2022, Cui

This article was downloaded by: [218.35.205.
47] On: 25 April 2023, At: 18:11

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA
Manufacturing & Service Operations Management

Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org
AI and Procurement
Ruomeng Cui, Meng Li, Shichen Zhang
To cite this article:

Ruomeng Cui, Meng Li, Shichen Zhang (2022) AI and Procurement. Manufacturing & Service Operations Management
24(2):691-706. https://doi.org/10.1287/msom.2021.0989
Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.
Copyright © 2021, INFORMS
Please scroll down for article—it is on subsequent pages
With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
MANUFACTURING & SERVICE OPERATIONS MANAGEMENT
Vol. 24, No. 2, March–April 2022, pp. 691–706
http://pubsonline.informs.org/journal/msom ISSN 1523-4614 (print), ISSN 1526-5498 (online)
AI and Procurement
Ruomeng Cui,a Meng Li,b Shichen Zhangc,*
a
Goizueta Business School, Emory University, Atlanta, Georgia 30322; b Bauer College of Business, University of Houston, Houston, Texas
Downloaded from informs.org by [218.35.205.47] on 25 April 2023, at 18:11 . For personal use only, all rights reserved.
77204; c Business School, Nankai University, Tianjin 300071, China

*Corresponding author
Contact: ruomeng.cui@emory.edu, https://orcid.org/0000-0003-0939-8759 (RC); mli41@uh.edu, https://orcid.org/0000-0002-8476-2189
(ML); sczhang@tju.edu.cn, https://orcid.org/0000-0003-3688-3717 (SZ)
Received: April 8, 2020 Abstract. Problem definition: In this research, we study how buyers’ use of artificial intelli-
Revised: September 17, 2020; November gence (AI) affects suppliers’ price quoting strategies. Specifically, we study the impact of
30, 2020; December 7, 2020 automation—that is, the buyer uses a chatbot to automatically inquire about prices instead
Accepted: December 16, 2020 of asking in person—and the impact of smartness—that is, the buyer signals the use of a
Published Online in Articles in Advance: smart AI algorithm in selecting the supplier. Academic/practical relevance: In a world ad-
July 20, 2021
vancing toward AI, we explore how AI creates and delivers value in procurement. AI has
https://doi.org/10.1287/msom.2021.0989 two unique abilities: automation and smartness, which are associated with physical ma-
chines or software that enable us to operate more efficiently and effectively. Methodology:
Copyright: © 2021 INFORMS We collaborate with a trading company to run a field experiment on an online platform in
which we compare suppliers’ wholesale price quotes across female, male, and chatbot buy-
er types under AI and no recommendation conditions. Results: We find that, when not
equipped with a smart control, there is price discrimination against chatbot buyers who re-
ceive a higher wholesale price quote than human buyers. In fact, without smartness, auto-
mation alone receives the highest quoted wholesale price. However, signaling the use of a
smart recommendation system can effectively reduce suppliers’ price quote for chatbot
buyers. We also show that AI delivers the most value when buyers adopt automation and
smartness simultaneously in procurement. Managerial implications: Our results imply
that automation is not very valuable when implemented without smartness, which in turn
suggests that building smartness is necessary before considering high levels of autonomy.
Our study unlocks the optimal steps that buyers could adopt to develop AI in procurement
processes.
Funding: R. Cui gratefully acknowledges the financial support provided by a research grant from Goi-
zueta Business School of Emory University.
Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2021.0989.
Keywords: artificial intelligence • procurement • wholesale pricing • automation • smartness
Forty-five percent of chief procurement officers are can also facilitate smarter control rules by continuously
using, piloting, or planning to use AI. learning, reasoning, deciding, and acting to drive a
(Deloitte 2018, p. 32; HICX Solutions 2018, p. 4) business outcome. As AI enables companies to reach
unprecedented levels of scale, scope, and learning
speed, organizations around the world are eager to
1. Introduction participate in this AI transformation. However, the rise
Artificial intelligence (AI) is related to making ma- of AI is posing new challenges for organizations to un-
chines or software mimic human behavior and intelli- derstand how it works, when it is the most powerful,
gence and eventually supersede or augment human and how to optimize their AI strategies.
work. AI is becoming the new operational foundation AI has created new business opportunities and de-
of business and has transformed the very nature of
livered value to organizations in numerous ways. For
companies—how they operate and how they compete
example, a chatbot is an AI application that can auto-
(Iansiti and Lakhani 2020). AI has two unique capabili-
ties: automation and smartness, which are associated mate basic, repeatable, standardized interactions be-
with physical machines or software that replace manu- tween customers and sellers. Specifically, chatbots
al work through automated processes or augment hu- such as IKEA’s Anna use voice or texts to automate
man work through smart decisions (Boute and Van communications and create personalized customer ex-
Mieghem 2021). AI can automate simple, tedious, and periences. The chatbot market size is predicted to ex-
repetitive tasks to perform them faster and cheaper. AI pand from $250 million in 2017 to $1.34 billion in 2024
691
Cui, Li, and Zhang: AI and Procurement
692 Manufacturing & Service Operations Management, 2022, vol. 24, no. 2, pp. 691–706, © 2021 INFORMS
(Pise 2018), and the adoption of the chatbot feature is Alibaba’s trading platform 1688 (1688.com), which is
predicted to save businesses $11 billion annually by the largest domestic trading platform in China. It
2023 (Hampshire 2018). serves millions of buyers and suppliers who use an in-
AI has also been applied to automate procurement tegrated chat program called Aliwangwang to com-
tasks and assist strategic sourcing in business-to-busi- municate with each other. The trading company’s pro-
ness (B2B) markets, which is referred to as cognitive curement managers are required to quote prices from
procurement (Loo and Santhiram 2018). Surveys re- suppliers on 1688. We design a 3 × 3 field experiment.
veal that 60% of companies use AI to automate the re- The procurement representatives are (1) a female hu-
quest-for-quotation process and 50% of companies man, (2) a male human, or (3) a chatbot, where the
use AI to recommend new suppliers (Tata Consultan- chatbot automatically sends inquiry messages without
cy Services 2016). human involvement. The quoting messages indicate
that the supplier is (1) not informed of any recommen-
There are two ways in which AI can be used for
smarter sourcing in procurement. The first is the
dation, (2) recommended by a peer, or (3) recom-
automation … . For example, AI-powered … bots … . mended by an AI algorithm. We test the effect of auto-
The second—and more important—use relates to AI- mation and smartness in procurement by comparing
powered tools helping to rapidly collect, present and suppliers’ wholesale price quotes across the aforemen-
even analyse commodity, market, and supply intelli- tioned three buyer types and three recommendation
gence to inform market strategies. conditions.
We find that when automation is not equipped
—Nicholas Walden, Senior Director at The Hackett
Group (HICX Solutions 2018) with a smart control, it negatively affects the quoted
wholesale price. Specifically, chatbot buyers receive
On one hand, chatbots have been used to automate a higher price quote than human buyers. This is be-
the request-for-quotation process in procurement by cause a supplier might believe that a chatbot buyer
mimicking human interactions, thereby relieving lacks the expertise in product specifics, and in turn,
workers from tedious and repeatable tasks. For exam- has a higher reservation price and a higher willing-
ple, SAP Ariba—an information technology services ness to pay than human buyers. Moreover, a suppli-
company in the United States—uses a procurement AI er does not have to lower his wholesale price in
assistant to request price quotations and draft con- order to develop a professional relationship with a
tracts. Chatbots have been shown to reduce labor chatbot buyer. Consequently, the supplier prices dis-
costs by 39% for a global energy company by auto- criminate against chatbot buyers. In addition, as a
mating procurement processes (Papa et al. 2019). On side finding, our results reveal that the wholesale
the other hand, procurement managers can also use prices quoted to male and female buyers are not sig-
AI to identify potential suppliers, which is referred to nificantly different.
as AI recommendation. Traditionally, procurement We find that signaling the use of AI algorithms in
companies often identify potential suppliers based on selecting the supplier reduces the wholesale price for
their colleagues’ recommendation, which is referred chatbot buyers, but it cannot reduce the price for hu-
to as human recommendation. AI adds the component man buyers. This is because, for chatbot buyers, sup-
of smartness to procurement manager’s supplier se- pliers believe in AI’s capability to collect and learn
lection decisions by using its extraordinary capability from market information and in AI’s complete influ-
to collect and analyze market information. To summa- ence on chatbot buyers’ decisions, thereby changing
rize, in the procurement context, automation helps their perception of chatbot buyers’ reservation price
buyers automatically inquire about prices instead of and willingness to pay. However, human buyers are
asking in person, and smartness aids buyers by using not machines. They are susceptible to their own judg-
an AI algorithm to recommend suppliers. ment and heuristics, thereby making them reluctant
Given that procurement is the core decision in B2B to strictly follow algorithm-suggested decisions (Cui
businesses, it is critical to study how AI creates and et al. 2015, Dietvorst et al. 2018, Ibanez et al. 2018,
delivers value along its automation and smartness ca- Tan and Staats 2020, Sun et al. 2021). Because of such
pabilities. We investigate how buyers’ AI strategies af- decision deviations, suppliers may perceive that hu-
fect suppliers’ wholesale pricing decisions. Specifi- man buyers do not follow AI’s recommendations,
cally, we study the effect of automation—that is, thereby ignoring these buyers’ use of AI and not al-
whether the buyer inquires about prices using an au- tering the wholesale price accordingly. In contrast,
tonomous chatbot or in person. We also study the ef- we find that the traditional recommendation without
fect of smartness—that is, whether the buyer signals smart controls—that is, human recommendation—
the use of AI recommendations in selecting suppliers. cannot reduce the price quotes for either chatbot
In this study, we run a field experiment by collabo- buyers or human buyers. This allows us to tease out
rating with a trading company that operates on the effect of recommendation and attribute the
Manufacturing & Service Operations Management, 2022, vol. 24, no. 2, pp. 691–706, © 2021 INFORMS 693
overall effect of AI recommendation to the effect of Our paper is closely related to Boute and Van
smartness. Mieghem (2021). The authors propose a framework
In summary, in the absence of smart controls, the that synthesizes automation and smartness for compa-
buyer suffers from automation by receiving a higher nies who transform operations digitally. They argue
wholesale price, whereas having smart controls leads that having a smart control is necessary before high
to a lower wholesale price for these autonomous levels of autonomy are considered. Our paper follows
buyers. This implies that when automation is imple- this framework to study the value and synergies be-
mented without smart controls, it is not very valuable, tween automation and smartness in procurement pro-
which suggests that building smartness is necessary cesses. Our findings echo the insights of Boute and
before implementing high levels of autonomy. Van Mieghem in that we empirically show that auto-
Last, we study the combined value of automation mation, when implemented without smart controls,
and smartness. We find that chatbot buyers aided by a does not bring value and can even backfire, whereas
smart recommendation system receive the lowest price smartness is valuable. Specifically, we find that auto-
quote among all conditions. In other words, AI delivers mation causes suppliers to increase their wholesale
the most value when buyers use both automation and prices, but AI recommendations can effectively lower
recommendation in price inquiry. This finding high- the price quotes. Consequently, AI delivers the most
lights the value of using autonomous agents aided by a value when automation and smartness are adopted in
smart recommendation system in procurement. combination with each other.
2. Literature Review 2.2. Procurement and Wholesale Pricing

2.1. AI Automation and Recommendation Procurement is a critical business decision. The litera-
Prior research indicates that automation creates value ture has studied various mechanisms, such as inven-
in inventory replenishment (Van Donselaar et al. 2010) and tory investment (Jain et al. 2014), information provi-
financial services (Köhler et al. 2011, Luo et al. 2019, Aci- sion (Engelbrecht-Wiggans and Katok 2008), timing of
movic et al. 2021). An application of automation is a chatbot, auctions (Bimpikis et al. 2020), and trust (Fugger et al.
which helps human workers automate communications 2019) to improve procurement effectiveness. We fol-
with consumers. Extant literature reveals that consumers of- low suit to study the integration of AI as a market
ten dislike communicating with a chatbot, despite the fact mechanism to affect request-for-quotation outcomes.
that automation can improve agents’ productivity. We com- The procurement outcome that we measure is the
plement this literature by investigating suppliers’ reactions wholesale price charged by sellers. Wholesale pricing
to the procurement managers’ usage of chatbot. is one of the central topics of supply chain manage-
Prior research has also shown that AI’s recommen- ment (Cachon 2003, Cachon and Netessine 2006). In
dations add value in various contexts, such as disease supply chains, the wholesale price that suppliers
diagnosis (Leachman and Merlino 2017), wholesale charge for downstream buyers is an important deter-
pricing (Karlinsky-Shichor and Netzer 2019), and minant of suppliers’ profit margins and buyers’ pri-
product recommendations (Häubl and Trifts 2000). ces, which in turn affects profitability. A supplier may
For example, Karlinsky-Shichor and Netzer (2019) cre- charge different wholesale prices to retailers based on,
ate an AI version of B2B salespersons’ pricing deci- for example, buyer intermediation (Tunca and Zhu
sions that mimics their past pricing behavior, which 2018), supplier–buyer relationships (Zhang et al.
improves profits by 10%. However, human decision- 2014), or buyers’ race (Cui et al. 2020). We contribute
makers often choose to rely on their own judgment, to the literature by studying whether suppliers price
making them reluctant to strictly follow algorithm-in- against or in favor of chatbot buyers and, if so, which
structed decisions. Such decision deviation behavior features of AI allow it to deliver the most value.
has been widely documented in the literature. For ex-
ample, managers tend to use their own demand fore- 3. Research Hypotheses
casts rather than forecasts provided by machines (Cui We study how suppliers vary their wholesale prices
et al. 2015, Dietvorst et al. 2018); doctors prioritize tasks to buyers with and without the use of AI on an online
in a manner that deviates from system recommenda- B2B platform. Before purchasing a product, buyers re-
tions (Ibanez et al. 2018); workers pack orders in boxes search its market price by asking for price quotes from
larger than the size suggested by the system (Sun et al. suppliers. Suppliers then provide a price quote to
2021); and restaurant managers deviate from the rout- buyers based on buyers’ characteristics and the inqui-
ing rules that they are instructed to follow (Tan and ry message. In this section, we develop a framework
Staats 2020). We add to this literature by studying sup- that predicts the effect of automation and smartness
pliers’ reactions when B2B buyers tell them that they in procurement. We discuss whether suppliers price
(the suppliers) are recommended by AI algorithms. against or in favor of (1) buyers’ autonomous
characteristic—whether the buyer is a chatbot or hu- 3.2. Effect of Smartness

man, (2) buyers’ smartness characteristic—whether In this section, we study the effect of having smartness
the buyer signals the use of AI recommendations in in the process of wholesale price inquiries. Smartness
selecting suppliers, and (3) buyers’ autonomous and means that the buyer uses AI recommendations to se-
smartness characteristics—whether the buyer is a lect suppliers. Without claiming the use of AI recom-
chatbot with a smart control or a human without a mendations, suppliers would not be able to know and
smart control. react to this. Therefore, in our research context, smart-
ness is signaled to the suppliers. Specifically, when

3.1. Effect of Automation asking for a price quote, the buyer tells the supplier
When deciding on a wholesale price to charge buyers that the company was recommended to the buyer
in a B2B setting, a supplier’s most pivotal consider- by AI’s market search and data analysis. The sup-
ation is the buyer’s best alternative to a negotiated plier can use this information to update beliefs
agreement (BATNA). BATNA refers to the most ad- about the buyer and alter the offered wholesale
vantageous alternative action that the buyer can take price accordingly.
if the negotiation reaches an impasse (Fisher and Ury The information on the use of AI recommendations
1981, Fisher et al. 2011, Pinkley et al. 1994). Conse- can be a key determinant for suppliers, because AI
quently, the buyers’ BATNA determines the suppli- has an extraordinary capability to collect and learn
ers’ pricing strategy: buyers with a stronger BATNA from market information (Häubl and Trifts 2000).
have better outside options, and in turn, they have a Knowing that the buyer has comprehensive knowl-
lower reservation price and a lower willingness to pay edge of the market aided by AI, the supplier would
(Korobkin 2014), which results in a lower wholesale believe that the buyer has a stronger BATNA—that is,
price charged by suppliers. a better walk-away option—and in turn would con-
We consider the scenario that the chatbot or human sider lowering the price. In addition to the capability
buyer asks for prices without providing any recom- of AI, suppliers would also evaluate whether the AI
mendation information to the supplier, that is, automa- recommendations have a complete influence on
tion without smartness. Autonomous chatbots are an buyers’ decisions. If a buyer does not follow AI-gener-
effective tool to automate repeated inquiries and pre- ated recommendations, then the buyer’s decisions
programmed responses. In our research context, a chat- will not heavily rely on AI, which suggests that the
bot is used to automatically send inquiry messages to a supplier could ignore the buyer’s AI use.
group of suppliers, saving buyers’ time spent in send- We first consider the scenario where the procure-
ing messages to each supplier personally. However, ment manager uses autonomous chatbots to ask for
these traditional chatbots, when their main objective is prices and signal that the supplier was selected by
to repeat tasks without smart controls, are not AI’s market search. Because chatbots are machines
equipped to address the complex requirements of B2B programmed to follow the AI’s recommendations, the
suppliers, who expect in-depth communications and supplier would believe in the thorough knowledge of
negotiations with buyers (Swanson 2015). We inter- the market gained by AI and the influence that AI has
viewed several highly experienced trading managers on the buyer. Therefore, we expect that chatbot
who confirm that procurement requires a significant buyers’ use of AI recommendations will improve the
level of professional knowledge in product specifics, supplier’s perception of their BATNA, thereby reduc-
such as product materials, size, functionality, and after- ing the supplier’s wholesale price.
sales service, which preprogrammed chatbots might be Next, we consider the scenario where the human
less knowledgeable in.1 Consequently, suppliers may buyer asks for prices in person and informs the sup-
plier that the company was recommended by AI.
believe that chatbots lack expertise in product specifics
Humans are not machines. They are susceptible to
and, in turn, have a worse BATNA and thus a higher
their own judgment and heuristics, thereby making
reservation price than human buyers.
them reluctant to strictly follow algorithms and rules.
Furthermore, because chatbots lack personal feelings
This is known as decision deviation (Cui et al. 2015,
and empathy, suppliers do not need to lower the whole-
Dietvorst et al. 2018, Ibanez et al. 2018, Tan and Staats
sale price in order to develop a serious relationship with
2020, Sun et al. 2021). Such deviation behavior from
chatbot buyers (Dietvorst et al. 2018, Luo et al. 2019).
algorithm-instructed decisions has been widely docu-
Therefore, we expect that without smart controls, chat-
mented in the literature. For example, managers are
bot buyers will be price discriminated against and re-
shown to use human forecasts rather than algorithmic
ceive a higher price quote than human buyers.
forecasts (Cui et al. 2015, Deloitte 2018); doctors are
Hypothesis 1 (Automation). Without smart controls, shown to prioritize tasks in a manner that deviates
chatbot buyers receive a higher wholesale price quote than from system recommendations (Ibanez et al. 2018);
human buyers. workers are shown to pack orders in boxes larger than
the system-recommended size (Sun et al. 2021); and would perceive that chatbot buyers have more compre-
restaurant managers are shown to deviate from the hensive market knowledge, thereby a stronger BATNA
routing rules that machines instruct them to follow with a lower reservation price than human buyers. We
(Tan and Staats 2020). Given this widespread recogni- hypothesize this relation in Hypothesis 3(a).
tion that humans often deviate from algorithmic Next, we study the difference between chatbot
recommendations, we expect that suppliers would an- buyers with AI recommendations and human buyers
ticipate human buyers to not strictly follow AI without any recommendation. According to the previ-
recommendations. ous theories, AI enables buyers to have comprehen-

To further confirm that decision deviations exist in sive knowledge about the market and exerts a full in-
procurement, we interviewed nine professional B2B fluence on chatbot buyers. As a result, suppliers
suppliers with an average of 12 years of trading expe- would perceive chatbot buyers with smartness to
rience. In the interviews, we asked them to share have a stronger BATNA and thus would offer them a
whether they believe in AI recommendations’ influ- lower wholesale price than human buyers without
ence on a human buyer or a chatbot buyer. We sum- smartness. Therefore, we expect that the effect of AI is
marize their quotes in Table EC.1 in the online appen- the strongest when both automation and smartness
dix. Most suppliers indicate that they believe that are in place.
such an AI recommendation can help chatbot buyers Hypothesis 3 (Automation and Smartness). (a) When
learn about market knowledge and can dictate their informing suppliers that they are selected by smart AI algo-
sourcing choices. However, eight of nine suppliers do rithms, chatbot buyers receive a lower wholesale price quote
not believe that a human buyer will follow AI recom- than human buyers. (b) Chatbot buyers, when informing
mendations because they think that human buyers suppliers that they are selected by smart AI algorithms, re-
would make their own judgment about the market ceive a lower wholesale price quote than human buyers
and are likely to deviate from algorithms. There- without AI recommendations.
fore, we expect that suppliers would assume that
human buyers do not follow the AI’s recommenda- We summarize the effect of automation and smart-
tions, thereby ignoring buyers’ use of AI and not ness on suppliers’ price decisions in Figure 1. We fol-
altering their perception on the human buyers’ low the framework of Boute and Van Mieghem (2021)
BATNA. In other words, the use of AI becomes in- to classify buyers’ AI strategies into four groups: hu-
effective in reducing the wholesale price quote for man buyer without the help of AI, automation en-
human buyers. abled by chatbot buyers, smart control enabled by AI
recommendations, and the joint application of auto-
Hypothesis 2 (Smartness). (a) Chatbot buyers, when in- mation and smartness. In this framework, Hypothesis 1
forming suppliers that they are selected by smart AI algo- describes the pure effect of automation when we
rithms, receive a lower wholesale price quote than chatbot move from the Human Buyer zone to the Automated
buyers without AI recommendations. (b) Human buyers, zone; Hypothesis 2 describes the effect of smartness
when informing suppliers that they are selected by smart on human buyers and chatbot buyers separately
AI algorithms, receive a similar wholesale price quote as hu- when we move from the Human Buyer zone to the
man buyers without AI recommendations. Smart zone, and from the Automated zone to the
3.3. Automation and Smartness

In this section, we study the effect of having both au- Figure 1. Framework of Automation and Smartness in
tomation and smartness. We first discuss the effect of Procurement
automation under smart controls. That is, we compare
the difference between chatbot buyers with AI recom-
mendations and human buyers with AI recommenda-
tions. When both human and chatbots are equipped
with AI recommendations, the effect boils down to
who would follow the AI’s recommendations. Ac-
cording to Hypothesis 2(a), chatbots are programmed
to follow the AI’s recommendations. According to Hy-
pothesis 2(b), human buyers may not fully follow the
AI’s recommendations because of their tendency to
deviate from AI-instructed decisions. Therefore, sup-
pliers will react to chatbot buyers’ and ignore human
buyers’ use of AI recommendations. Taken together, Notes. +, –, and represent higher, lower, and similar price quotes,
when equipped with a smart control, suppliers respectively. H1–H3 represent Hypotheses 1–3, respectively.
Automated+Smart zone, respectively; Hypothesis 3 5.1. Study Design

describes the effect of automation under smartness In order to study the effect of automation, we design
when we move from the Smart zone to the Automa- the sender who asks for the price quote to be a female
ted+Smart zone and the joint effect of automation human, a male human, or an autonomous chatbot.
and smartness when we move from the Human Buy- We identify the value of pure automation by compar-
er zone to the Automated+Smart zone. ing the price quote received by a chatbot buyer and a
human buyer. In order to study the effect of smart-
4. Research Context ness, we design the sender to signal that the supplier
Alibaba Group launched 1688 in 1999, which is the is recommended by AI or human peers, or to not sig-
largest domestic online B2B platform in China (Aliba- nal any recommendation at all. We identify the value
ba 2020a). This platform connects 30 million enterprise of AI recommendations by comparing the price quote
buyers and suppliers (China Daily 2019); the suppliers received with AI recommendations and without any
provide products in 49 major categories, including recommendation. We also introduce a treatment with
apparel, general merchandise, electronics, and car ac- human recommendations, in which the buyer signals
cessories (CNXtrans 2020). The 1688 platform has a that the supplier was recommended by a (human)
built-in instant chat program called Aliwangwang peer, in order to disentangle the pure impact of hav-
that enables buyers to contact suppliers for product ing recommendations and the pure impact of having
specifics and prices. Buying companies are permitted smart controls. If the effect of human recommenda-
to embed autonomous chatbot features in Aliwang- tions is weak, we can attribute the overall effect of AI
recommendations to smartness. Consequently, we use
wang in order to automate communications.
a 3 × 3 experiment design by considering three types
On 1688, a supplier introduces company informa-
of buyers (female buyer, male buyer, and chatbot buy-
tion on a profile page and lists product information on
er) and three recommendation conditions (no recom-
a product page. The supplier’s profile page displays
mendation, human recommendation, and AI recom-
basic company information (e.g., name, location,
mendation). We outline how our experiment design
membership status, and credibility) and trading
matches our AI framework in Figure EC.2 in the
performance on the platform (e.g., number of transac-
online appendix.
tions, number of buyers, repeat purchase rate, and
The company has multiple procurement representa-
refund rate). Suppliers can pay to have an elite mem-
tives whose routine job is to keep track of market dy-
bership in order to obtain advantages in product pro-
namics by collecting wholesale price information. The
motion and exposure. A supplier’s credibility has five
company also uses chatbots to assist in this task. In
levels. The product page displays product characteris-
our study, the procurement representatives follow our
tics—for example, description, picture, price, and op-
scripts and guidelines when quoting wholesale prices
tions—and transaction details—for example, number from suppliers. The trading company asks for price
of reviews, review rating, and transaction volume in quotes via three buying representatives: a female buy-
the past 30 days. er, a male buyer, and a chatbot buyer. We tailor the
A buyer also has a personal profile that includes the messages to incorporate different recommendation
buyer’s name, gender, date of birth, location, photo, conditions. Thereafter, we record and compare suppli-
phone number, and email address. Buyers can search ers’ responses. Table 1 summarizes the study design.
for a specific product and choose one from a list of We select a sample of 3,960 products from 3,960
suppliers displayed by the platform. The buyer can suppliers in the car accessories sector.2 This sector,
then view the supplier’s profile and product details. which is the backbone of China’s industrial ascent
The buyer sends a price quote to the supplier on (Hong and Einhorn 2018), has a large number of sup-
Aliwangwang either through a personal message or pliers on 1688. Car-related products have also been
using autonomous chatbots to automate the inquiry studied to test price discrimination behavior in previ-
process. After receiving an inquiry from a buyer, the ous literature (Busse et al. 2017). In our sample, there
supplier chooses whether to follow up and how much are 14 product subcategories including, for example,
to quote. After transaction details are settled, the buy- automobile data recorders, car cameras, car MP3, ve-
er makes a payment, the supplier ships the order, and hicle displays, vehicle bluetooth headsets, vehicle
the transaction is completed. bluetooth speakers, vehicle-mounted mobile holders,
vehicle chargers, vehicle lockers, car vacuum cleaners,
5. Identification Design GPS locators, vehicle air purifiers, vehicle refrigera-
We aim to study the effect of the buyer’s usage of au- tors, and vehicle-mounted inverters.3 Each supplier
tomation and smartness on the suppliers’ price quot- usually sells a wide selection of products (e.g., a vehi-
ing strategy. We collaborate with a trading company cle refrigerator in capacities of 6, 12, or 20 liters). From
that operates on 1688 to conduct a field experiment. each supplier’s listed products, we select a product
Table 1. Field Experiment Design
Automation × recommendation condition
No recommendation Human recommendation AI recommendation
Design Chatbot Female Male Chatbot Female Male Chatbot Female Male
Planned sample size 440 440 440 440 440 440 440 440 440
Actual sample size 440 439 437 435 436 439 440 439 439
Notes. The planned sample size was 3,960—that is, 440 suppliers per treatment arm. The actual sample size is 3,944 after excluding unavailable
listings.
model that is the most common and standard in the buyers sent price inquiries to the selected suppliers
market. Suppliers are randomly assigned to one of the during the period December 18, 2019, to January 20,
nine (3 × 3) treatment arms. Consequently, we obtain 2020.5 Each message asks for a price quote per unit for
1,320 suppliers per buyer type, 1,320 suppliers per rec- 1,000 units of the preselected product. The message
ommendation condition, and 440 suppliers per treat- content varies based on the recommendation condi-
ment arm. This means that each supplier is quoted tions. In the “no recommendation” condition, the buy-
only once. er includes the most basic information in the inquiry
All our studied products are commodity products. message without indicating any human or AI recom-
Relative to noncommodities that are custom and mendation. In particular, all buyers in this condition
unique, commodities are standard and basic goods. sent a message that said, “Hello, I am [a procurement
One might question that whether procuring standard manager or a chatbot buyer]. We are interested in
commodities requires buyers to have extensive exper- your product: [the specific product name and link of
tise in product specifics. Our interviews with several this product]. Could you please quote us your best
highly experienced trading managers confirm that price per piece for an order of 1,000 units?” The AI
buying commodity products also requires significant chatbot buyers disclose their machine identity in order
professional knowledge such as product materials, to comply with China’s regulation regarding AI
size, functionality, and after-sales service, which ena- (Laskai and Webster 2019). Quoting a price including
bles suppliers to exert in-depth communications and the packaging fee is the industry norm. In order to en-
negotiations. Their exact interview quotes are summa- sure that the quoted prices are not confounded by the
rized in Table EC.2 in the online appendix. On the oth- value-added tax or shipping fee, the buyers ask sup-
er hand, when procuring noncommodity products, pliers to quote a price excluding these fees. The origi-
chatbots might be less knowledgeable in product spe- nal inquiry messages in the field experiment are in
cifics due to their uniqueness. Therefore, the estimated Chinese and are carefully translated and presented in
effect of automation for noncommodities products Figure EC.1 in the online appendix.
might be larger than the effect identified in our study. In the “human recommendation” condition, the
In order to ensure that suppliers are randomly as- buyer discloses that the supplier is recommended by a
signed to treatment arms, we conduct a randomization peer. In the inquiry message, the buyer signals a hu-
check across the following supplier characteristics: (1) man recommendation prior to requesting the price
membership status (i.e., the number of years that the quote: “Your company was recommended to us by a
supplier has been an elite member), (2) credibility (i.e., peer.” We follow the common practice and the indus-
the supplier’s credibility based on the Alibaba credit try norm to not include the peers’ name in the inquiry
system), (3) number of transactions in the past 90 days, message.6 In the “AI recommendation” condition, the
(4) number of buyers in the past 90 days, (5) repeat buyer reveals that the supplier is recommended by
purchase rate in the past 90 days, (6) refund rate in the AI’s market search and data analysis: “Your company
past 90 days, (7) listed price of the selected product, (8) was recommended to us by an AI system’s market in-
trading volume of the selected product in the past 30
formation collection and data processing.”
days; (9) number of reviews for the selected product,
Within a week after the inquiry, we record and
and (10) review rating for the selected product. Table 2
compare the initial price quotes.7 We received 1,807
presents the summary statistics of these variables. Fur-
responses that included a price quote from the 3,944
thermore, the p-values in Table 3 are all larger than
suppliers that we sent messages to.
0.05, which ensures the randomized assignment.
5.2. Study Procedure 6. Results

Buyers’ characteristics (male, female, or chatbot) are In this section, we study the effect of automation and
signaled by their names and profile pictures.4 The smartness on suppliers’ price quoting strategy. We
698
Table 2. Summary Statistics
No. of No. of Repeat Refund Listed No. of Review Trading

Membership Credibility transactions buyers purchase rate rate price reviews rating volume Observations
Chatbot 4.62 3.26 503.8 170.0 28.63 5.88 140.9 25.89 2.46 205.9 1,320
(3.28) (0.90) (1,759) (482.2) (18.19) (10.01) (202.3) (155.5) (2.47) (1,940)
Female 4.47 3.24 597.4 174.9 27.47 6.50 142.0 18.54 2.41 166.3 1,320
(3.08) (0.93) (2,959) (529.6) (17.52) (11.57) (198.4) (102.4) (2.47) (1,625)
Male 4.44 3.20 536.6 171.6 27.58 6.50 140.2 33.96 2.48 140.1 1,320
(3.12) (0.92) (2,144) (495.6) (16.96) (14.28) (197.6) (391.1) (2.47) (1,590)
N 4.79 3.29 502.1 163.7 29.01 5.62 139.7 32.24 2.51 340.5 440
(3.46) (0.91) (1,940) (428.3) (18.40) (9.17) (205.7) (194.4) (2.48) (3,132)
Chatbot H 4.62 3.26 494.1 169.7 29.07 6.08 133.9 25.24 2.35 121.5 440
(3.39) (0.93) (1,391) (477.7) (18.68) (11.25) (191.8) (157.7) (2.47) (650.4)
A 4.45 3.23 515.3 176.4 27.80 5.93 149.1 20.20 2.51 155.8 440
(2.98) (0.85) (1,896) (535.6) (17.48) (9.51) (209.1) (99.68) (2.46) (1,020)
N 4.47 3.25 492.0 167.4 26.12 7.09 132.7 21.84 2.42 145.6 440
(3.05) (0.95) (1,799) (456.1) (16.88) (12.6) (173.6) (148.6) (2.48) (1,057)
Female H 4.50 3.23 683.3 174.2 28.40 5.95 153.3 20.96 2.41 121.2 440
(3.20) (0.97) (3,516) (514.6) (18.55) (10.13) (215.2) (81.47) (2.47) (743.8)
A 4.44 3.23 617.0 183.2 27.90 6.45 140.0 12.83 2.40 232.1 440
(2.98) (0.87) (3,269) (608.0) (17.05) (11.77) (204.0) (52.30) (2.47) (2,501)
N 4.40 3.18 523.1 165.2 27.97 6.57 141.1 30.97 2.58 106.2 440
(2.82) (0.94) (1,743) (439.1) (16.50) (13.11) (193.9) (181.7) (2.48) (824.3)
Male H 4.43 3.19 632.2 171.5 27.50 6.29 140.1 41.95 2.34 105.9 440
(3.14) (0.95) (2,953) (489.7) (17.52) (10.60) (197.3) (636.1) (2.47) (892.2)
A 4.51 3.24 454.3 178.2 27.28 6.64 139.5 28.97 2.53 208.2 440
(3.02) (0.87) (1,429) (552.6) (16.88) (18.13) (202.0) (147.7) (2.47) (1,443)
Note. N, H, and A represent no recommendation, human recommendation, and AI recommendation, respectively.
Manufacturing & Service Operations Management, 2022, vol. 24, no. 2, pp. 691–706, © 2021 INFORMS
Table 3. Randomization Check (p-Value)
Chatbot Female Male
C vs. F C vs. M F vs. M N vs. H N vs. A H vs. A N vs. H N vs. A H vs. A N vs. H N vs. A H vs. A
Membership 0.22 0.15 0.83 0.48 0.13 0.43 0.89 0.88 0.77 0.87 0.56 0.70
Credibility 0.51 0.09 0.30 0.66 0.34 0.62 0.86 0.82 0.97 0.89 0.34 0.42
No. of transactions 0.32 0.67 0.55 0.94 0.92 0.85 0.31 0.48 0.77 0.51 0.52 0.26
No. of buyers 0.80 0.93 0.87 0.84 0.70 0.85 0.84 0.67 0.81 0.84 0.70 0.85
Repeat purchase rate 0.10 0.13 0.87 0.96 0.32 0.30 0.06 0.12 0.68 0.68 0.54 0.85
Refund rate 0.14 0.20 1.00 0.51 0.62 0.83 0.14 0.44 0.50 0.73 0.94 0.72
Listed price 0.89 0.93 0.82 0.67 0.50 0.26 0.12 0.57 0.35 0.94 0.90 0.96
No. of reviews 0.15 0.49 0.17 0.56 0.25 0.57 0.91 0.23 0.08 0.73 0.86 0.68
Review rating 0.65 0.78 0.46 0.34 1.00 0.33 0.97 0.91 0.94 0.17 0.80 0.26
Trading volume 0.57 0.28 0.63 0.15 0.24 0.55 0.69 0.51 0.37 1.00 0.20 0.21
Notes. C, F, and M represent chatbot buyer, female buyer, and male buyer, respectively. N, H, and A represent no recommendation, human rec-
ommendation, and AI recommendation, respectively.
examine the effect of automation by comparing the 6.1.1. Automation Without Smartness. We first focus
price quotes between female (or male) buyers and on the no recommendation condition and investigate
chatbot buyers in Section 6.1, the effect of smartness the effect of automation on suppliers’ price quoting
by comparing the price quotes between the no recom- strategy. Panel A of Table EC.3 in the online appendix
mendation and AI recommendation conditions in Sec- presents the summary statistics of the suppliers’ price
tion 6.2, and the joint effect of automation and smart- discounts for chatbot, female, and male buyers. Figure 2
ness by comparing the price quotes between female presents a visual illustration. In particular, chatbot, fe-
(or male) buyers under the no recommendation condi- male, and male buyers receive an average price dis-
tion and chatbot buyers under the AI recommenda- count of 18.01%, 19.15%, and 20.96%, respectively—that
tion condition in Section 6.3. is, both female and male buyers receive a lower price
quote than chatbot buyers. Moreover, the difference of
6.1. Effect of Automation the price discount between male and chatbot buyers is
In a B2B setting, it is an industry norm and a common statistically significant (p 0.07).
practice that suppliers privately quote a lower price In addition, we formally examine the price differ-
than their publicly listed prices (Cui et al. 2020). In or- ence between chatbot buyers and human buyers:
der to conduct a fair comparison of the amount of
price discount offered by suppliers, we follow previ-
Discounti α + βTypei + γControlsi + i , (2)
ous literature (Cui et al. 2020) to compare the price
discount percentage relative to the listed price:
where Typei is a categorical variable that represents
Discount 100%
whether a buyer is a chatbot, female, or male; Controlsi
Listed Price − Supplier s Quoted Price is a vector of control variables regarding supplier
× :
Listed Price characteristics, including membership status, number
(1) of transactions, listed price, repeat purchase rate, and
number of reviews.
Figure 2. (Color online) Effect of Automation and Smartness

The estimation results are presented in the first col- Hypothesis 3(a). In other words, automation is helpful
umn of Table 4, where the omitted buyers’ type is the in the presence of smartness. This finding echoes the
chatbot group. The coefficients of Female (Male) rep- conjecture of Boute and Van Mieghem (2021): in the
resent the additional price discounts offered to female presence of smart controls, it is conceivable that trust in
(male) buyers relative to chatbot buyers. The coeffi- the algorithm increases and risk is contained, which
cient of Male is weakly positively significant (p < 0.1), opens up the possibility of higher levels of autonomy.
which implies that the supplier quotes a significantly
lower wholesale price to human, particularly male, 6.1.3. Automation Under Human Recommendation. In
buyers than chatbot buyers, which weakly supports addition, from column II of Table 4, we can observe
Hypothesis 1. We conduct several analyses in order to that under the human recommendation condition, the
confirm the robustness of this coefficient: a combined coefficient of Female is weakly positively significant
regression with all the observations in Section 7.1 and (p < 0.1), which implies that the supplier quotes a sig-
an analysis with time fixed effects in Section 7.2. All nificantly lower wholesale price to human—particu-
these analyses support Hypothesis 1. In other words, larly female—buyers than chatbot buyers. In other
the implementation of pure automation does not help words, the implementation of automation still results
buyers and can even backfire in a procurement set- in a higher price even when human recommendations
ting. This is because a chatbot buyer, due to its auton- are adopted. This highlights the importance of having
omous and unsmart nature, signals a higher willing- smart controls when implementing automation in a
ness to pay than human buyers, and human suppliers procurement setting.
are less interested in building a professional relation-
ship with a chatbot buyer. 6.1.4. Gender. A natural extension that we can study
is whether suppliers price discriminate based on
buyers’ gender. Table EC.3 in the online appendix and
6.1.2. Automation Under Smartness. Next, we discuss Figure 2 summarize the price discounts for female
the effect of automation on suppliers’ pricing strategy and male buyers under different recommendation
in the presence of smartness. Panel C of Table EC.3 in conditions. In the no recommendation condition, we
online appendix presents the summary statistics of find that female and male buyers receive an average
the suppliers’ price discounts for chatbot, female, and price discount of 19.15% and 20.96%, respectively;
male buyers under the “AI recommendation” condi- there is no statistically significant difference between
tion. In particular, chatbot, female, and male buyers, male and female buyers (p 0.26). This result also
when equipped with smart recommendations, receive holds under the human recommendation condition
a price discount of 22.57%, 18.76%, and 21.04%, re- and the AI recommendation condition. We also
spectively. The difference between chatbot buyers formally test the price difference based on buyers’
(having automation and smartness) and human gender:
buyers (only smartness) is significant (p-value 0.02).
Discounti α + βGenderi + γControlsi + i , (3)
We also test this effect by using Equation (2) and re-
port the results in column III of Table 4. We can where Genderi is a binary variable that equals one
see that chatbot buyers receive a significantly lower when the buyer is male or equals zero when the buyer
price quote than (particular female) buyers when smart- is female. The estimations are presented in Table 5,
ness is adopted (p-value < 0.05), thereby supporting where the omitted variable is Female; the coefficient
Table 4. Effect of Automation on Price Quote
Dependent variable: Discount
No recommendation Human recommendation AI recommendation All data

(I) (II) (III)I (IV)
Male 0.028* 0.009 −0.015 0.009

(0.016) (0.014) (0.016) (0.009)
Female 0.010 0.026* −0.037** 0.004
(0.016) (0.014) (0.016) (0.009)
Supplier Controls Yes Yes Yes Yes
Observations 595 665 547 1,807
R2 0.047 0.054 0.031 0.033
Notes. This table tests the effect of automation on the price discount for four different samples. Results from columns I–III are based on the sam-
ple under the no recommendation condition, under the human recommendation condition, and under the AI recommendation condition, respec-
tively. Results from column IV are based on the full sample.
*p < 0.1; **p < 0.05.
of Male represents the additional price discount of- condition. The estimation results are presented in Table 6,
fered to male buyers, compared with female buyers, where the omitted variable is the no recommendation
which is not significant. condition.
We show that there is no gender discrimination in The coefficient of AI recommendation represents the
the B2B procurement context. This result differs from additional price discount that a buyer can obtain when
the findings in the business-to-consumer (B2C) set- signaling that the supplier is recommended by an AI
tings—that female consumers receive a higher price algorithm compared with the no recommendation con-
than male consumers because they are perceived to be dition. The coefficient of AI recommendation is signifi-
less knowledgeable (Busse et al. 2017, Mejia and Par- cant and positive (p < 0.05) for a chatbot buyer but not
ker 2021). Unlike B2C consumers whose purchasing significant for female or male buyers. These results con-
decisions are often emotional and irrational, B2B firm that a smart recommendation is effective for low-
buyers are professional procurement managers whose ering prices for chatbot buyers but not for human
job is to negotiate with suppliers. Consequently, male buyers, thereby supporting Hypothesis 2. Because of
and female procurement managers are perceived to AI’s ability to search and learn about market informa-
have a similar willingness to pay (Goldberg 2018). tion, suppliers believe that chatbot buyers have a lower
reservation price and a lower willingness to pay, and
6.2. Effect of Smartness therefore reduce their wholesale price. However, hu-
6.2.1. AI Recommendation. We investigate how AI man buyers are deemed to not fully follow algorithms’
recommendation affects suppliers’ price quoting strat- recommendations and would not be able to benefit
egy for chatbot, female, and male buyers, respectively. from claiming the use of AI recommendations.
Table EC.4 in the online appendix summarizes the In summary, having a purely autonomous process
suppliers’ price discounts for chatbot, female, and leads to a higher wholesale price, putting buyers in a
male buyers. Figure 2 presents an illustration. In par- disadvantageous position, whereas having a smart
ticular, for chatbot buyers, the average price discount control leads to a lower wholesale price. In other
is 18.01% under the no recommendation condition words, automation is not very valuable when imple-
and 22.57% under the AI recommendation condition, mented without smart controls, which suggests that
respectively. This implies that, compared with the no building smartness is necessary before high levels of
recommendation condition, AI recommendation sig- autonomy are to be considered.
nificantly reduces the wholesale price quoted for chat-
bot buyers (p 0.01). For female (male) buyers, the av- 6.2.2. Human Recommendation. Recall that we intro-
erage price discount is 19.15% (20.96%) under the no duced a treatment with human recommendation in
recommendation condition and 18.76% (21.04%) un- order to disentangle the pure impact of having any
der the AI recommendation condition, respectively. recommendation at all and the pure impact of having
This implies that, compared with the no recommenda- smart controls. Next, we study this human recommen-
tion condition, AI recommendation cannot reduce the dation effect. If this effect is weak, then we can con-
wholesale price quoted for female or male buyers. clude that the effect of AI recommendation stems
We also formally examine the impact of recommen- from having smart controls. Table EC.4 in the online
dation conditions on price discounts: appendix and Figure 2 indicate that the average price
discount for chatbot buyers is 18.01% under the no
Discounti α + βConditioni + γControlsi + i , (4)
recommendation condition and 17.39% under the hu-
where Conditioni is a binary variable that represents the man recommendation condition, respectively. We
no recommendation condition or AI recommendation perform a t test and find that the difference is
Table 5. Effect of Gender on Price Quote
No recommendation Human recommendation AI recommendation All data

(I) (II) (III) (IV)
Male 0.020 −0.017 0.022 0.006

(0.016) (0.014) (0.015) (0.009)
Supplier controls Yes Yes Yes Yes
R2 0.046 0.040 0.045 0.033
Notes. This table tests the effect of gender on the price discount for four different samples. Results from columns I–III are based on the sample un-
der the no recommendation condition, under the human recommendation condition, and under the AI recommendation condition, respectively.
Results from column IV are based on the full sample.
Table 6. Effect of Smartness on Price Quote Table 7. Joint Effect of Automation and Smartness on Price
Quote
Chatbot Female Male All data
(I) (II) (III) (IV) Male −0.014
(0.018)
AI recommendation 0.042** −0.003 0.000 0.012
Female −0.034**
(0.017) (0.015) (0.015) (0.009)
(0.017)
Supplier controls Yes Yes Yes Yes

Supplier controls Yes
Observations 562
R2 0.040 0.032 0.058 0.033
R2 0.034
Notes. This table tests the effect of smartness on price discounts for
Notes. This table tests the joint effect of automation and smartness on
four different samples. Results from columns I–III are based on the
price discount. Results are based on the sample of chatbot buyers un-
sample of chatbot, female, and male buyers, respectively. Results
der AI recommendation pooled with human buyers without
from column IV are based on the full sample.
recommendation.
**p < 0.05.
**p < 0.05.
insignificant (p 0.68). We find a similar result for hu-

We summarize the results of buyers’ AI strategies
man buyers that human recommendations cannot
in our framework in Figure 3. First, when a buyer
help human buyers lower the wholesale price re-
adopts pure automation but without smartness by
ceived from suppliers. We conjecture that this is driv-
moving from the Human Buyer zone to the Automat-
en by two reasons. First, because of humans’ limita-
ed zone, the buyer suffers from automation by receiv-
tions in information processing (Payne 1982, Payne
ing a higher price. However, when the buyer adopts
et al. 1988), the supplier may believe that the peer is
automation in the presence of smartness by moving
not capable of providing a valid recommendation.
from the Smart zone to the Automated+Smart zone,
Second, suppliers often distrust buyers when they
the buyer benefits from automation by receiving a
present soft social information like recommendations
lower price. Second, when a human buyer is equipped
(Özer et al. 2014, Cui et al. 2020). Therefore, human
with a smart algorithm by moving from the Human
recommendations, in general, are less effective in
Buyer zone to the Smart zone, smartness does not
changing suppliers’ belief regarding buyers’ willing-
change the price. However, when a chatbot buyer in-
ness to pay and their pricing strategy.
corporates a smart recommendation system by mov-
ing from the Automated zone to the Automated+S-
6.3. Joint Effect of Automation and Smartness mart zone, smartness becomes helpful in reducing the
Thus far, we have demonstrated that automation price. Last, when the buyer adopts both automation
brings a negative effect to buyers, whereas smartness and smartness by moving from the Human Buyer
brings a positive effect. Next, we study the joint value zone to the Automated+Smart zone, the buyer can
of automation and smartness. Table EC.5 in the online also benefit from receiving a lower price quote.
appendix summarizes the price discounts received with
and without automation and smartness. This table re- 7. Robustness Check
veals that autonomous chatbot buyers, when informing
In this section, we conduct additional analysis to
suppliers that they are selected by smart algorithms, re-
check the robustness of our key insights regarding the
ceive a lower wholesale price quote than human (partic-
individual and joint effects of automation and
ularly female) buyers without any recommendation (p
smartness.
0.05). We also formally test this joint effect:
Discounti α + βJointi + γControlsi + i , (5) 7.1. Combined Regression
where Jointi is a categorical variable that represents a In our main analysis, we studied the effect of automa-
chatbot buyer aided by AI recommendations, a female tion and the effect of smartness in separate regres-
buyer without recommendations, or a male buyer sions. To show the robustness of our results, we now
without recommendations. The estimation results are combine all observations with all nine experiment
presented in Table 7, where the omitted variable is conditions into a single regression:
when a buyer is equipped with both automation and Discounti α + β0 Typei + β1 Recommendationi
smartness. Table 7 reveals that having both automa- + β2 Typei × Recommendationi + γControlsi + i ,
tion and smartness can effectively reduce the price for
(6)
(particularly female) buyers (p < 0.05). This implies
that we should improve the levels of autonomy and where Recommendationi is a categorical variable that
smartness simultaneously. represents the no recommendation condition, human
Figure 3. Framework and Results of Automation and Smart- The estimation results with time fixed effects are
ness in Procurement shown in Tables EC.7 and EC.8 in the online appen-
dix. As shown in column I of Panel A in Tables EC.7
and EC.8, the coefficients of Male are weakly positive-
ly significant (p < 0.1), which implies that suppliers
quote a lower wholesale price to human—particularly
male—buyers than chatbot buyers in the absence of
smartness. However, column III of Panel A in Tables

EC.7 and EC.8 shows that the coefficients of Female
are negatively significant (p < 0.05), which implies
that chatbot buyers receive a significantly lower price
quote than (particularly female) buyers when smart-
ness is adopted. This is consistent with our main re-
sults regarding the effect of automation, thereby sup-
Note. +, –, and represent higher, lower, and similar price quotes, porting Hypotheses 1 and 3(a).
respectively. As shown in Panel B of Tables EC.7 and EC.8, the
coefficients of AI recommendation are positively sig-
nificant (p < 0.05) for a chatbot buyer but not signifi-
recommendation condition, or AI recommendation
cant for female or male buyers. This is consistent with
condition.
our main results regarding the effect of smartness,
Table EC.6 in the online appendix reports the esti-
thereby supporting Hypotheses 2(a) and 2(b).
mation results, where the omitted buyers’ type is
As shown in Panel C of Tables EC.7 and EC.8, hav-
the chatbot buyer and the omitted recommendation
ing both automation and smartness can effectively re-
condition is the no recommendation condition. This
duce the price for (particularly female) buyers (p <
table shows three observations. First, the coefficient of
0.05). This is consistent with our main results regard-
Male is weakly positively significant (p < 0.1). That is,
ing the joint effect of automation and smartness, there-
chatbot buyers receive a significantly higher price
by supporting Hypothesis 3.
quote than human buyers without recommendations;
this is consistent with our result on the effect of
automation without smartness, thereby supporting 7.3. Simulated AI Recommendation
Hypothesis 1. Second, the coefficient of AI Recom- In our design, the signal that the supplier is recom-
mendation is positively significant (p < 0.05). That is, mended by AI is randomly assigned to each supplier.
AI recommendations help chatbot buyers receive a lower In practice, it may be true that only some (high-quali-
price quote; this is consistent with our result on the effect ty) suppliers would receive such signals. In order to
of smartness on chatbot buyers, thereby supporting simulate such a scenario, we follow our collaborative
Hypothesis 2(a). Third, the coefficients of Male × Human company’s guide to score 10 supplier/product charac-
Recommendation and Female × Human Recommen- teristics (as shown in Section 5) according to how
dation are not significant, but the coefficients of Male much they determine buyers’ perceptions of suppli-
× AI Recommendation and Female × AI Recommen- ers’ quality. We then apply these scores to compute
dation are weakly negatively significant (p < 0.1). That the perceived quality of each supplier. We define sup-
is, human recommendations cannot reduce price dis- pliers above the average score as high-quality suppli-
crimination against chatbot buyers, but AI recommen- ers and the rest as low-quality suppliers. We then sim-
dations are effective in reducing such price discrimina- ulate an AI recommendation condition where buyers
tion; these results are consistent with our main result equipped with AI recommendation only contact the
on the effect of automation under smartness, thereby high-quality suppliers. In this way, we can simulate
supporting Hypothesis 3(a). the situation where only high-quality suppliers are se-
lected by AI algorithms and approached by buyers.
We next identify the effect of smartness in practice
7.2. Time Fixed Effects by comparing suppliers’ wholesale prices across the
We test our key results by including the time fixed ef- no recommendation condition and the simulated AI
fects at two levels: the inquiries’ request date and the recommendation condition. The average supplier
inquiries’ quote date. Because different suppliers may quality score is 0.25 under the no recommendation
take different amounts of time to respond to a price condition, which is lower than 0.31 under the simulat-
inquiry, the quote dates for the same batch of inqui- ed AI recommendation condition, confirming that
ries might differ. To ensure rigor and robustness, we only high-quality suppliers are included in the
test for both time fixed effects. sample.
Table EC.9 in the online appendix summarizes the For the joint effect of automation and smartness, we
suppliers’ price discounts for chatbot, female, and male use the following estimation. Table EC.13 in the online
buyers under the simulated AI recommendation condi- appendix presents the estimation results.
tion and the no recommendation condition. In particu-
Discounti α + β1 Jointi + β2 Moderatori + β3 Moderatori
lar, for chatbot buyers, the average price discount is
18.01% without recommendations and 23.91% with the × Jointi + γControlsi + i : (10)
simulated AI recommendation. This means that the Overall, none of the studied characteristics (except for
simulated AI recommendation significantly reduces the the listed price) has an impact on the individual and
wholesale price quoted for chatbot buyers (p 0.01). joint effects of automation and smartness. A higher
However, consistent with our main result, the simulat- listed price weakens the effectiveness of smartness for
ed AI recommendation cannot reduce the wholesale chatbot buyers, probably because suppliers are more
price quoted for human buyers; for female (male) prudent when selling expensive products and are less
buyers, the average price discount is 19.15% (20.96%)
likely to regard AI-driven price quotations as a serious
without AI recommendation and 18.76% (21.34%) with
negotiation.
simulated AI recommendation, respectively.
We also formally examine the impact of the simulat-
ed AI recommendation on price by 8. Conclusion
Discounti α + βAIRecommendationi + γControlsi + i , AI is transforming the very nature of procurement—
how to operate and how to interact with supply chain
(7)
partners. According to the Roland Berger’s survey on
where AIRecommendationi is a binary variable that rep- Fortune Global 500 companies, 67% of chief procure-
resents the no recommendation condition or the simu- ment managers rank AI among their top three priori-
lated AI recommendation condition. The estimation ties for the next 10 years (Marlinghaus 2018). Thus, we
results are presented in Table EC.10 in the online ap- explore how a buyer’s AI strategy would affect the
pendix, where the coefficient of simulated AI recom- wholesale price received from suppliers. By designing
mendation is significant (p < 0.05) and positive for and conducting a randomized field experiment, we
chatbot buyers but not significant for human buyers. find that having a purely autonomous request-for-
These results again confirm that a smart recommenda- quotation process results in a higher price quote—that
tion is effective in lowering prices for chatbot buyers is, suppliers price discriminate a not-so-smart chatbot
but not for human buyers. buyer. Furthermore, we find that introducing a smart
7.4. Heterogeneous Treatment Effect control—signaling that the supplier is recommended
We next test whether any supplier or product charac- by a smart system—can reduce the price quoted for
teristics (i.e., the number of transactions, listed price, chatbot buyers. Last, we show that automation and
review rating, and trading volume) could change the smartness can jointly reduce the wholesale price quot-
effect of automation and smartness. ed by suppliers, thereby highlighting the potential of
For the effect of automation, we use the following a smart automation in procurement.
estimation:
Discounti α + β1 Typei + β2 Moderatori + β3 Moderatori 8.1. Managerial Implications
Our work can provide implications for the manage-
× Typei + γControlsi + i , (8)
ment of B2B platforms and buyers aiming to embrace
where β2 represents how a supplier or product char- AI in procurement.
acteristic moderates the effect of automation on the For procurement companies, our study provides
wholesale price quotes. Moderatori represents the strategic guidance for them moving toward in auto-
number of transactions for the supplier, product’s mating their standard and routine processes, such as
listed price, review rating, or trading volume. Con- price quoting and new supplier selection. In fact, ex-
trolsi includes all other control variables except for cessive and duplicated processes can comprise up to
the tested moderator. Table EC.11 in the online 40%–60% of a procurement company’s capacity (Papa
appendix presents the estimation results. et al. 2019). AI is capable of unlocking employees’
For the effect of smartness, we use the following es- workload for more strategic pursuits, thereby trans-
timation. Table EC.12 in the online appendix presents forming the transaction-oriented procurement toward
the estimation results. the strategy-oriented procurement, which is known as
Discounti α + β1 Conditioni + β2 Moderatori Procurement 4.0 (Loo and Santhiram 2018, Marling-
+ β3 Moderatori × Conditioni + γControlsi + i : haus 2018). Our results indicate that in addition to the
advantages of AI in releasing workload, AI also cre-
(9) ates value by reducing the wholesale price.
Our findings further shed light on how to imple- their effect size. Based on a two-sided t test with a power level of
ment AI strategies for procurement companies. First, 0.8 and a significance level of 0.05, we require 99 observations with
a 0.40 effect size between chatbot and female buyers under the no
in the absence of AI smartness, automation alone can recommendation condition, 38 observations with a 0.65 effect size
backfire. This implies that a company should first initi- between chatbot and male buyers under the no recommendation
ate and strengthen its smart control algorithms, such condition, 393 observations with a 0.20 effect size between the no
as improving the data quality, analytics capability, and recommendation and human recommendation conditions, and 164
prediction accuracy of its recommendation systems, observations with a 0.31 effect size between the no recommendation
and AI recommendation conditions. We determined the sample
before considering a high level of autonomy. Second, size per treatment arm to be 440 (>393) to further ensure the validi-
when implementing AI smartness, in order to ensure ty of the experiment.
the effectiveness of smartness, companies should help 3
In order to explore new markets, the trading company specifies
their employees get along with AI—that is, reduce these 14 product categories from which our research team indepen-
their biases and enhance their trust in algorithms. dently selects the supplier and product sample. We validate with
Third, our results suggest that, to obtain the most val- the company that there is no previous supplier in the sample.
4
ue from AI, a company should eventually adopt auto- The chatbot buyer has a standard robotic profile picture. We edit
mation and smartness together in procurement. To the photos of human buyers using Photoshop to ensure their photos
have a similar attractiveness.
conclude, our work unlocks the optimal steps for 5
Our experiment (which was from December 18, 2019 to January
buyers to develop AI in procurement: first build
20, 2020) was conducted before the outbreak of COVID-19 (which
smartness, then sharpen the effectiveness of smartness, caused the first lockdown measure to take place on January 23,
and finally build automation. 2020) and before the Chinese New Year (which was from January
For online trading platforms, our work provides the 24, 2020, to January 30, 2020). As a result, our experiment was not
following managerial implications. Platforms such as affected by the pandemic or the holiday.
6
Alibaba have initiated the automatic request-for-quo- In our human recommendation message design, a buyer does not
tation systems as a premium service provided for provide the name of the peer who recommended the supplier, and
it has been validated that such a design format conforms to norms
buyers (Alibaba 2020b). Our study suggests that such regarding both confidentiality and industry practice (Cui et al.
automatic systems should be facilitated with a smart 2020).
supplier identification system in order to reduce the 7
Following the literature (Ayres and Siegelman 1995, Busse et al.
wholesale price charged to downstream buyers and 2017, Cui et al. 2020), our study focuses on the initial price quote be-
reducing the inefficiencies of supply chains arising cause (1) the initial price quote reflects the supplier’s perception of
from the double marginalization issue. In addition, the buyer’s willingness to pay; (2) suppliers could easily lose cus-
our result highlights that the value of such a smart tomers to competitors if they do not offer an attractive initial price
in an online trading platform; and (3) the initial price quote, unlike
supplier identification system can be much more sig- a second price quote or price concession, is not confounded by any
nificant than human recommendation systems that of- bargaining or negotiation techniques.
ten facilitate peer recommendations between buyers.
AI has become the universal engine of execution,
driving the explosive growth of new business models, References
Acimovic J, Parker C, Drake D, Balasubramanian K (2021) Show or
but there is limited empirical research to understand
tell? Improving agent decision making in a Tanzanian mobile
and quantify how AI works and when it is the most money field experiment. Manufacturing Service Oper. Managment
powerful (Terwiesch 2019, Terwiesch et al. 2020). Our Forthcoming.
study is among the first to research how AI creates and Alibaba (2020a) 1688.com: Leading integrated domestic wholesale
delivers value in a critical business process, namely, marketplace in China. Accessed November 27, 2020, www.
procurement. We hope that our paper will serve as a alibabagroup.com/en/about/businesses.
Alibaba (2020b) Let the right supplier find you with RFQ: Fast re-
stepping stone for future AI-related business research.
sponses from gold suppliers. Accessed November 27, 2020,
https://rfq.alibaba.com.
Ayres I, Siegelman P (1995) Race and gender discrimination in bar-
Acknowledgments gaining for a new car. Amer. Econom. Rev. 85(3):304–321.
The authors thank Hau Lee, the anonymous associate edi- Bimpikis K, Elmaghraby WJ, Moon K, Zhang W (2020) Managing
tor, and anonymous referees for constructive and helpful market thickness in online B2B markets. Management Sci.
feedback. The authors also appreciate the valuable feed- 66(12):5783–5822.
back provided by colleagues at University of Hong Kong. Boute RN, Van Mieghem JA (2021) Digital Operations: Autonomous
automation and the smart execution of work. Management Bus.
Rev. 1(1):177–186.
Endnotes
Busse MR, Israeli A, Zettelmeyer F (2017) Repairing the damage:
1
We discuss our interviews in detail in Section 5.1. The effect of price knowledge and gender on auto repair price
2
The sample size is determined by the statistics power calculation. quotes. J. Marketing Res. 54(1):75–95.
By running a pilot experiment with 40 chatbot buyers, 40 female Cachon GP (2003) Supply chain coordination with contracts. Hand-
buyers, and 40 male buyers under the no recommendation, human book Oper. Res. Management Sci. 11:227–339.
recommendation, and AI recommendation conditions, respectively, Cachon GP, Netessine S (2006) Game theory in supply chain
we compare the price discounts across treatment arms and obtain analysis. Johnson MP, Norman B, Secomandi N, eds. Models,
Methods, and Applications for Innovative Decision Making (IN- Laskai L, Webster G (2019) Translation: Chinese expert group offers
FORMS, Catonsville, MD), 200–233. ‘governance principles’ for ‘responsible AI.’ Accessed Novem-
China Daily (2019) Alibaba’s cross-border e-commerce sees fast ber, 27, 2020, https://perma.cc/V9FL-H6J7.
growth in 2018. Accessed November 27, 2020, www.chinadaily. Leachman SA, Merlino G (2017) Medicine: The final frontier in can-
com.cn/a/201904/04/WS5ca5e33ba3104842260b4819.html. cer diagnosis. Nature 542(7639):36–38.
CNXtrans (2020) Buy any product from 1688.com & ship interna- Loo SK, Santhiram RR (2018) Emerging Technologies for Supply Chain
tionally to your doorstep. Accessed November 27, 2020, www. Management (WOU Press, Malaysia).
cnxtrans.com/1688-agent. Luo X, Tong S, Fang Z, Qu Z (2019) Frontiers: Machines vs. hu-
Cui R, Allon G, Bassamboo A, Van Mieghem JA (2015) Information mans: The impact of artificial intelligence chatbot disclosure on
sharing in supply chains: An empirical and theoretical valua- customer purchases. Marketing Sci. 38(6):937–947.
tion. Management Sci. 61(11):2803–2824. Marlinghaus S (2018) AI and the future of procurement. Accessed
Cui R, Li J, Li M, Yu L (2020) Wholesale price discrimination in November, 27, 2020, www.rolandberger.com/en/Publications/
global sourcing. Manufacturing Service Oper. Management, ePub AI-and-the-future-of-procurement.html.
Mejia J, Parker C (2021) When transparency fails: Bias and financial
ahead of print May 7, https://doi.org/10.1287/msom.2019.0862.
incentives in ridesharing platforms. Management Sci. 67(1):
Deloitte (2018) The global chief procurement officer survey 2018—
166–184.
Leadership: Driving innovation and delivering impact. Ac-
Özer Ö, Zheng Y, Ren Y (2014) Trust, trustworthiness, and informa-
cessed November 27, 2020, www2.deloitte.com/content/dam/
tion sharing in supply chains bridging China and the United
Deloitte/at/Documents/strategy-operations/deloitte-global-cpo-
States. Management Sci. 60(10):2435–2460.
survey-2018.pdf. Papa T, Kaufman A, Maxwell C (2019) When bots do the buying:
Dietvorst BJ, Simmons JP, Massey C (2018) Overcoming algorithm Procurement at half the cost. Accessed June 14, https://www.
aversion: People will use imperfect algorithms if they can (even accenture.com/t00010101t000000z__w__/au-en/_acnmedia/
slightly) modify them. Management Sci. 64(3):1155–1170. accenture/conversion-assets/dotcom/documents/global/pdf/
Engelbrecht-Wiggans R, Katok E (2008) Regret and feedback infor- strategy_8/accenture-transcript-bots-tom-papa.pdf.
mation in first-price sealed-bid auctions. Management Sci. Payne JW (1982) Contingent decision behavior. Psych. Bull.
54(4):808–819. 92(2):382.
Fisher R, Ury W (1981) Getting to Yes: How to Negotiate Without Giv- Payne JW, Bettman JR, Johnson EJ (1988) Adaptive strategy selection
ing In (Arrow Books, London). in decision making. J. Experiment. Psych. Learning Memory Cogni-
Fisher R, Ury WL, Patton B (2011) Getting to Yes: Negotiating Agree- tion 14(3):534.
ment Without Giving In (Penguin, New York). Pinkley RL, Neale MA, Bennett RJ (1994) The impact of alternatives
Fugger N, Katok E, Wambach A (2019) Trust in procurement inter- to settlement in dyadic negotiation. Organ. Behav. Human Deci-
actions. Management Sci. 65(11):5110–5127. sion Processes 57(1):97–116.
Goldberg M (2018) B2B and B2C advertising is night and day. Ac- Pise R (2018) Chatbot market size is set to exceed USD 1.34 billion
cessed November, 27, 2020, www.dnb.com/perspectives/ by 2024. Accessed November 27, 2020, www.clickz.com/
marketing-sales/b2b-b2c-advertising-differences.html. chatbot-market-size-is-set-to-exceed-usd-1-34-billion-by-2024/
Hampshire (2018) Chatbots to deliver $11bn in annual cost savings. 215518.
Accessed November, 27, 2020, www.juniperresearch.com/ Sun J, Zhang D, Hu H, Jan A Van M (2021) Predicting human dis-
press/press-releases/chatbots-to-deliver-11bn-cost-savings- cretion to adjust algorithmic prescription: A large-scale field ex-
2023. periment in warehouse operations. Management Sci.
Häubl G, Trifts V (2000) Consumer decision making in online shop- Forthcoming.
ping environments: The effects of interactive decision aids. Mar- Swanson L (2015) Why B2B needs knowledge bots instead of chat
keting Sci. 19(1):4–21. bots. Accessed November, 27, 2020, https://medium.com/@
HICX Solutions (2018) The AI revolution in procurement. Accessed lswanson/why-b2b-needs-knowledge-bots-instead-of-chat-bots-
November, 27, 2020, www.raconteur.net/wp-content/uploads/ 9eb271c7f542.
2018/09/AI_Revolution_in_Procurement_HICX-1.pdf. Tan TF, Staats BR (2020) Behavioral drivers of routing decisions: Ev-
Hong J, Einhorn B (2018) Trade war with China may hit car part idence from restaurant table assignment. Production Oper. Man-
agement 29(4):1050–1070.
makers first. Accessed November 27, 2020, www.newequipment.
Tata Consultancy Services (2016) Getting smarter by the day: How
com/industry-trends/article/22060167/trade-war-with-china-
artificial intelligence is elevating the performance of global
may-hit-car-part-makers-first.
companies. Accessed November 27, 2020, http://sites.tcs.com/
Iansiti M, Lakhani KR (2020) Competing in the Age of AI: Strategy and
artificial-intelligence/wp-content/uploads/TCS-GTS-how-AI-
Leadership When Algorithms and Networks Run the World (Har-
elevating-performance-global-companies.pdf.
vard Business Review Press, Cambridge, MA). Terwiesch C (2019) OM forum—Empirical research in operations
Ibanez MR, Clark JR, Huckman RS, Staats BR (2018) Discretionary management: From field studies to analyzing digital exhaust.
task ordering: Queue management in radiological services. Manufacturing Service Oper. Management 21(4):713–722.
Management Sci. 64(9):4389–4407. Terwiesch C, Olivares M, Staats BR, Gaur V (2020) A review of em-
Jain N, Girotra K, Netessine S (2014) Managing global sourcing: In- pirical operations management over the last two decades.
ventory performance. Management Sci. 60(5):1202–1222. Manufacturing Service Oper. Management 22(4):656–668.
Karlinsky-Shichor Y, Netzer O (2019) Automating the B2B salesper- Tunca TI, Zhu W (2018) Buyer intermediation in supplier finance.
son pricing decisions: Can machines replace humans and Management Sci. 64(12):5631–5650.
when? Working paper, Northeastern University, Boston. Van Donselaar KH, Gaur V, Van Woensel T, Broekmeulen RACM,
Köhler CF, Rohm AJ, de Ruyter K, Wetzels M (2011) Return on in- Fransoo JC (2010) Ordering behavior in retail stores and impli-
teractivity: The impact of online agents on newcomer adjust- cations for automated replenishment. Management Sci.
ment. J. Marketing 75(2):93–108. 56(5):766–784.
Korobkin R (2014) Negotiation: Theory and Strategy, 3rd ed. (Wolters Zhang JZ, Netzer O, Ansari A (2014) Dynamic targeted pricing in
Kluwer Law & Business, New York). B2B relationships. Marketing Sci. 33(3):317–337.

2022, Cui

Uploaded by

Copyright:

Available Formats

2022, Cui

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2022, Cui

Uploaded by

Copyright:

Available Formats

This article was downloaded by: [218.35.205.

47] On: 25 April 2023, At: 18:11

Manufacturing & Service Operations Management

To cite this article:

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Copyright © 2021, INFORMS

Please scroll down for article—it is on subsequent pages

77204; c Business School, Nankai University, Tianjin 300071, China

Keywords: artiﬁcial intelligence • procurement • wholesale pricing • automation • smartness

2. Literature Review 2.2. Procurement and Wholesale Pricing

characteristic—whether the buyer is a chatbot or hu- 3.2. Effect of Smartness

ness is signaled to the suppliers. Speciﬁcally, when

recommendations. ous theories, AI enables buyers to have comprehen-

3.3. Automation and Smartness

Automated+Smart zone, respectively; Hypothesis 3 5.1. Study Design

Table 1. Field Experiment Design

Automation × recommendation condition

No recommendation Human recommendation AI recommendation

5.2. Study Procedure 6. Results

Table 2. Summary Statistics

No. of No. of Repeat Refund Listed No. of Review Trading

Table 3. Randomization Check (p-Value)

Chatbot Female Male

Figure 2. (Color online) Effect of Automation and Smartness

Table 4. Effect of Automation on Price Quote

Dependent variable: Discount

No recommendation Human recommendation AI recommendation All data

Male 0.028* 0.009 −0.015 0.009

Table 5. Effect of Gender on Price Quote

Dependent variable: Discount

No recommendation Human recommendation AI recommendation All data

Male 0.020 −0.017 0.022 0.006

Supplier controls Yes Yes Yes Yes

insigniﬁcant (p 0.68). We ﬁnd a similar result for hu-

smartness. However, column III of Panel A in Tables

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.