
spearheaded a closed loop of inventory optimization
mathematical program P:
in which tactical inventory models are used to gen-
erate more rigorous strategic analysis that can then
N



feed the next-generation design of the supply chain. P min cj SIj
Sj 
j=1
Furthermore, P&G continues to work in related areas,
such as improving the forecast through point-of-sale- s.t. Sj − SIj ≤ Tj ∀j ∈ N
based demand sensing. SIj − Si ≥ 0 ∀ i
j ∈ A
Sj ≤ sj ∀ j
k ∈ N
j
k ∈ A
Appendix. Multiechelon Modeling
Sj
SIj ≥ 0
integral ∀j ∈ N
Framework
The guaranteed service (GS) model of inventory place- The objective function minimizes the total stage
ment forms the underpinnings of the multiechelon cost. The first constraint imposes a rational operating
inventory optimization (MEIO) tool. Graves and constraint on a stage’s outgoing service time: a stage’s
Willems (2003) summarize the specifics of the GS outgoing service time cannot exceed the longest ser-
framework and compare it to other MEIO approaches. vice time quoted to the stage plus its own lead time.
 Farasyn et al.: Inventory Optimization at P&G
78 Interfaces 41(1), pp. 66–78, © 2011 INFORMS

140 Therefore, we based our approach to constructing

Expected inventory level

Stage B Stage A Stage C

120 an algorithm on two simple observations: first, that
100 the spanning tree algorithm from Graves and Willems
(2000) could handle these complicated cost structures
(units)

80
60 elegantly with a minor reformulation; second, that the
40 solution to a well-chosen spanning tree relaxation of
20 a general network was very often (when we observed
0
it qualitatively) close to the true optimal solution. The
0 5 10 15 20 25 30 35 problem was how to move from the relaxed solution
Net replenishment lead time (days) to the network-optimal solution.
The answer was not obvious, because the span-
Figure A.1: Generalized stage cost functions occur in the presence of ning tree algorithm cannot be used to get an optimal
advanced supply chain dynamics (e.g., review periods).
solution by bounding a single variable, as Humair
and Willems (2011) discuss. By observing the exis-
The second constraint defines the incoming service tence of another relationship between the inbound
time at a stage to be at least as large as the longest out- service times and network partitioning, we were able
going service time quoted to the stage. The third con- to construct an algorithm that could move efficiently
straint enforces an upper bound on a demand stage’s from the relaxed solutions to the network-optimal
outgoing service time; the final constraints require ser- solutions. The efficiency of the resulting algorithm
vice times to be nonnegative and integer. was surprising, and its attractive property was its
More complex forms of the GS model capture provable optimality. Humair and Willems (2011) pro-
advanced supply chain dynamics, such as nonnested vide details.
review periods (Bossert and Willems 2007) or non-
stationary demand (Neale and Willems 2009). These
advanced dynamics require more data to fully specify References
the model; however, the general structure is similar Bossert, J. M., S. P. Willems. 2007. A periodic-review modeling
approach for guaranteed service supply chains. Interfaces 37(5)
in form to P. 420–435.
Farasyn, I., K. Perkoz, W. Van de Velde. 2008. Spreadsheet models
Novelty of Solution Approach for inventory target setting at Procter & Gamble. Interfaces 38(4)
If the model incorporates dynamics such as review 241–250.
periods or batching, the stage cost function is neither Graves, S. C., S. P. Willems. 2000. Optimizing strategic safety stock
placement in supply chains. Manufacturing Service Oper. Man-
concave nor monotone. Figure A.1 illustrates why agement 2(1) 68–83.
standard GS optimization approaches cannot be used Graves, S. C., S. P. Willems. 2003. Supply chain design: Safety stock
for generalized cost functions, because they assume placement and supply chain configuration. A. G. de Kok, S. C.
Graves, eds. Supply Chain Management: Design, Coordination and
either an extreme-point property of the optimal solu- Operation. Handbooks in Operations Research and Management Sci-
tion or piecewise linearity of the cost function. ence, Vol. 11. North-Holland Publishing Company, Amsterdam,
This figure also demonstrates why standard noncon- 95–132.
vex optimization approaches fail—because of their Humair, S., S. P. Willems. 2006. Optimizing strategic safety stock
placement in supply chains with clusters of commonality. Oper.
assumptions about the structure of the objective func- Res. 54(4) 725–742.
tion (e.g., submodularity or quasi concavity). Search Humair, S., S. P. Willems. 2011. Optimizing strategic safety stock
algorithms also typically either require an objective placement in general acyclic networks. Oper. Res. Forthcoming.
Murphy, R. A. 1975. Inventory control with gamma demand and
function structure (e.g., conjugate gradient methods)
gamma lead times. Internat. J. Systems Sci. 6(1) 81–85.
or use an unstructured strategy (e.g., genetic algo- Neale, J. J., S. P. Willems. 2009. Managing inventory in supply
rithms) and are usually less efficient. chains with nonstationary demand. Interfaces 39(5) 388–399.

View publication stats