Opportunity Management for Business-to-Business (B2B) Service Organizations: A Theory-Informed Decision Support Framework

Abstract

To meet sales targets with limited resources, business-to-business service firms must prioritize promising opportunities within large pipelines. Yet, both theory and practice indicate that such decisions often rely on intuition or ad hoc rules, resulting in suboptimal sales and operations planning. Drawing on the relationship management and organizational buying literature, we develop a theory-informed sales-operations framework that links buyer typology (e.g., new vs. rebid) and opportunity characteristics (e.g., size and relationship strength) to the firm’s bid and win decisions. Using archival data from a global on-site services provider encompassing 4,574 opportunities across 23 countries (2010–2021), we document a persistent tradeoff: while low-risk, relationship-based opportunities yield higher win probabilities, they are insufficient to achieve regional sales goals. We address this challenge through an ensemble machine-learning model that predicts win likelihood and a combinatorial optimization model that allocates bidding capacity strategically. The integrated framework improves predictive accuracy by 11% and could have increased realized sales by 21% while bidding on 38% fewer opportunities. Extensions incorporating stochastic programming and a Heckman-style two-stage correction enhance the framework’s robustness to uncertainty and data selection bias, providing managers with a rigorous, data-driven approach to opportunity management.

1. Introduction

Business-to-business (B2B) transactions constitute a substantial share of the U.S. economy, accounting for 51% of total revenues (Bonde and Bruno, 2019). Realizing B2B sales opportunities requires considerable effort due to the complex and high-value nature of transactions, extensive buying-center involvement, and heterogeneous customer requirements (Grewal et al., 2015). Approximately 50% of new B2B deals take more than seven months to complete (Miller Heiman Group, 2019), and the cost of a direct B2B sales representative is estimated at $1.9K per business day. ¹ Given these high costs, firms must strategically allocate their resources in pursuing sales opportunities.

Sales opportunity management is central to this strategic allocation. It involves prioritizing and optimizing opportunities within the sales pipeline to maximize revenue and resource efficiency (Söhnchen and Albers, 2010). Effective opportunity management allows firms to concentrate efforts on high-potential leads, generate reliable sales forecasts to support strategic planning, and enhance overall conversion rates (Mittal and Sridhar, 2021). However, two interrelated challenges complicate this task. First, the inherent complexity and uncertainty of B2B sales processes make it difficult to accurately assess the likelihood of success for each opportunity. As a result, firms often fall back on heuristic stop/go decision rules that introduce systematic biases, such as overweighting mid-sized deals and underestimating true win probabilities by as much as 100% (Xu et al., 2022). Second, and equally important, the empirical data available to managers and researchers are inherently biased because firms only observe and record opportunities they choose to pursue, resulting in a non-random sample that complicates both predictive modeling and prescriptive optimization. This selection bias represents a core empirical challenge for data-driven decision-making in B2B contexts. Together, these behavioral and data-driven biases help explain why B2B sales opportunity conversion rates remain low, typically ranging from 1% to 5% (D’Haen and Van den Poel, 2013), with 63% of sales managers reporting dissatisfaction with their firm’s ability to manage such opportunities effectively (Miller, 2019).

Extant research across three domains provides important and partial insights into the challenges of sales opportunity management. First, the literature on salesperson judgment and decision-making reveals the presence of cognitive biases in how sales opportunities are prioritized (Hall et al., 2015; Sabnis et al., 2013; Xu et al., 2022). These biases can lead to suboptimal allocation of selling effort and distort the assessment of opportunity quality. Second, research in relationship marketing consistently finds that stronger buyer–seller relationships improve the likelihood of winning bids (Dwyer et al., 1987). Despite this, relational variables are rarely incorporated into predictive models used in sales opportunity management. Third, a growing body of work in marketing and operations research has employed machine learning (ML) techniques to estimate bid success probabilities based on observable buyer characteristics, such as firm size and industry classification (D’Haen and Van den Poel, 2013; Jahromi et al., 2014; Mortensen et al., 2019). However, these models are often atheoretical, typically assuming unbiased data and stopping short of prescribing how firms should act on predictive insights.

This study integrates these three research streams to develop a unified decision support framework for sales opportunity management. We begin by proposing a theoretical structure that identifies key bid-specific attributes and market-level factors influencing the outcomes of sales opportunities. Grounded in this framework, we develop an ensemble ML model to estimate the probability of winning a bid using historical sales opportunity data. We then embed this predictive model into a prescriptive optimization tool that enables firms to strategically prioritize and allocate resources across competing opportunities under strategic capacity constraints—forming the core of our “predict + optimize” approach. While this model provides practical guidance for data-driven bidding decisions, we also propose an extension using a Heckman-style two-stage model to correct for potential selection bias in settings where observed data are limited to previously bid opportunities. Finally, to enhance practical applicability, we introduce a further extension that incorporates uncertainty in opportunity characteristics and resource requirements through stochastic programming, supporting more resilient decision-making under uncertainty.

To empirically validate our modeling framework, we leverage bid-level data from a global on-site services provider. The dataset includes 4,574 B2B sales opportunities spanning 2010 to 2021, each representing a competitive bidding situation between the focal firm and rival bidders. We construct an ensemble ML model to estimate the focal firm’s probability of winning a bid. Specifically, we stack five base learners—random forest (RF), generalized boosted machines (GBMs), logistic regression, Naïve Bayes classifier, and support vector machines (SVMs)—using an RF as the meta-learner to aggregate predictions. These models represent a diverse set of algorithmic classes (tree-based, probabilistic, linear, and margin-based), ensuring predictive robustness and breadth. The ensemble significantly outperforms individual learners both in-sample and out-of-sample evaluations, improving predictive accuracy by 11% over the best-performing single model. To interpret model outputs, we use partial dependence plots (PDPs) (Friedman, 2001), which highlight key marginal effects: larger project size lowers win likelihood, while stronger buyer–seller relationships increase it—underscoring the importance of relational factors in B2B sales.

We embed these predicted win probabilities into a prescriptive optimization model that identifies the optimal subset of opportunities to pursue given strategic capacity constraints. This integrated approach enables firms to make resource-efficient bidding decisions based on predicted outcomes. In a retrospective simulation, we find that the focal firm could have increased total sales by 21% while bidding on 38% fewer opportunities, illustrating the substantial efficiency gains achievable through a data-driven “predict + optimize” framework.

This study makes three primary contributions to operations theory and practice. First, we contribute to the literature on sales opportunity management (D’Haen and Van den Poel, 2013; Söhnchen and Albers, 2010) by empirically integrating well-established constructs from the B2B buyer–seller relationship literature (Andreassen and Lervik, 1999; Dwyer et al., 1987) into a normative decision support framework for service operations. This framework enables firms to systematically incorporate relational factors—often overlooked in algorithmic models—into their bidding decisions. Our findings also reinforce prior evidence that selling organizations tend to avoid larger deals and disproportionately favor opportunities involving existing relationships. Recognizing and correcting for such behavioral biases is essential when designing prescriptive tools.

Second, we contribute to the predictive analytics literature in operations management (D’Haen and Van den Poel, 2013; Mortensen et al., 2019; Rezazadeh, 2020) by demonstrating a novel application of ML in B2B sales opportunity management. ² Our setting responds directly to recent calls to expand the use of big data analytics in underexplored domains such as revenue management and marketing (Choi et al., 2018). In particular, we address the gap identified by Habel et al. (2023), who argue that existing research has not adequately explored how firms should develop and operationalize decision rules based on predictive models in sales contexts. We further contribute to the literature on optimization in sales management (Table EC.1) by embedding ML-based win probabilities into a prescriptive model for opportunity prioritization. This integrated decision support framework allows firms to make informed bidding decisions under strategic constraints. Importantly, we extend this framework to correct for non-random sampling in historical data using a Heckman-style two-stage model. This extension is particularly valuable in settings where data on lost or unbid opportunities are unavailable, as it improves both prediction and prescription quality. We also present a stochastic programming extension that accommodates uncertainty in opportunity characteristics, making the model more flexible and robust to dynamic sales environments.

Finally, we contribute to the decision support systems literature (Aggarwal et al., 1992; Blackmon et al., 2021; Kesavan and Kushwaha, 2020; Ramirez-Nafarrate et al., 2015; Silva-Risso and Ionova, 2008) by presenting a real-world application of integrated predictive and prescriptive analytics in the B2B service sales domain. Building on prior research that combines empirical model predictions with optimization methods to guide operational decisions (Kanuri et al., 2018; Naik et al., 1998, 2005), our framework demonstrates how data-driven decision tools can be effectively implemented in complex sales environments. Our results indicate that prioritizing opportunities based on model-informed recommendations can lead to substantial performance gains. More broadly, the proposed framework serves as a practical decision support tool for senior commercial leaders, including Chief Operations Officers, Chief Commercial Officers, and Chief Executive Officers, who rely on accurate forecasting and resource allocation to manage opportunity pipelines and drive strategic growth.

2. Conceptual Framework for Sales Opportunity Management

Building on the literature insights discussed above, we next develop a conceptual framework for B2B opportunity management. This framework serves as the theoretical foundation for our empirical models. In this section, we define key variables that influence opportunity pursuit and outcomes, drawing from prior literature. These include market characteristics, competitive dynamics, historical performance, and buyer–seller relationships.

B2B sales opportunities often involve long buying cycles and multiple decision-makers, requiring significant time and effort from sales teams. Due to resource constraints, firms cannot pursue every opportunity and must strike a balance between revenue goals and margin targets across regions and product categories. As such, identifying high-potential opportunities is critical for effective sales and operations planning.

Research suggests that sales managers often rely on heuristics and subjective judgment when making pursuit decisions (Scherm and Scherm, 2021; Söhnchen and Albers, 2010). This can lead to an overestimation of win probabilities and missed sales targets. Moreover, salespeople tend to prefer opportunities perceived as more certain, even when riskier alternatives may offer higher returns.

We focus on three core attributes known to shape both the decision to pursue and the likelihood of winning a B2B opportunity: opportunity size, relationship strength, and bid/competitive characteristics.

2.1. Size of the Opportunity

Research on salesperson judgment has documented the effect of opportunity size on pursuit decisions. Using a scenario experiment, Xu et al. (2022) demonstrate an inverted U-shaped relationship between opportunity size and willingness to pursue. As opportunity size increases from low to medium, pursuit willingness rises, but as size increases further, pursuit willingness declines. This effect is mediated by resource slack, indicating that salespeople balance cost–benefit considerations at the portfolio level. A similar inverted U-shaped relationship is observed between opportunity size and actual win rates.

This suggests that extremely large projects may deter bids due to perceived risk and resource constraints. In our empirical analysis, we examine whether such a pattern is reflected in the actual outcomes.

2.2. Relationship With the Customer

The relationship marketing literature emphasizes the financial value of strong buyer–seller relationships. Dwyer et al. (1987) and subsequent work show that relational exchanges enhance customer commitment and repurchase likelihood (Bosukonda et al., 2020). However, salespeople often misjudge relationship strength, relying on surface cues such as frequency or duration of contact (Homburg et al., 2014).

To operationalize relationship strength, we propose a typology (Table 1) that classifies direct and indirect connections between the seller and customer. This complements the Buy-Grid framework (Robinson et al., 1967), adding a relationship-based lens to sales decision-making.

Table 1.

Relationship typology.

Type of relationship	Description
Direct relationship	The focal firm has a direct relationship with the customer at the focal site.
Indirect relationship and new location	The focal firm has a relationship with the customer’s parent company at other locations, and the focal site is either a new location or a self-serviced location.
Indirect relationship	The focal firm has a relationship with the customer’s parent company at other locations, and the focal site is currently serviced by competitors.
No relationship and new location	The focal firm has no relationship with the customer’s parent company at other locations, and the focal site is either a new location or a self-serviced location.
No relationship	The focal firm has no relationship with the customer’s parent company at other locations, and the focal site is serviced by competitors.

Notes. The focal firm refers to the bidding firm making the decision to pursue the opportunity. The focal site is the specific customer location under consideration. A new location (also referred to as a greenfield site) is a customer site that has not previously been serviced. A self-serviced location is one where the customer currently handles the service in-house without using external vendors.

In short, this typology distinguishes between: (a) direct relationships at the focal site; (b) indirect relationships via the customer’s parent company, with further distinctions based on whether the site is new or competitor-served; and (c) no existing relationship, again split into new or competitor-served sites. These categories play a central role in our modeling.

2.3. Bid Characteristics and Competitive Dynamics

Beyond opportunity size and relationships, bid characteristics also impact both the seller’s decision to pursue an opportunity and the buyer’s decision to award the contract. Factors such as pricing, delivery timelines, product differentiation, and competitive landscape shape sales outcomes.

In addition to the variables outlined above, other factors, such as the incentive structure for sales teams (Claro et al., 2023) and internal organizational priorities (Cui et al., 2020), can also significantly impact sales opportunity management decisions. These additional variables could potentially enhance the predictive efficiency of models used in sales opportunity management. However, due to constraints on available data, we restrict our model to the variables explicitly included in our framework.

While we lack data on factors such as detailed pricing or salesforce incentives, our framework captures the most salient predictors identified in prior research (size, relationship strength, etc.). We focus on these core variables, and note that managers and practitioners looking to improve predictive accuracy can incorporate these additional variables into their models, tailoring their decision-making frameworks to better align with their organizational context and strategic priorities.

As depicted in Figure 1, we posit that opportunity characteristics, particularly size, relationship strength, and bid context, influence two key outcomes: (1) the firm’s decision to pursue the bid, and (2) the likelihood of winning, conditional on pursuit. We expect strong relationships and moderately sized opportunities to increase both, while extremely large or unfamiliar opportunities may dampen them.

Figure 1.

A theoretical framework of factors affecting opportunity pursuit and winning the opportunity.

Next, we develop a decision support model, consisting of a predictive (ML) component and a prescriptive (optimization) component, grounded in the key factors outlined here.

3. Modeling Framework

The objective of this section is to develop a decision support system that enables firms to optimize sales opportunity pursuit. In our context, this involves two decision stages: (1) predicting the likelihood of winning each opportunity using an ensemble ML model, and (2) optimally allocating bidding capacity using an optimization model informed by those predictions. Essentially, the ML predictions feed into the optimization model, which guides bidding decisions under capacity constraints.

3.1. Ensemble ML Model

We employ an ensemble-learning approach to predict the likelihood of winning a sales opportunity. The model integrates multiple ML techniques to enhance predictive accuracy, accounting for complex interactions between market conditions, customer attributes, and competitive factors.

Ensemble approaches are widely used to improve predictive performance by combining multiple ML models. More specifically, we use an ensemble-stacking approach, where multiple ML models are trained in parallel, and a meta-learner is trained to predict the outcome using the predictions of the earlier models. Figure 2 presents a visualization of the ensemble-stacking approach used in our study. It is important to note that our predictive model estimates the probability of winning conditional on the focal firm having submitted a bid, that is, $P(\text{win}\mid \text{bid})$ . The process by which the firm decides whether or not to bid on an opportunity is not modeled in our current framework. This distinction is important because applying these predictions to opportunities that were not historically pursued may introduce bias if the bid/no-bid decision is systematically related to the likelihood of winning. To explicitly address this data limitation, we implement a Heckman-style two-stage correction model (detailed in Section 5) that accounts for the non-random nature of observed bids. This approach provides a bias-corrected estimate of win likelihood, offering managers a conservative benchmark against which to compare the baseline ML predictions.

Figure 2.

Ensemble machine-learning model.

The ensemble-learning model includes the following base learners:

RF: The RF algorithm, developed by Breiman (2001), constructs multiple decision trees using bootstrapped samples of training data and a randomly selected subset of features. Predictions are aggregated through majority voting in classification tasks. We implement the RF algorithm using the R package described in Kuhn et al. (2020).

GBMs: The GBM algorithm iteratively builds an ensemble of decision trees, where each successive tree corrects residual errors from previous iterations. This boosting process enhances prediction accuracy. We employ the implementation provided by Kuhn et al. (2020) in R.

Logistic Regression: Logistic regression models the log-odds of an instance belonging to a particular class as a linear function of predictor variables. Estimated coefficients are then used to compute class probability estimates.

Naïve Bayes: The Naïve Bayes classifier applies Bayes’ theorem under the assumption of conditional independence among predictor variables. Despite this simplifying assumption, it performs effectively in various classification settings. We use the implementation provided by Meyer et al. (2015) in R.

SVM: The SVM classifier identifies a hyperplane that maximizes the margin between classes in a high-dimensional space, making it particularly suitable for handling complex, nonlinearly separable data. We implement the SVM model using the R package introduced by Karatzoglou et al. (2004).

We selected these five base models to reflect diverse algorithmic paradigms: RF and GBM represent tree-based ensemble methods; logistic regression is a linear classifier; Naïve Bayes is a probabilistic model; and SVM is a kernel-based method. This diversity enhances the ensemble’s ability to capture complex patterns in the data by leveraging the complementary strengths of different learning philosophies.

RF Meta-Learner: The final step in our ensemble model involves using an RF model as a meta-learner. The meta-learner aggregates the predictions from the base models and generates the final prediction of the likelihood of winning a sales opportunity.

Mathematically, let $M_1,M_2,\dots ,M_k$ be $k$ base models that produce probability estimates $p_1,p_2,\dots ,p_k$ for a given opportunity. The meta-learner learns a function $f(\mathbf{p})$ that maps these predictions to a final probability estimate:

$$\hat{y}=f(p_1,p_2,\dots ,p_k)=\sum _{i=1}^k{w}_i{p}_i,$$

where $w_i$ are learned weights that determine the contribution of each base model. These weights are computed using a separate RF model trained on the training set, similar to running a second-stage regression. The input variables for this secondary RF model are the probability estimates $p_1,p_2,\dots ,p_k$ from the base models, and the output variable is the actual observed outcome (win/loss) of the sales opportunity. The RF meta-learner learns to assign higher weights to the base models that perform better on the training set while reducing the influence of weaker models.

For example, suppose three base models (RF, logistic regression, and Naïve Bayes) predict the likelihood of winning a sales opportunity as follows: $M_1$ (RF) predicts 0.65; $M_2$ (logistic regression) predicts 0.60; $M_3$ (Naïve Bayes) predicts 0.55

If the meta-learner assigns weights of $w_1=0.5$ , $w_2=0.3$ , and $w_3=0.2$ , the final predicted probability of winning the opportunity is:

$$\hat{y}=(0.5\times 0.65)+(0.3\times 0.60)+(0.2\times 0.55)=\mathrm{0.605.}$$

This approach effectively combines multiple models’ strengths, improving prediction robustness and accuracy.

We use an RF for the meta-learner, which in practice learns a nonlinear function of the base predictions. While the actual function is more complex than a weighted average, this formulation conveys the intuition that each base model contributes to the final prediction based on its learned importance.

The meta-learning approach offers several key strengths:

Enhanced Predictive Performance: By leveraging diverse base models, the meta-learner can capture complex patterns and relationships in the data that individual models might miss (Dong et al., 2020).

Robustness to Overfitting: The ensemble nature of the meta-learner helps to mitigate overfitting, as it combines predictions from multiple models, reducing the impact of any single model’s biases (Sollich and Krogh, 1995).

Adaptability: The meta-learner can dynamically adjust the importance of each base model’s predictions based on their performance across different types of sales opportunities (Sollich and Krogh, 1995).

Improved Generalization: By learning from the predictions of multiple models, the meta-learner often generalizes better to new, unseen data compared to individual models (Ahmed et al., 2013).

Handling of Complex Decision Boundaries: The combination of diverse base models allows the meta-learner to capture intricate decision boundaries that may be challenging for any single model to represent.

We evaluate predictive performance using accuracy, sensitivity (recall), and specificity on a held-out test set (20% of the data). A five-fold cross-validation procedure is applied during training for hyperparameter tuning and to mitigate overfitting.

3.2. Optimization Model: Optimizing Opportunity Pursuit Portfolio

The optimization model is designed to allocate available capacity for competing projects to be bid on, with the objective of maximizing revenue for a firm. The predictive ML model developed earlier provides input parameters, such as the win probabilities of projects required for the optimization model. We formulate the model for a single region $\ell$ , consistent with the firm’s region-specific capacity planning structure. The model can be executed independently for each region. The model can be run for each region separately.

3.2.1. Optimization Model: Single Region Service Scenario

We define the following key parameters and decision variables for a single region $\ell$ :

$t$ —Current period $t$ denoting a planning horizon.

$m$ —Number of project types.

$n_{j,t}^{\ell }$ —Number of projects of type $j$ available in region $\ell$ during period $t$ .

$r_{i,j,t}$ —Revenue from project $i$ of type $j$ in period $t$ .

$p_{i,j,t}$ —Predicted probability of winning project $i$ of type $j$ in period $t$ .

$a_{i,j,t}$ —Capacity required to complete project $i$ of type $j$ in period $t$ .

$C_{j,t}^{\ell }$ —Available capacity for project type $j$ in region $\ell$ at time $t$ .

The decision variables are defined as:

$X_{i,j,t}$ —Binary variable; 1 if the firm bids on project $i$ of type $j$ , 0 otherwise.

$Y_{i,j,t}$ —Binary variable; 1 if a project of type $m$ is assigned to a team handling project type $j=1,2,\dots ,m-1$ , 0 otherwise.

$\mathrm{\Psi }$ —Expected total revenue during the planning horizon.

The optimization problem is formulated as follows:

$$max\mathrm{\Psi }=\sum _{j=1}^m\sum _{i=1}^{n_{j,t}^{\ell }}{r}_{i,j,t}\cdot p_{i,j,t}\cdot X_{i,j,t},$$

(1)

subject to:

Constraints (2) enforce the capacity constraint for period $t$ in region $\ell$ .

$$\sum _{i=1}^{n_{j,t}^{\ell }}{a}_{i,j,t}\cdot X_{i,j,t}+\sum _{i=1}^{n_{m,t}^{\ell }}{a}_{i,m,t}\cdot Y_{i,j,t}\le C_{j,t}^{\ell },j=1,2,\dots ,m-1.$$

(2)

Constraints (3) enforce that if the project of type $m$ is bid for period $t$ then it should be assigned to a project team handling any type, $j=1,2,\dots ,m-1$ .

$$\sum _{j=1}^{m-1}{Y}_{i,j,t}=X_{i,m,t},i=1,2,\dots ,n_{m,t}^{\ell }.$$

(3)

Constraints (4) indicate the decision variables assuming value $X_{i,j,t}=1$ if the focal firm bids on project $i$ of type $j$ in period $t$ ; otherwise, $X_{i,j,t}=0$ , where, $j=1,2,\dots ,m-1$ . Constraints (5) indicate the decision variables $Y_{i,j,t}=1$ if the focal firm bids on project $i$ of type $m$ in time period $t$ then it should be assigned to one of the sales team handling project type $j$ , $j=1,2,\dots ,m-1$ ; otherwise $Y_{i,j,t}=0$ .

$$\begin{array}{rl}{X}_{i,j,t}\in \{0,1\},& i=1,2,\dots ,n_{j,t}^{\ell };j=1,2,\dots ,m-1.\end{array}$$

(4)

$$\begin{array}{rl}{Y}_{i,j,t}\in \{0,1\},& i=1,2,\dots ,n_{m,t}^{\ell };j=1,2,\dots ,m-1.\end{array}$$

(5)

This model ensures optimal allocation of sales capacity, maximizing revenue while accounting for constraints on available resources. This framework can be applied by any company that aims to optimize its bidding decisions for sales opportunities across various regions and product categories. In essence, the model embodies a risk–reward balance: it allows some pursuit of risky (large or new) opportunities, but ensures capacity limits are not violated, and overall expected revenue is maximized.

To further illustrate how the optimization framework can be adapted to decision-making under uncertainty, we also develop a stochastic programming extension using a sample-average-approximation (SAA) approach. This extension considers uncertainty in project revenues and capacity requirements and demonstrates how managers can stress-test allocation decisions when key parameters are uncertain. Because the primary focus of the paper is the baseline data-driven “predict + optimize” framework, the full mathematical formulation and implementation of this SAA model are presented in the Electronic Companion (EC.8).

We next outline the design criteria that informed the development of our opportunity management framework, emphasizing principles that enhance its predictive accuracy, strategic flexibility, and ease of application.

4. Criteria for Developing a Model for Opportunity Management

An effective opportunity management model should be grounded in principles that enhance predictive accuracy, decision-making efficiency, and strategic flexibility. Our modeling framework integrates ML and optimization techniques to create a system that not only predicts the likelihood of winning opportunities, but also optimally allocates resources. The following criteria ensure the model’s practicality, adaptability, and effectiveness in guiding firms’ sales pursuit strategies:

Parsimonious Design. The model should balance simplicity and effectiveness, avoiding unnecessary complexity. We intentionally limited our variable set to theoretically grounded and practically observable features, rather than including a wide array of loosely relevant predictors. This design choice strengthens the theoretical contribution and makes the model easier to apply across firms. The ensemble model further streamlines prediction by selecting the most informative features, while the optimization component simplifies decision-making through a binary bid/no-bid structure rather than more granular or harder-to-implement allocation rules.

Incorporation of Interaction Effects. The model should account for nonlinear relationships between factors influencing bidding and buying decisions. Our ensemble-learning approach naturally captures complex interactions by combining multiple predictive models, most of which do not impose functional form restrictions. This allows for greater flexibility in identifying patterns within the data. The stacking methodology ensures that interdependencies between opportunity attributes, market conditions, and competitive factors are effectively modeled. Indeed, as noted later, we observe interaction effects in our data, such as the interplay between opportunity size and relationship strength on win likelihood, which our model architecture is designed to capture.

High Predictive Validity. The model must demonstrate strong predictive performance using empirical data. Our ensemble-stacking approach enhances predictive accuracy by leveraging the strengths of multiple base learners, with an RF meta-learner aggregating their outputs to improve generalizability. To validate model performance, we employ rigorous cross-validation techniques, ensuring that the model maintains high predictive validity across different types of sales opportunities while avoiding overfitting.

Customizability for Firms’ Objectives. Firms should be able to tailor the model to optimize different objectives, such as revenue maximization, profit optimization, or region-specific goals. Our modeling framework provides flexibility through an adaptable optimization structure that allows firms to adjust constraints and objectives based on their strategic priorities. Additionally, the stochastic optimization extension enables firms to incorporate uncertainty into their decision-making processes, making the model applicable across various industries and market conditions.

By adhering to these criteria, our opportunity management model ensures predictive accuracy, adaptability, and effective resource allocation, supporting firms in making data-driven sales pursuit decisions. In the following sections, we demonstrate an empirical application of this model at ServiceCo, evaluating its performance and implications.

5. Empirical Application

This section applies our proposed framework to the sales opportunity management process of a leading B2B on-site services provider (hereinafter “ServiceCo”) with global operations. We first describe the institutional context and the customer relationship management (CRM) dataset used for the analysis. Next, we provide model-free evidence highlighting key patterns in ServiceCo’s bidding and winning behaviors. We then develop and evaluate an ML model to predict ServiceCo’s win likelihood across opportunities, followed by an optimization model that recommends an improved bidding strategy. Together, these analyses demonstrate how predictive analytics and optimization can enhance ServiceCo’s opportunity selection and revenue outcomes.

5.1. Institutional Context and Data

ServiceCo follows a structured multistage internal process to evaluate which opportunities to pursue. First, regional sales managers conduct an initial qualification of each lead based on client fit and service feasibility. Second, opportunities undergo conditional approval that accounts for projected margins, operational complexity, and client profile. Finally, a formal go/no-go decision is made by senior executives. ServiceCo competes primarily with four major players globally and several local service providers. ServiceCo’s sales opportunities can be categorized into two broad types: rebid opportunities and new-bid opportunities, each representing distinct competitive and strategic dynamics. A rebid opportunity arises when ServiceCo is already providing services at a specific site, and the contract for those services is up for renewal or recompetition. In such cases, ServiceCo competes to retain the business against both its existing contract terms and competing bids from other service providers. Rebid scenarios can be advantageous for ServiceCo because of its existing relationship, operational familiarity, and historical service performance at the site. However, competitors may attempt to undercut ServiceCo’s pricing or offer enhanced services to win the contract. In contrast, new-bid opportunities arise when ServiceCo is not the current provider, and these can be further classified into three subcategories. First, some new-bid opportunities involve sites where a competitor is the incumbent service provider. In these cases, ServiceCo must displace the competitor by demonstrating superior value, service quality, or cost-effectiveness. Such opportunities are particularly competitive, as the existing provider may leverage its established relationship and service history to retain the contract. Second, some sites are currently self-managed by the customer instead of being outsourced to a third-party provider. ServiceCo must convince such customers that outsourcing will result in cost savings, efficiency improvements, service enhancements, or risk reduction. Lastly, a subset of new-bid opportunities emerges from customers opening new locations that require service providers. These opportunities present a unique situation where there is no incumbent provider, and ServiceCo competes on equal footing with other bidders.

Most of ServiceCo’s clients operate multiple sites across different geographic locations, creating opportunities for cross-site relationships between ServiceCo and its customers. Even if ServiceCo is not currently servicing a particular site, it may already have an established relationship with the parent company or another subsidiary at a different location. These existing relationships can influence bidding decisions, as prior engagements may provide insights into the customer’s preferences, procurement processes, and service expectations. To leverage this dynamic, our relationship typology categorizes customer interactions based on prior engagement levels. This typology captures ServiceCo’s strength in its connection with a given client and its impact on winning. For instance, customers with long-standing engagements at other sites may be more inclined to consider ServiceCo for new contracts, while those with no prior relationship might require a more aggressive sales approach.

The CRM system data provide a comprehensive snapshot of 4,574 sales opportunities recorded between fiscal years 2010 and 2021, capturing both successful and unsuccessful engagements along with historical customer account interactions. Of these opportunities, 550 opportunities (12.04%) were withdrawn by customers, meaning the sales process was terminated before reaching the bidding stage. This could occur due to budget constraints, strategic shifts, or operational changes on the customer’s end. A withdrawal does not necessarily indicate a failure on ServiceCo’s part, as it reflects a customer-driven decision rather than a competitive loss. In total, 1,203 opportunities (26.03%) were not pursued by ServiceCo, indicating strategic decisions to walk away, likely influenced by factors such as low profitability, high competition, or operational constraints. The remaining 2,821 opportunities (61.67%) resulted in submitted bids, where ServiceCo actively competed for the contract. ServiceCo’s walkaway rate, calculated as the percentage of no-bid cases among nonwithdrawn opportunities, stands at 30%. Figure 3 provides a distribution of old providers and new providers (who won the bid) across different opportunities ServiceCo was involved in. Rebid opportunities are identified with ServiceCo as the old provider.

Figure 3.

Distribution of old service provider and new service provider across all opportunities.

ServiceCo’s opportunities are grouped into different Strategic Segments, with the majority coming from the mining and energy sectors. ServiceCo offers services in three primary domains (referred to as service levels), and some customers award contracts that combine services across different service levels (referred to as integrated services). Most of ServiceCo’s customers have opportunities at multiple locations, with each location operating as an individual entity.

In our dataset, we observe the parent account associated with each opportunity and use this information to develop a variable that captures the relationship typology discussed earlier. Additionally, ServiceCo designates some of these customers as strategic accounts, and we incorporate this classification into our modeling framework.

For the 4,024 opportunities that were not withdrawn by customers, Figure 4 illustrates the number of opportunities per year, while Table 2 provides a description of each variable along with the corresponding descriptive statistics. We observe a sudden increase in the number of opportunities from 2017 onward. However, this increase is not a result of a strategic decision to pursue more business, but rather due to a company-wide initiative to adopt the CRM system more rigorously to improve data documentation. To ensure that our findings are not biased by differences in data recording over time, we replicate our analysis using only data from 2017 to 2021. The results remain consistent, and we provide these as a robustness check in the EC.3.

Figure 4.

Number of opportunities in each year.

Table 2.

Summary statistics.

Variable	Definition	Mean (SD)	Min	Max
Dependent variable
Customer withdraws the opportunity	Yes = 1, otherwise = 0	12.02%*	0	1
ServiceCo bids for the opportunity	Yes = 1, otherwise = 0	70.10%	0	1
ServiceCo wins the opportunity	Yes = 1, otherwise = 0	48.71%**	0	1
Independent variables
Size of the project
Revenue	Expected revenue from the opportunity in millions of dollars	1.57 (3.73)	0	66.05
Relationship type
Direct relationship with buyer on site	Yes = 1, otherwise = 0	22.94%	0	1
Indirect relationship at old location	Yes = 1, otherwise = 0	23.14%	0	1
Indirect relationship at new location	Yes = 1, otherwise = 0	6.21%	0	1
No relationship at new location	Yes = 1, otherwise = 0	12.10%	0	1
No relationship at old location	Yes = 1, otherwise = 0	35.61%	0	1
Strategic segment
Energy	Energy = 1, otherwise = 0	72.06%	0	1
Mining	Mining = 1, otherwise = 0	27.00%	0	1
Other segments	Other segments = 1, otherwise = 0	0.94%	0	1
Service level
1	Service level 1 = 1, otherwise = 0	9.82%	0	1
2	Service level 2 = 1, otherwise = 0	25.72%	0	1
3	Service level 3 = 1, otherwise = 0	36.98%	0	1
4 (integrated services)	Service level 4 = 1, otherwise = 0	27.49%	0	1
Strategic account
Yes	Customer is a strategic account = 1, otherwise = 0	16.55%	0	1
Region
Asia Pacific & Africa	Yes = 1, otherwise = 0	18.23%	0	1
Europe	Yes = 1, otherwise = 0	14.49%	0	1
North America	Yes = 1, otherwise = 0	32.75%	0	1
South America	Yes = 1, otherwise = 0	34.52%	0	1

* $N=4,574$ ; ** $N=2,821.$

5.2. Model-Free Evidence

In this section, we present model-free evidence to examine how ServiceCo’s bidding behavior and success rates vary across different types of opportunities. Specifically, we analyze three key metrics: walkaway rate, defined as the percentage of opportunities where ServiceCo chose not to bid among all opportunities not withdrawn by customers; absolute win rate, which represents the percentage of opportunities ServiceCo won among all nonwithdrawn opportunities; and win rate conditional on bidding, which captures the percentage of opportunities won among those where ServiceCo actively submitted a bid. By comparing these metrics across different opportunity characteristics, we provide initial, descriptive insights into how factors such as opportunity size, relationship type, service level, strategic segment, and customer status (strategic vs. nonstrategic accounts) influence ServiceCo’s bidding decisions and success rates.

To examine the impact of opportunity size, we group opportunities into low-revenue (first quartile), medium-revenue (second and third quartiles), and high-revenue (fourth quartile) categories. Panel A in Figure 5 illustrates how walkaway rate, absolute win rate, and win rate conditional on bidding vary across these groups. The walkaway rate is significantly higher for low-revenue opportunities compared to high-revenue opportunities ( $p<.01$ ), and similarly, the difference between medium-revenue and high-revenue opportunities is also statistically significant ( $p<.05$ ). The absolute win rate is significantly lower for low-revenue opportunities compared to high-revenue opportunities ( $p<.01$ ), and this pattern holds when comparing medium-revenue and high-revenue opportunities ( $p<.01$ ). Additionally, the win rate conditional on bidding is significantly lower for medium-revenue opportunities compared to high-revenue opportunities ( $p<.01$ ). These findings suggest that ServiceCo is more likely to walk away from low-revenue projects, and when it does bid on them, its chances of winning are lower compared to higher-revenue opportunities.

Figure 5.

Model-free evidence.

Panel B in Figure 5 examines how these metrics vary based on relationship type. The walkaway rate declines as the strength of the relationship between ServiceCo and the customer increases, with differences in walkaway rates across relationship types being statistically significant ( $p<.05$ ). The absolute win rate is significantly different across relationship types ( $p<.01$ ), except for the comparison between the indirect relationship group and the no-relationship-at-new-location group. Similarly, the win rate conditional on bidding differs significantly across relationship types ( $p<.01$ ), except for the difference between the direct relationship group and the no-relationship-at-new-location group. EC.4 shows the walkaway rate, absolute win rate, and win rate conditional on bidding among opportunities with different relationship types within low-revenue, medium-revenue, and high-revenue groups. These findings suggest that ServiceCo’s bidding success is closely tied to the strength of its relationships with customers, indicating that existing relationships provide a competitive advantage and that the firm prioritizes opportunities where it has established ties.

Panel C in Figure 5 compares walkaway rate, absolute win rate, and win rate conditional on bidding across different service levels. The walkaway rate varies significantly between service levels ( $p<0.1$ ), except for the differences between service levels 1 and 3, and service levels 3 and 4. The absolute win rate differs significantly between service levels 3 and 4 ( $p<.01$ ). Similarly, differences in win rate conditional on bidding are statistically significant ( $p<.01$ ) between service levels 2 and 4, and between service levels 3 and 4. These results indicate that ServiceCo’s likelihood of walking away, as well as its win probability, varies based on the complexity and nature of the service being offered, suggesting that certain service levels may be more competitive or better aligned with ServiceCo’s capabilities.

Panel D in Figure 5 shows the walkaway rate, absolute win rate, and win rate conditional on bidding among opportunities of different strategic segments. There are no statistically significant differences in walkaway rate, absolute win rate, and win rate conditional on bidding among opportunities of these different strategic segments. This suggests that industry type alone is not a primary driver of ServiceCo’s bidding decisions or success rates, indicating that other factors, such as customer relationships or opportunity size, play a more critical role.

Finally, Panel E in Figure 5 compares strategic-account customers with nonstrategic-account customers. The walkaway rate is significantly lower for strategic-account customers compared to nonstrategic-account customers ( $p<.01$ ), and the absolute win rate is also significantly higher for strategic-account customers ( $p<.05$ ). However, there is no statistically significant difference in win rate conditional on bidding between the two groups. These results reinforce the importance of strategic relationships in ServiceCo’s bidding strategy, as the firm is more likely to bid on and win opportunities with strategic accounts, even though its success rate in competitive bidding situations does not significantly differ between strategic and nonstrategic customers.

To complement the model-free analysis, we formally estimate logit models to quantify the effects of opportunity size, relationship type, strategic segment, and customer status on three key decisions in the bidding process: (1) the customer’s decision to withdraw the opportunity, (2) ServiceCo’s decision to bid for the opportunity, and (3) ServiceCo’s likelihood of winning the opportunity if a bid is submitted. These models provide a structured statistical assessment of how these factors influence decision-making beyond the descriptive insights from the model-free evidence. The results of these logit models reinforce the observed patterns, confirming that larger opportunities have a higher likelihood of being withdrawn by customers, ServiceCo is more likely to bid on opportunities where it has an existing relationship, and stronger customer relationships significantly increase the likelihood of winning a bid. Full details of these analyses are provided in the EC.5.

These descriptive findings highlight clear patterns in ServiceCo’s bidding behavior and success rates, suggesting that factors such as opportunity size, relationship type, and customer strategic status play a key role in influencing decision-making and outcomes. Taken together, the evidence indicates that ServiceCo’s decision-making approach is driven more by experience-based heuristics than by a structured evaluation of revenue potential and competition.

ServiceCo is more likely to bid on smaller- and medium-revenue opportunities, suggesting a potential risk aversion or capacity constraint that discourages the pursuit of large-scale projects. Additionally, ServiceCo favors opportunities where it has an existing relationship with the customer, particularly in direct relationships and strategic accounts. In contrast, the firm is more likely to walk away from opportunities involving new customers, especially those at new locations with no prior relationship, reinforcing a conservative, relationship-driven bidding strategy.

Borrowing from Ansoff’s matrix (Ansoff, 1957), a strategic framework used to assess business growth strategies, one could argue that ServiceCo primarily follows a market penetration strategy, which focuses on expanding sales within existing markets using familiar products and customers while avoiding the risks associated with diversification or new market entry. A more structured, optimization-driven bidding strategy could help ServiceCo better align its resource allocation with profitability and long-term growth, potentially enabling it to engage in market development, expanding into new customer segments or geographic markets, while maintaining a balanced risk approach.

5.3. Application of the ML Model to Predict ServiceCo’s Likelihood of Winning the Bid

In this section, we apply the modeling framework explained earlier to develop a supervised ML model to predict ServiceCo’s likelihood of winning the bid. We train the ensemble model using historical-bidding data from ServiceCo, incorporating key opportunity attributes such as opportunity size, relationship type, strategic segment, and service level. We split the 2,821 opportunities ServiceCo bid on into a training set that includes a stratified random sample of 80% of the opportunities and a test set that includes the remaining 20% of the opportunities. To avoid overfitting, we used a five-fold cross-validation procedure for hyperparameter tuning. We use the size of the project (continuous variable), relationship type, strategic segment, strategic account, and service level for projects across different locations as inputs for the ML models.

5.3.1. Model Evaluation

We compare the predictive performance of different ML models stacked in the ensemble model with the ensemble model in both train and test samples. Any binary classification problem usually has two classes—a positive class (ServiceCo won the bid) and a negative class (ServiceCo lost the bid)—and the model predictions can be classified into true positives where the model predicts a win and ServiceCo actually won the bid, false positives where the model predicts a win but ServiceCo actually lost the bid, true negatives where the model predicts a loss and ServiceCo actually lost the bid, and false negatives where the model predicts a loss but ServiceCo actually won the bid. We use these four values to calculate Accuracy, Sensitivity, and Specificity to compare the performance of different models. Detailed definitions of each of these measures are provided in EC.6.

The test sample performance evaluates how well the model predicts outcomes for unseen opportunities. Table 3 presents the results for the five individual models and the ensemble model. The ensemble model outperforms all the individual models in Accuracy, Sensitivity, and Specificity for the training sample. The next best model is the RF model. For the test sample, the ensemble model outperforms all the individual models in Accuracy and Specificity, followed by the RF model. On Sensitivity in the test sample, the ensemble model and the RF model perform equally well.

Table 3.

Performance of different machine learning models.

Machine learning model	Training sample (80% data)			Test sample (20% data)
	Accuracy	Sensitivity	Specificity	Accuracy	Sensitivity	Specificity
Random forest	94.72	92.97	96.37	76.42	75.18	77.59
Generalized boosted machines	76.03	74.64	77.36	72.16	67.88	76.21
Logistic regression	69.95	62.5	77.01	68.79	59.85	77.24
Naïve Bayes	64.67	72.45	57.3	65.25	72.99	57.93
Support vector machine	74.61	71.9	77.18	72.16	68.25	75.86
Ensemble model	99.25	98.81	99.65	77.48	75.18	79.66

5.3.2. Marginal Dependence of Predicted Outcome of the Ensemble Model on Opportunity Characteristics

To understand how different factors influence ServiceCo’s likelihood of winning bids, we analyze PDPs following Friedman (2001), with methodological details provided in EC.7. Figure 6 shows how opportunity size, relationship type, service level, strategic segment, and strategic-account status affect predicted win propensity. Tukey’s honest significance test is used to assess differences across categorical variables.

Figure 6.

Partial dependence plots following Friedman (2001).

For Project Size, the first plot shows a clear nonlinear relationship: win propensity fluctuates around 50% for projects below $8 million but declines steadily thereafter, reaching its lowest levels beyond $40 million. The overall slope is negative ( $\beta =-.156$ , $p<.01$ ), indicating that larger projects are significantly less likely to be won.

For Relationship Type, win propensity varies systematically with relational proximity. Opportunities involving a direct customer relationship exhibit the highest success rates ( $\approx 69\mathrm{\%}$ ), followed by indirect relationships with new locations ( $\approx 87\mathrm{\%}$ ), while bids involving no prior relationship or indirect ties with older locations show substantially lower success probabilities ( $\approx$ 35% to 50%). These differences are statistically significant ( $p<.01$ ), underscoring the centrality of relational strength in competitive bidding.

For Service Level, average win propensities range between 48% and 52% across categories, with only small and inconsistent differences among levels. This suggests that the service tier does not meaningfully influence win likelihood once other factors are accounted for. For Strategic Segment, average win propensities are 53.7% for mining, 46.7% for energy, and 55.3% for other segments, with modest but significant variation ( $p<.10$ ). Finally, for strategic-account status, bids from strategic-account customers show higher win propensities (61%) compared to nonstrategic accounts (48%), a difference significant at $p<.01$ .

Overall, the marginal dependence plots highlight that project size, relationship strength, and strategic-account status have substantial and systematic effects on ServiceCo’s win probability. Specifically, smaller projects, stronger existing relationships, and bids from strategic accounts are associated with significantly higher chances of winning. In contrast, for factors, such as service level and strategic segment, the predicted win propensity hovers around 50% across different categories, suggesting that these variables play a comparatively limited role in shaping ServiceCo’s likelihood of success.

5.4. Optimizing ServiceCo’s Opportunity Pursuit Portfolio

ServiceCo pursues sales opportunities across 23 countries, organized into four regions—Asia Pacific & Africa, Europe, North America, and South America—each managed by a dedicated sales director and supported by a regional sales team. Given the lengthy bidding process, sales planning is conducted annually, with each planning cycle referred to as period $t$ .

To optimize ServiceCo’s bidding decisions, we develop an optimization model that maximizes expected revenue for each region $\ell$ during period $t$ . The available sales team capacity in each region is denoted by $C_t^{\ell }$ , representing the total monetary value (in dollars) of projects that can be pursued in region $\ell$ at time $t$ . Since some bids initiated in prior periods continue to consume capacity in the current period, we assume that ServiceCo can estimate the remaining available capacity at the start of each new period. Initially, we assume that each region’s sales capacity is used exclusively for projects within that region, but this constraint can be relaxed to allow a flexible allocation of a percentage of capacity across regions.

ServiceCo primarily handles projects in three domains: Energy (73.4%), Mining (26.2%), and Other (0.4%). The Energy and Mining sectors account for 99.6% of sales opportunities, while the Other category includes segments such as Healthcare, Corporate Services, Government Services, and Schools. We assume that each region has two dedicated sales teams, one specializing in Energy and the other in Mining, with both teams capable of handling projects classified under Other domains as well. We define $C_{j,t}^{\ell }$ , which is the remaining capacity at a focal company in period, $t$ , in region $\ell$ for product type $j$ , where $j=1$ indicates the sales team’s capacity for pursuing products under Energy domain, while $j=2$ and $j=3$ indicate Mining and Other domains, respectively. To be general, we assume, $j=1,2,\dots ,m-1$ are major product types, $j=m$ is classified as Others. We assume these sales teams handling project type $j$ , $j=1,2,\dots ,m-1$ , can also handle sales opportunities in the domain classified as Others (product type $m$ ). In our case study, $m=3$ . Thus, the capacity $C_{j,t}^{\ell }$ is defined only for the major product types, $j=1,2,\dots ,m-1$ .

The input data for the optimization model are derived from the predictive ML model developed earlier section, which serves as a sales response function. We assume that the predicted win probabilities generated by this model provide a reasonable approximation of the firm’s likelihood of success across both historically bid and unbid opportunities. This assumption enables us to simulate counterfactual scenarios and evaluate the potential impact of alternative bidding strategies. The optimization model computed the optimal revenue-maximizing portfolio of projects to pursue from a large pool of opportunities available for the firm to pursue for each region. Model $M_{\mathrm{SM}}$ describes the optimization problem to be solved to arrive at a revenue-maximizing portfolio for period $t$ in a single region $\ell$ . The objective function maximizes the prospect revenue during the planning horizon $t$ , and the capacity constraint restricts the capacity of the total monetary value of projects bid on in the time period $t$ in region $\ell$ .

Our predictive model estimates the probability of winning an opportunity conditional on the focal firm having submitted a bid, that is, $P(\text{win}\mid \text{bid})$ . Applying these predictions to opportunities that were not historically pursued requires an implicit assumption that the same win-probability relationship holds for both bid and unbid opportunities. If the firm’s no-bid decisions are primarily driven by an assessment of low win likelihood, then predicted probabilities for unbid opportunities are likely to be downward-biased. Conversely, if non-performance-related factors, such as temporary capacity constraints, incomplete opportunity information, or shifts in strategic focus, explain the no-bid decision, then win probabilities for unbid opportunities could be upward-biased. In either case, the lack of observed outcomes for unbid opportunities introduces uncertainty into the counterfactual estimates.

To address this concern, we also implement a two-stage Heckman-style framework, detailed in the Robustness Check section, that explicitly models the bid decision in the first stage and the win probability conditional on bidding in the second stage, enabling adjusted estimates of the joint probability of bidding and winning. In addition to benchmarking against historical bidding behavior, we evaluate our model-based optimization against two simple rule-based heuristics often used in practice: (i) bidding only on opportunities with revenue above the sample median, and (ii) bidding only on opportunities in the top or bottom quartile of revenue. As shown in EC.11, both heuristics result in substantially fewer bids and lower total revenue won, highlighting the value added by our predictive and optimization framework.

We apply this optimization approach to ServiceCo’s 4,024 sales opportunities to identify the optimal subset of opportunities that the company should have pursued each year by region. By comparing the actual bidding decisions made by ServiceCo with the recommendations from the optimization model, we assess the potential improvements in revenue generation and resource efficiency.

Table 4 presents a side-by-side comparison of the projects ServiceCo actually bid on each year and the opportunities identified by the optimization model. The results reveal that ServiceCo’s current bidding strategy involved substantially more bids but with lower revenue efficiency. Specifically, between 2010 and 2021, ServiceCo bid on 2,821 projects, with a total bid value of $3,819.27 million, resulting in an actual revenue of $1,663.44 million. In contrast, the optimization model recommends bidding on only 1,756 projects, with a total bid value of $2,011.97 million, but achieving a higher expected revenue of $2,011.97 million. This indicates that ServiceCo could have realized a 20.95% increase in revenue while bidding on 37.75% fewer projects, highlighting significant inefficiencies in its current opportunity selection process.

Table 4.

Optimization results with capacity constraint on total number of projects ServiceCo can bid in a year by region.

		True values			Projected values based on optimization			Percentage change
Year	Region	Projects bid	Revenue bid	Revenue won	Projects should bid	Revenue bid	Conditional revenue won	Revenue	Bids
2010	North America	6	1.95	1.95	1	0.57	0.57	−70.95%	−83.33%
2011	North America	61	78.57	24.27	30	40.71	40.71	67.74%	−50.82%
2012	North America	53	71.77	41.07	29	33.62	33.62	−18.13%	−45.28%
2013	Europe	1	0.23	0.23	1	0.23	0.23	−0.98%	0.00%
2013	North America	49	110.87	61.03	48	56.88	56.88	−6.80%	−2.04%
2014	North America	60	143.77	16.22	16	20.36	20.36	25.54%	−73.33%
2014	South America	1	0.64	0.64	1	0.64	0.64	0.20%	0.00%
2015	North America	125	59.03	22.38	14	16.67	16.67	−25.51%	−88.80%
2015	South America	2	0.64	0.43	1	0.43	0.43	1.10%	−50.00%
2016	Asia Pacific & Africa	15	11.72	3.19	9	6.97	6.97	118.43%	−40.00%
2016	Europe	25	26.13	11.89	9	13.81	13.81	16.15%	−64.00%
2016	North America	73	111.99	36.45	27	44.92	44.92	23.23%	−63.01%
2016	South America	8	27.07	25.51	3	9.00	9.00	−64.70%	−62.50%
2017	Asia Pacific & Africa	141	159.59	58.39	82	62.88	62.88	7.70%	−41.84%
2017	Europe	56	66.52	21.88	32	24.59	24.59	12.38%	−42.86%
2017	North America	72	70.45	18.17	27	20.92	20.92	15.12%	−62.50%
2017	South America	182	257.38	77.54	94	108.87	108.87	40.40%	−48.35%
2018	Asia Pacific & Africa	111	155.88	54.63	103	74.22	74.22	35.86%	−7.21%
2018	Europe	89	122.47	86.67	75	107.88	107.88	24.47%	−15.73%
2018	North America	86	128.13	29.27	42	70.86	70.86	142.09%	−51.16%
2018	South America	276	298.27	58.80	89	83.02	83.02	41.19%	−67.75%
2019	Asia Pacific & Africa	67	122.33	87.23	54	90.76	90.76	4.05%	−19.40%
2019	Europe	99	48.29	26.79	77	28.86	28.86	7.72%	−22.22%
2019	North America	64	65.00	21.75	33	22.71	22.71	4.43%	−48.44%
2019	South America	183	260.91	83.36	76	78.25	78.25	−6.13%	−58.47%
2020	Asia Pacific & Africa	82	128.57	64.89	45	74.95	74.95	15.51%	−45.12%
2020	Europe	98	117.59	62.26	69	72.81	72.81	16.95%	−29.59%
2020	North America	154	278.89	149.15	121	174.42	174.42	16.95%	−21.43%
2020	South America	127	241.04	147.46	112	148.87	148.87	0.96%	−11.81%
2021	Asia Pacific & Africa	99	135.90	93.58	77	112.42	112.42	20.13%	−22.22%
2021	Europe	91	120.52	67.49	72	117.83	117.83	74.60%	−20.88%
2021	North America	63	106.87	56.31	61	89.69	89.69	59.29%	−3.17%
2021	South America	202	290.29	152.56	226	202.32	202.32	32.62%	11.88%
Total		2,821	3,819.27	1,663.44	1,756	2,011.97	2,011.97	20.95%	−37.75%

To further understand the drivers behind these gains, we examined the distribution of predicted win probabilities for all opportunities included in the counterfactual simulations, disaggregated by revenue category (Figure 7). The overall distribution is bimodal, with peaks at very low and very high predicted win probabilities. Low-revenue opportunities are concentrated at both extremes, medium-revenue opportunities follow a similar pattern but with relatively fewer very-high-probability cases, and high-revenue opportunities are more prevalent in the low-to-moderate-probability range. This distributional evidence indicates that the observed revenue increase from the optimization model stems from both improved targeting of high-probability opportunities and the strategic inclusion of moderate-probability opportunities with high-revenue potential, opportunities that were often overlooked in ServiceCo’s historical bidding.

Figure 7.

Density of predicted win probabilities for all opportunities in the counterfactual simulation, disaggregated by revenue category. The top-left panel shows the overall distribution. Other panels show low-, medium-, and high-revenue categories separately.

A deeper analysis of year–region trends shows that the optimization model does not uniformly recommend fewer bids in every instance. Instead, it identifies situations where bidding should have been more selective and other instances where more aggressive bidding could have been beneficial.

For instance, in 2010, ServiceCo bid on six projects in North America, accounting for a total revenue of $1.95 million. However, the optimization model recommends bidding on only one project, with an expected revenue of $0.57 million, resulting in a 71% decrease in revenue compared to the actual outcome. This suggests that ServiceCo overcommitted resources to projects with low return potential, spreading its capacity too thin. Similarly, in 2012, ServiceCo pursued 53 projects in North America, generating $41.07 million in total revenue. The optimization model, however, recommends focusing on only 29 projects, with a slightly lower expected revenue of $33.62 million, an 18.13% decrease compared to the actual revenue. These examples illustrate how a more focused approach could have optimized resource allocation and improved profitability.

Conversely, the optimization model also identifies cases where ServiceCo underinvests in high-value opportunities. For example, in 2011, ServiceCo bid on 61 projects in North America, generating $24.27 million in revenue. The optimization model suggests bidding on only 30 projects, but these would have yielded $40.71 million, representing a 67.71% increase in revenue compared to reality. Similarly, in 2021, ServiceCo bid on 99 projects in Asia Pacific & Africa, generating $93.58 million in revenue. The optimization model recommends bidding on only 77 projects, but with an expected revenue of $112.42 million, representing a 20.13% revenue increase.

These findings highlight that ServiceCo’s current bidding strategy often prioritizes quantity over quality, leading to unnecessary resource expenditure on low-value projects while missing higher-value opportunities. By using the optimization model, ServiceCo could have strategically refined its opportunity selection, achieving higher revenue with fewer bids, ultimately improving its profitability and operational efficiency. The optimization model’s recommendations provide actionable guidance for refining future bidding strategies, ensuring that sales teams focus on the most promising opportunities, aligning resources more effectively, and enhancing overall financial performance.

5.5. Heckman-Style Bias Correction

A key concern with the main counterfactual analysis is that the win-probability model is estimated only on opportunities that the firm historically chose to bid on. If the firm’s bid decisions are systematically related to unobserved factors that also affect win likelihood, such as relationship quality, perceived fit, or resource constraints, then directly applying this model to historically unbid opportunities may introduce selection bias. In particular, the distribution of features for unbid opportunities may differ from that of bid opportunities, potentially leading to biased win-probability estimates and overstated or understated revenue effects.

To address this concern, we implement a two-stage Heckman-style framework that explicitly models the bid decision in the first stage and the win outcome in the second stage. This correction accounts for selection effects when estimating win probabilities, which are then integrated into the optimization routine under the same capacity constraints as the baseline ML + optimization model. The results, presented in Table 5, show that the Heckman-style correction yields a more conservative estimate of revenue improvement (+14.65%) compared to the baseline model (+20.95%). This approach provides managers with a practical benchmark when historical data are incomplete or selectively recorded, helping them adjust expectations and strategy depending on data quality. Full implementation details of the Heckman-style model are provided in EC.10.

Table 5.

Comparison of Heckman-style bias-corrected and cost-minimization optimization models.

Strategy	Projects bid	$\mathrm{\Delta }$ Projects (%)	Revenue bid	Revenue won	$\mathrm{\Delta }$ Revenue won (%)
Baseline ML + optimization	1,756	−37.75	2,011.97	2,011.97	+20.95
Heckman-style ML + optimization	1,770	−37.25	1,907.37	1,907.37	+14.65
Baseline ML + cost-minimization optimization	1,372	−51.36	1,614.33	1,614.33	−2.95
Historical (true values)	2,821	–	3,819.27	1,663.44	–

Note. ML = machine learning.

5.6. Cost-Minimization Optimization

We also test the robustness of our results by reformulating the optimization problem as a cost-minimization exercise. In this specification, the focus is limited to the opportunities the firm historically pursued, and the objective function minimizes the expected cost of bidding subject to budget and capacity constraints. This specification removes any counterfactual extrapolation to unbid opportunities and isolates the efficiency gains purely from improved allocation among historically bid projects.

As shown in Table 5, this cost-minimization framework produces results consistent with the main analysis: optimized allocation improves efficiency relative to historical performance and substantially outperforms heuristic benchmarks, even when focusing solely on the observed subset of opportunities.

6. Discussion and Conclusion

In this research, we seek to improve the state of research and practice in B2B sales and operations planning by focusing on the important problem of opportunity management for service organizations. Motivated by the ad hoc approach to this topic by executives, we build a theory-informed framework that relates bid characteristics, relationship strength, and opportunity size to outcomes and overlay it with an optimization model. The results of the empirical analysis document the biases held by the sales teams in opportunity pursuit. To alleviate these inherent biases, we propose and develop a data-driven decision support framework informed by extant marketing theory that can help firms prioritize the opportunities to pursue from a larger opportunity pool. These results have important theoretical and practical implications.

6.1. Theoretical Implications

Our findings offer several important contributions to the emerging literature on salesforce decision-making, particularly within the relatively understudied B2B context.

First, we find that firms are systematically less likely to pursue larger opportunities. Given that opportunity size often proxies for project complexity, our results suggest that sales decision-makers may exhibit a bias toward avoiding more complex engagements. This extends the growing literature on salesperson judgment and decision-making by providing evidence of risk-avoidance behavior in the context of opportunity pursuit, a phenomenon that has received limited attention in B2B settings. Our results align with prior work documenting cognitive biases among salespeople (e.g., boundary conditions in opportunity selection) and highlight how such biases can lead to suboptimal resource allocation.

Second, we find that firms are more likely to pursue opportunities where they have a pre-existing relationship with the customer. Conversely, they are more likely to forgo opportunities with unfamiliar customers. Interestingly, we observe no significant interaction between opportunity size and prior relationship, indicating that the aversion to large opportunities persists regardless of the firm’s relational history. This adds nuance to existing theories on relationship marketing and customer familiarity by showing that even strong relational ties do not fully mitigate complexity-related avoidance.

Third, our prescriptive optimization model reveals that firms are potentially leaving value on the table by not bidding on opportunities where their likelihood of success is higher. This insight bridges behavioral sales research with decision science by demonstrating how data-driven decision support tools can correct for human biases and improve firm outcomes. We thus contribute to the theoretical conversation around sales enablement by introducing optimization-based insights into opportunity selection.

Taken together, our study advances theoretical understanding of how B2B firms evaluate and act upon sales opportunities, especially under conditions of complexity and uncertainty. By grounding our contributions in actual decision data and modeling, we provide an empirically robust foundation for extending theories of sales behavior, decision support, and opportunity management.

6.2. Managerial Implications

This research demonstrates how integrating predictive analytics with prescriptive optimization can improve resource allocation and revenue outcomes in complex B2B environments. These results offer several important implications for practice, particularly for senior B2B executives such as Chief Sales Officers, Chief Operations Officers, and Chief Executive Officers.

First, the findings highlight that sales managers must recognize and correct for behavioral and data-driven biases when prioritizing sales opportunities. The prescriptive framework quantifies the revenue impact of these biases: in our context, the focal firm could have increased its total revenue by approximately 21% through a data-driven optimization of opportunity pursuit, compared to historical performance.

Second, the optimization framework provides executives with a flexible decision support tool that can be customized to align with specific strategic goals. For example, a manager seeking to prioritize certain service levels, customer segments, or regions can introduce these preferences as additional constraints in the optimization model. The resulting recommendations then reflect both revenue-maximizing and strategy-consistent opportunity allocations.

Third, by incorporating the Heckman-style bias correction into the analysis, managers are equipped with a conservative benchmark that accounts for potential selection bias in historical data. This version of the model allows decision-makers to adjust their expectations and strategies based on the reliability of their firm’s data systems—using the bias-corrected estimates when data completeness or consistency is a concern.

Finally, this integrated “predict + optimize” framework can also enhance sales forecasting accuracy and operational planning by linking expected wins to downstream production and resource capacity needs. Together, these insights help senior executives balance growth ambitions with operational feasibility in complex B2B service environments.

6.3. Limitations and Future Research

This study has several limitations that offer opportunities for future research. First, the analysis relies on CRM data from a single firm, capturing only the opportunities in which the focal firm submitted bids. As such, the dataset does not represent all opportunities available in the broader market, introducing potential selection bias. This limitation may affect the generalizability of our findings. Future work could explore richer contexts, such as business-to-government contracting, where publicly available data enables researchers to observe the full universe of opportunities and develop more comprehensive models of opportunity pursuit.

Second, our current predictive model estimates win probability at a single point in time. A promising avenue for future research would be to develop dynamic models that update win probabilities as opportunities progress through the sales pipeline, enabling more informed decision-making throughout the sales process. Third, due to data limitations, we were unable to include certain potentially influential variables, such as competition intensity, bidding costs, and detailed compensation schemes. While our models demonstrate high-predictive accuracy and we anticipate that the marginal gain from including such variables would be limited in this empirical context, future research using richer datasets could explore how these factors shape bidding behavior and outcomes.

Fourth, our main counterfactual simulation assumes that competitors do not alter their bidding strategies in response to changes in the focal firm’s approach. While this assumption simplifies the analysis, it is important to acknowledge its implications in a competitive B2B setting. In reality, competitors could respond to the focal firm’s strategy shifts by reallocating sales resources, lowering prices, targeting different market segments, or defending high-value accounts more aggressively. Such strategic reactions could reduce the focal firm’s win probabilities and, consequently, attenuate the gains estimated in our counterfactual analysis. Accordingly, our results should be interpreted as upper-bound estimates under the assumption of fixed competitor behavior. Future research could more fully assess the dynamic impact of predictive decision support tools by embedding them in field experiments, thereby allowing direct observation of both focal firm and competitor responses.

Finally, our study focuses on structured CRM data. Future research could incorporate unstructured data, such as text from email exchanges, audio or video from sales interactions, or notes from internal sales meetings, to enhance predictive power and provide a more holistic understanding of the sales pursuit process. Integrating such multimodal data with probabilistic forecasting techniques could enable more robust, real-time optimization of sales strategies in complex and uncertain environments.

Supplemental Material

sj-pdf-1-pao-10.1177_10591478251407559 - Supplemental material for Opportunity Management for Business-to-Business (B2B) Service Organizations: A Theory-Informed Decision Support Framework

Supplemental material, sj-pdf-1-pao-10.1177_10591478251407559 for Opportunity Management for Business-to-Business (B2B) Service Organizations: A Theory-Informed Decision Support Framework by Muzeeb Shaik, Shrihari Sridhar, Chelliah Sriskandarajah and Vikas Mittal in Production and Operations Management