Consumer-driven e-commerce: A literature review, design framework, and research agenda on last-mile logistics models

Planning

At the outset of this study, a panel was formed comprising academic scholars and industry practitioners who are experts in the field of LML and retail operations. These experts participated in structured meetings in which the critical research needs related to LML design were discussed. The following research questions were then formulated based upon the emergent themes:

Searching

The academic material in this study covers the period from 2000 to 2017. This period coincides with critical industry events, such as the emergence and subsequent demise of the online grocer, Webvan. The review is limited to peer-reviewed publications to ensure the quality of the corpus (Saenz and Koufteros, 2015) and considers 16 journals, including one practitioner journal (MIT Sloan Management Review), to capture theoretical perspectives on industry best practices. Only articles from the selected journals have been included in this review, with one exception, where we included the article by Wang et al. (2014), published in Mathematical Problems in Engineering. The article was deemed critical as it represents the only piece of work to date that connects and extends prior research on the evaluation of CDPs.

The 16 journals were selected based on their primary focus on empirical and conceptual development, rather than on their discussion of analytical modelling. Although we appreciate that there are significant research studies in this area (e.g. operations research), the focus of this review led us to primarily consider how scholars conceptualise LML distribution structures and apply theoretical variables to LML design through quantitative, qualitative, or conceptual approaches, rather than through mathematic-based models. The mathematic-based model literature focuses on the development of stylised and optimisation models in areas of multi-echelon distribution systems, vehicle routing problems (Savelsbergh and Van Woensel, 2016), buy-online-pick-up-in-store services (Gao and Su, 2017), pricing and delivery choice, inventory-pricing, delivery service levels, discrete location-allocation, channel design, and optimal order quantities via newsvendor formulation for different fulfilment options (Agatz et al., 2008), amongst others. These studies typically employ a series of assumptions to simplify real-world operations in order to provide closed-form or heuristic-based prescriptive solutions (Agatz et al., 2008; Savelsbergh and Van Woensel, 2016). Therefore, this review excluded journals with a primarily mathematical modelling or operations research focus. However, it included relevant mathematical modelling articles – published in any of the 16 selected journals – as long as they explicated types of LML distribution structure and/or the associated contingency variables. Finally, this study also excluded general management journals in order to fit the operational focus of this research.

The literature search was conducted on the following databases: ISI Web of Science, Science Direct, Scopus, and ABI/Inform Global. Two search rounds were undertaken to maximise inclusion of all relevant articles. The first literature probe was performed using the following search terms: “urban logistics” OR “city logistics” OR “last-mile logistics” OR “last mile logistics”. To extend the corpus, we incorporated the “snowballing” technique of tracing citations backward and forward to locate leads to other related articles; this study used this process in the second round to supplement a protocol-driven methodology. This approach resulted in new search terms and scholar identification to refine the search strategy as the study unfolded (Greenhalgh and Peacock, 2005). The following new search terms were identified: “home delivery”, “B2C distribution”, “extended supply chain”, “final mile”, “distribution network”, “distribution structure”, and “grocery delivery”. These new keywords were then used to create additional search strings with Boolean connectors (AND, OR, AND NOT). Finally, in order to cross-check the searches, we consulted with a supply chain professor from Arizona State University and one from the University of Cambridge. It is therefore posited that the review coverage is reasonably comprehensive.

Screening

For the purpose of this review, the abstract of each paper selected was studied in order to assess its relevance to LML in the e-commerce context. When an abstract proved inconclusive or unavailable, the full paper was examined to determine the relevant contribution to the LML topic. We evaluated article relevance using the following criteria in order to ensure fidelity and objectivity:

1. Exclusion criteria: paper titles bearing the terms “urban logistics”, “city logistics”, “last-mile logistics”, or “last-mile” but with limited coverage on distribution structures and the associated design variables were excluded (e.g. public policy, urban traffic regulations, logistics infrastructure, urban sustainability, environment), as were editorial opinions, conference proceedings, textbooks, book reviews, dissertations, and unpublished working papers.

2. Inclusion criteria: papers with coverage of distribution structures and design variables in the e-commerce context were included, regardless of their actual study focus. We included multiple research methods to have both established findings as well as emergent theorising.

During the search phase, we identified 425 articles referencing our subject terms. We eliminated duplicates based on titles and name of authors and rejected articles matching the exclusion criteria. For example, while the paper by Gary et al. (2015) holds the keyword “urban logistics” in the title, it focuses on logistics prototyping, rather than LML models, as a method to engage stakeholders. This paper, therefore, was excluded. The elimination stage resulted in 100 articles being considered relevant for further review. Results were exported to reference management software, EndNote version X8, for further review and to facilitate data management. We then adopted the inclusion criteria to select the final articles. Finally, we grouped the articles into two categories: LML distribution structures and the associated contingency variables. Ultimately, a total of 47 journal articles form the corpus of this review.

Extraction, synthesis, and reporting

Following an initial review of the 47 articles, a summary of each article was prepared using a spreadsheet format organised under descriptive (year, journal, title), methodology (article type, theoretical lens, sampling protocol), and thematic categories (article purpose, context, LML distribution structures, design variables, others) as adapted from Pilbeam et al. (2012).

Accordingly, we conducted three analyses: descriptive, methodological, and thematic (Richard and Beverly, 2014). The descriptive analysis summarises the research development over the period of interest, and the distribution statistics of the journals. The methodological analysis highlights the research methods employed in the domain, while the thematic analysis synthesises the main outcomes from the extracted literature and provides an overview of the review structure. Reporting structures were organised in a manner that sequentially responds to the research questions posed earlier.

Push-centric system: n-tier direct to home

This study found that the push-centric system is the most commonly adopted distribution form. It typically comprises a number of intermediate stages (n-tier) between the source and destination in order to create distribution efficiencies. The literature classifies three picking variants according to fulfilment (inventory) location: manufacturer-based (or “drop-shipping”), DC-based, or local B&M store-based (i.e. retailer’s intermediate warehouse or store). The destination can either be consumers’ homes or, increasingly, their workplaces (McKinnon and Tallam, 2003). The mode of delivery can be in-sourced (using retailer’s own vehicle fleet), outsourced to a third-party logistics provider (3PL) (Boyer and Hult, 2005), or crowd-sourced using independent contractors (Wang et al., 2016).

When selecting a distribution channel, retailers need to trade-off between fulfilment capabilities, inventory levels (Netessine and Rudi, 2006), product availability and variety (Agatz et al., 2008), transportation cost (Rabinovich et al., 2008), and responsiveness (Chopra, 2003). The nearer the picking site is to the consumer segment, the more responsive is the channel. However, this responsiveness comes at the expense of lower-level inventory aggregation and higher risks associated with stock-outs (Netessine and Rudi, 2006).

Pull-centric system: consumer self-help

The literature also discussed two variants of the pull-centric system. Both variants require consumers to participate (or self-help) throughout the transaction process, from order fulfilment to order transportation. The first variant represents the traditional way of shopping at a local B&M store, with consumers performing the last-mile “delivery”. The second “information store” variant adopts a concept known as “dematerialisation”, substituting information flow for material flow (Lee and Whang, 2001). This variant recognises that material or physical flows are typically more expensive than information flows due to the costs of (un)loading, handling, warehousing, shipping, and product returns.

This study found that despite the popularity of online shopping, there are still occasions where consumers favour traditional offline shopping. Perceived or actual difficulties with inspecting non-digital products, the product returns process, or slow and expensive shipping can deter consumers from online shopping (Forman et al., 2009). This study also demonstrates other benefits of a pull-centric system, including lower capital investments and possible carry-over (or halo) effects into in-store sales (Johnson and Whang, 2002).

Hybrid system: n-tier to consumer self-help location

The rich literature here mainly compared different modes of reception. Variants typically entailed a part-push and part-pull configuration. For instance, the problem associated with “not-at-home” responses within attended home delivery (AHD) can be mitigated by delivering the product to a CDP for consumers to pick up. The literature discussed two CDP variants: CDP-A and CDP-U. It found that retailers establish CDP-A through developing new infrastructure development, through utilising existing facilities, or establishing partnerships with a third party (Wang et al., 2014). Other terminologies associated with CDP-A include “click-and-collect”, “pickup centre”, “click-and-mortar”, and “buy-online-pickup-in-store”. The literature showed that retailers establish CDP-U (or unattended reception) through independent RBs equipped with a docking mechanism, or shared RBs, whose locations range from private homes to public sites (e.g. petrol kiosks and train stations) accessible by multiple users (McLeod et al., 2006).

These CDP-A and CDP-U strategies are commonly adopted by multi/omnichannel retailers to exploit their existing store networks, to provide convenience to consumers through ancillary delivery services, and to expedite returns handling (Yrjölä, 2001). Moreover, the research showed that integrating online technologies with physical infrastructures enables retailers to achieve synergies in cost savings, improved brand differentiation, enhanced consumer trust, and market extension (Fernie et al., 2010). Studies have also investigated the cost advantage and operational efficiencies of using CDP-U over AHD and CDP-A (e.g. Wang et al., 2014). CDP-U reduces home delivery costs by up to 60 per cent (Punakivi et al., 2001), primarily by exploiting time window benefit (Kämäräinen et al., 2001).

Synthesising LML structural variables

The three dominant system dynamics identified (push-centric, pull-centric, and hybrid), reflect variations in LML distribution structures. By comparing these structures, we conceptualised a set of structural variables, whose variations can explain the structural forms of last-mile distribution. These structural variables consist of product source, geographical scope, mode of distribution, and number of nodes. These variables are expanded below:

• Product source refers to the location where products are stored when an order is accepted; it coincides with the start point of an LML network. It can be contextualised as a supply network member entity (manufacturer, distributor, or retailer). To illustrate, the computer manufacturer Dell (customisation services), online grocer Ocado (home delivery services), and the UK’s leading supermarket chain Tesco (click-and-collect services) source their products from manufacturer, distributor, and retailer sites, respectively.

• Geographical scope concerns the distance separating the start point (product source) and the end point (final consignee’s preferred destination point) of an LML network. An LML network can be classified as centrally based (e.g. Dell Services) or locally based (e.g. Tesco’s click-and-collect).

• Mode of distribution describes the delivery mode from the point where an order is fully fulfilled to the end point; it can be classified into three types: self-delivery (e.g. Tesco’s self-owned fleet for home deliveries), 3PL delivery including crowdsourcing (e.g. Dell Services), and consumer-pickup (e.g. Tesco’s click-and-collect services).

• Number of nodes concerns the operations in which products are “stationary”, residing in a facility for processing or storage. As opposed to nodes, links represent movements between nodes. There are two variations in respect to this variable: two-node and multiple-node. For example, a two-node structure can be found in Dell’s direct-to-consumer distribution channel, where computers are assembled and orders fulfilled at the factory prior to direct home delivery. In contrast, multiple-node structures are reflected in “in-transit merge” structure where an order comprising components sourced from multiple locations are assembled at a common node. As a case in point, when consumer order a computer processing unit (CPU) from Dell along with a Sony monitor, a parcel carrier would pick up the CPU from a Dell factory and the monitor from a Sony factory, then would merge the two into a single shipment at a hub prior to delivery (Chopra, 2003).

Synthesising LML contingency variables

The contingency variables exist in two modes: systematic development of LML contingency variables; and empirical analysis of particular variables or their interrelationships with other variables. To illustrate the former, Chopra (2003), Boyer and Hult (2005), and Gevaers et al. (2011) each produced their own lists of contingency variables. Regarding the latter, Boyer et al. (2009), for example, performed a simulation analysis of two independent variables (consumer density and delivery window size) and one dependent variable (delivery cost). The most common contingency-reported variables to have a direct impact on the structural form of LML distribution are listed below and commentary on them are summarised in Table III:

• Consumer geographical density: the number of consumers per unit area (Boyer and Hult, 2005; Boyer et al., 2009; Mangiaracina et al., 2015).

• Consumer physical convenience: the effort consumers exert to receive orders (Chopra, 2003; Harrington et al., 2016).

• Consumer time convenience: the time committed by consumers for the reception of orders. This variable fluctuates according to the structural form of last-mile distribution (Rabinovich and Bailey, 2004; Yuan and David, 2006).

• Demand volume: the number of products ordered by consumers relative to the distribution structure (Chopra, 2003; Boyer and Hult, 2005).

• Order response time: the time difference between order placement and order delivery (Kämäräinen et al., 2001; Mangiaracina et al., 2015).

• Order visibility: the ability of consumers to track their order from placement to delivery (Chopra, 2003; Harrington et al., 2016).

• Product availability and product variety: product availability is the probability of having products in stock when a consumer order arrives (Chopra, 2003; Yuan and David, 2006).

• Product variety is the number of unique products (or stock keeping units) offered to consumers (Punakivi et al., 2001; Punakivi and Saranen, 2001).

• Product customisability: the ability for products to be adapted to consumer specifications (Boyer and Hult, 2005).

• Product freshness: the time elapsed from the moment a product is fully manufactured to the moment when it arrives at the consumption point (Boyer and Hult, 2005).

• Product margin: the net income divided by revenue (Boyer and Hult, 2005; Campbell and Savelsbergh, 2005).

• Product returnability: the ease with which consumers can return unsatisfactory products (Chopra, 2003; Yuan and David, 2006).

• Service capacity: the ability of an LML system to provide the intended delivery service and to match consumer demand at any given point in time (Rabinovich and Bailey, 2004; Yuan and David, 2006).

Synthesising the relationship between LML structural and contingency variables

Following the identification of the structural and contingency variables, we draw a connection between structural and contingency variables and the relationships between them. This connection is achieved by capturing the differences between these variables and establishing the correlations between particular variations of structural variables and those of contingency variables. For example, Chopra (2003, p. 125) stated:

Firms that target customers who can tolerate a large response time require few locations that may be far from the customer and can focus on increasing the capacity of each location. On the other hand, firms that target customers who value short response times need to locate close to them.

This statement identifies the association between a structural variable, namely “geographical scope”, and a contingency variable, namely “order response time”. Within the literature, two variations emerged for each variable: centralised vs localised network for geographical scope and long vs short delivery period for order response time; i.e. centralised geographical scope corresponds to long response time, while localised scope is more responsive. As such, the findings demonstrate that by identifying connecting rationales and the variations at different levels for each variable, we can capture correlations between two sets of variables (i.e. structural and contingency). Continuing this procedure across relevant statements found in our corpus, Table IV summarises the outputs.

Reformulation from science-mode into design-mode knowledge

We adopted the approach by Romme (2003) to reformulate the descriptive knowledge (i.e. science-mode, developed in the previous section) into prescriptive (i.e. design-mode) knowledge so that the latter becomes more accessible to guide practitioners in their LML design thinking. This approach has previously been used to contextualise various design scenarios (e.g. Zott and Amit, 2007; Holloway et al., 2016; Busse et al., 2017). For example, Busse et al. (2017) employed a variant of the approach to investigate how buying firms facing low supply chain visibility can utilise their stakeholder network to identify salient supply chain sustainability risks.

Romme (2003) made a distinction between science and design modes of engaging in organisational research. The former is representational, descriptive, and analytic and it concerns the study of variances among variables over time and/or space. The latter is pragmatic, actionable, and exists in a more usable form in order to facilitate changing incumbent organisational systems and situations into the desired states. Furthermore, Romme (2003) proposed a framework exploiting synergies between science and design modes. He argued, suppose there is a science-mode proposition “if x, then y” and a design-mode proposition “to achieve O, do A”, to connect both modes, the former can be reformulated to the latter as follows:

• if necessary, redefine descriptive (properties of) variables into imperative ones (e.g. actions to be taken);

• redefine the probabilistic nature of a hypothesis into an action-oriented design proposition;

• add any missing context-specific conditions and variables (drawing on other research findings obtained in science- or design-mode); and

• in case of any interdependencies between hypotheses/propositions, formulate a set of propositions.

As an example, take the following science-mode statement:

[If order response time delivered by an LML network is short, then the geographical scope of the LML network should be localised].

We can reformulate this statement into the following design-mode statement:

[For an LML network to achieve short order response time, localise the geographical scope].

Following similar procedures, the science-mode knowledge describing the relationships between structural and contingency variables can be reformulated to the design-mode shown in Table V. Collectively, the resulting design-mode knowledge constitutes a set of design guidelines for LML practitioners.

Operational challenges in executing last-mile operations

The extant literature has focused on the planning aspect of LML, rather than exploring operational challenges. Consequently, research often takes a simplistic chain-level perspective of LML in order to develop simplistic design prescriptions for practitioners. While this approach seems suitable in the pre-digital era, it is inadequate to capture the complexities of last-mile operations in the omnichannel environment (Lim et al., 2017). The focus on LML nodes as solely unifunctional is also inadequate (Vanelslander et al., 2013). Not acknowledging the multi-functionality of individual nodes limits understanding of how this variant works.

To address the limitations of extant research, we propose extending the current research from addressing linear point-to-point LML “chains” (e.g. Chopra, 2003; Boyer and Hult, 2005) to also addressing the “networks” context, where multiple chains are intertwined and more widely practised in the industry. A study of LML systems using 3PL shared by multiple companies is an example of necessary future research. We also recommend future research to address the multi-functionality of individual nodes in an LML system. A study that addresses the ability of an LML node to simultaneously be a manufacturer and a distributor introduces more structural variance and needs to be theoretically addressed.

Additionally, existing literature typically focuses on comparing structural variants’ performance outcomes and their corresponding consumer and product attributes. However, we argue that such focus limits our understanding of how LML distribution structures interact as part of the broader omnichannel system. Accordingly, an avenue for future research would employ configuration perspectives (Miller, 1986; Lim and Srai, 2018) to complement the traditional reductionist approaches (e.g. Boyer et al., 2009) in order to more holistically examine LML models. Future studies could consider the structural interactions with relational governance of supply network entities, in order to promote information sharing and enhancing visibility, which are critical in omnichannel retailing (Lim et al., 2016).

Finally, while recent articles have started to examine the effects of online and offline channel integrations (e.g. Gallino and Moreno, 2014), limited contributions have been made to date to understand how retailers integrate their online and offline operations and resources to deliver a seamless experience for consumers (Piotrowicz and Cuthbertson, 2014; Hübner, Kuhn and Wollenburg, 2016). We propose revisiting the pull-centric system variants in the context of active consumer participation to understand the approaches retailers can use to attract consumers to their stores. In this regard, the subject can benefit from insightful case studies to advance our understanding of the challenges retailers face, as well as the operational processes retailers adopt to meet these challenges.

Intersection between last-mile operations and “sharing economy” models

With the exception of one paper (Wang et al., 2016), the majority of the extant literature discusses conventional LML models. Given the rapidly growing sharing economy that generates innovative business models (e.g. Airbnb, Uber, Amazon Prime Now) in several sectors (e.g. housing, transportation, and logistics, respectively) and exploits collaborative consumption (Hamari et al., 2016, p. 2047) and logistics (Savelsbergh and Van Woensel, 2016), there is an immense research scope at the intersection between LML and sharing economy models. First, we propose empirical studies to examine how retailers can effectively employ crowdsourcing models for the last-mile and to show how they can effectively integrate these models into their existing last-mile operations, such as combining in-store fulfilment through delivery using “Uber-type” solutions. This type of study is critical for understanding the impact of crowdsourcing models on retail operations and for promoting their adoption. Second, papers addressing omnichannel issues (Hübner, Kuhn and Wollenburg, 2016; Hübner, Wollenburg and Holzapfel, 2016; Ishfaq et al., 2016) are emerging. The emergence of new omnichannel distribution models demands theoretical development and the identification of new design variables. These models include on-demand delivery model (e.g. Instacart), distribution-as-a-service (e.g. Amazon, Ocado), “showroom” concept stores (e.g. Bonobos.com, Warby Parker), in-store digital walls (e.g. Adidas U.S. adiVerse), unmanned delivery (e.g. drones, ground robots), and additive printing (e.g. The UPS store 3D print). Increasingly, we also observe the growing convergence of roles and functions between online and traditional B&M retailers, which suggests new integrated LML models. These new roles and functions demand future research. Finally, while collaborative logistics enable the sharing of assets and capacities in order to increase utilisation and reduce freight, its success rests on developing a logistics ecosystem of relevant stakeholders (including institutions). Consequently, exciting research opportunities exist to explore new design variables that capture key stakeholders’ interests at various levels (Harrington et al., 2016).

Data harmonisation and analytics: collection and sharing platforms

The literature review revealed that, to date, there has been a tendency towards geographical-based studies and the use of simulated data. For example, this review reports studies based in Finland (Punakivi and Saranen, 2001), Scotland (McKinnon and Tallam, 2003), the USA (Boyer et al., 2009), England (McLeod et al., 2006), Germany (Wollenburg et al., 2017), and Brazil (Wanke, 2012), amongst others. While these studies contribute to generating a useful library of contexts, they are difficult to compare, given differences in geography and geographically based data collection and analysis methods. Moreover, the majority of the studies in this review (41.30 per cent) were based on modelling and simulated data with limited application to real-world data sets, which might suggest a lack of quality data sets. Simulated data limit the advancement of domain knowledge, thus the development of real-world data sets could significantly fuel progress. As such, more attention should be focused on developing data sets, e.g. through the use of transaction and consumer-level data, to gain insights into last-mile behaviours and to design more effective LML models.

Additionally, future studies should standardise data collection in order to address current trends in urbanisation and omnichannel retailing, which are changing retail landscapes and consumer shopping behaviours. This study recommends establishing a data collection framework to guide scholars in LML design, with scholars developing new competences in data mining analytics to exploit large-scale data sets.

Moving from prescriptive to predictive last-mile distribution design

Extant studies have derived correlations between variation of independent variables (e.g. order response time) and variation of dependent variables (e.g. degree of centralisation) to provide prescriptive solutions to the design of last-mile distribution structures. However, these relationships (both linear and non-linear) are often confounded by other factors due to the real-world complexities and they inherently face multicollinearity and endogeneity issues, including the omitted variable bias problem, which leads to biased conclusions. Moreover, model complexity increases as more variables are included, potentially causing overfitting. Given these complexities, researchers usually find immense challenges in untangling these relationships. In this regard, we offer several valuable future lines of research leveraging more advanced techniques for the design of last-mile distribution.

First, our review captured 13 contingency variables that influence the design of last-mile distribution. Future research could discuss other contingency variables and investigate the use of statistical machine and deep learning techniques to identify the most critical contingency variables and uncover hidden relationships to develop predictive models. Second, as urbanisation trends continue, more institutional attention is required on urban logistics focused on negative externalities (congestion and carbon emissions) driven by the intensification of urban freight. According to our review, there is insufficient attention paid to urban freight delivery, and we propose exploring archetyping of urban areas for the development of predictive models to guide the design of urban last-mile distribution systems.

Third, the developed design framework is based on the assumption that only one last-mile distribution structure may be adopted for a given scenario. As we observed in the omnichannel setting, it is common for retailers to concurrently operate multiple distribution structures. The interrelationships between the various structural combinations under the management of a single LML operator also present a potential future research direction.

Last, there is room for a combination of methods to more appropriately tackle the increasingly complicated and fragmented distribution networks in the omnichannel environment. Indeed, this research revealed only two papers in the corpus that have employed a mixed-method approach. Ishfaq et al. (2016) used case research and classification-tree analysis to understand the organisation of distribution processes in omnichannel supply networks, while Campbell and Savelsbergh (2006) combined analytical modelling with simulation to demonstrate the value of incentives in influencing consumer behaviour to reduce delivery costs.