Winning a complex B2B deal often depends on details that cannot be captured by a single sales metric. A quotation may involve demanding technical requirements, a specialized product configuration, commercial conditions shaped by previous negotiations and a customer relationship that the account manager understands through years of experience.
Much of this information leaves a record in CRM and quoting systems, making it possible to examine historical offers and identify patterns associated with successful negotiations. Machine learning models can use those records to predict win/loss outcomes, giving commercial teams another perspective on opportunities that might otherwise be evaluated largely through individual judgment.
What remains less straightforward is understanding why a model considers one offer more likely to succeed than another, particularly when its assessment differs from that of an experienced salesperson.
Explainable AI in B2B sales helps address this question by revealing how recorded characteristics contribute to predictive results. Rather than turning those explanations directly into commercial recommendations, the opportunity is to use them to examine assumptions, understand differences in judgment and make more informed decisions about customers, offers and pricing.
What Historical B2B Offers Can Reveal
Industrial sales provide a particularly interesting environment for predictive analysis because quotations often involve considerable technical work before the customer makes a purchasing decision.
In configure-to-order manufacturing, equipment is assembled from components and options selected to meet individual customer requirements. Configure-Price-Quote (CPQ) systems support this process by recording product configurations and related commercial information, while CRM platforms provide additional details about customers, sales organizations and the progress of negotiations.
Together, these records offer a way to examine how different characteristics of an opportunity relate to historical sales outcomes, although the usefulness of the analysis depends on the quality and completeness of the available information.
An August 2026 study by Tuomas Sormunen and colleagues, published in the Journal of Marketing Analytics, investigated this possibility using data from an industrial manufacturer of configurable equipment.
The researchers developed an XGBoost classification model based on 439 completed sales offers from two countries, combining information extracted from CRM and CPQ systems. The objective was to distinguish offers that had been won from those that had been lost, using characteristics available during the quotation process.
The model achieved a reported 83.9% balanced accuracy for the combined dataset and 74.1% for one regional subset. Sales personnel’s recorded win/loss estimates achieved 63.0% and 54.4%, respectively, suggesting that the model captured predictive relationships that were not fully reflected in their judgments.
The comparison has an important methodological limitation, however, because model performance was measured using held-out test observations, whereas the sales personnel benchmarks covered a broader collection of cases. The findings are therefore promising within the industrial environment examined, without establishing that predictive models will consistently outperform commercial professionals across other organizations or markets.
Perhaps more interesting than the difference in predictive performance is what the researchers discovered when they examined the model’s reasoning alongside the experience of the people working with those offers.
What SHAP Adds to a Win/Loss Prediction
A win/loss classification can help a sales manager identify an opportunity that deserves closer examination, but the result becomes more informative when it is possible to understand which characteristics influenced the model’s assessment.
Sormunen and colleagues used SHAP (SHapley Additive Explanations) to interpret the predictions generated by XGBoost. The technique assigns contributions to individual input variables relative to a reference model output, showing how different characteristics influence the prediction for a particular offer.
These explanations can also be examined across multiple cases, revealing which variables have contributed most strongly to the model’s predictions overall. This broader perspective is useful when a commercial team wants to understand recurring patterns rather than investigate an individual quotation.
In the industrial dataset, several pricing-related and product-related variables emerged as influential, including a feature representing the cost of specialized components within the proposed equipment. The relative importance of some variables also differed between the combined dataset and the regional model, reflecting the need to consider the context in which predictions are made.
For a commercial professional, an explanation of this kind provides a way to examine information that might otherwise remain hidden behind a probability or classification. It can highlight a product characteristic that deserves attention, reveal a pattern that was not immediately apparent and provide a basis for discussing how an offer compares with previous negotiations.
However, the contribution assigned to a feature describes its role in the trained model, which may not correspond directly to the importance that a salesperson attributes to the same characteristic when evaluating a real customer relationship.
When Sales Experience Tells a Different Story
To investigate whether the model’s explanations made sense to professionals familiar with the sales process, the researchers involved five business operations experts from the industrial partner in a structured evaluation.
The participants reviewed SHAP-based explanations for four selected offers, including two successful and two unsuccessful cases, alongside findings from the regional model. All four individual cases had been correctly classified by the algorithm, allowing the researchers to focus on how the explanations were understood rather than on the reasons behind incorrect predictions.
The comparison revealed partial agreement, but some of the differences were particularly revealing.
One pricing-related feature associated with specialized product configurations had a strong influence on the model’s predictions, yet the experts considered it only moderately important. Other product characteristics received relatively high importance ratings from the professionals despite contributing less strongly to the model’s output.
The customer’s identity offered another example of how the two perspectives differed. Sales professionals regarded the customer’s name as relevant to their evaluation, even though that information was not included among the model’s input variables.
Their reasoning may have reflected knowledge of previous negotiations, established relationships or customer-specific circumstances that were difficult to represent through the structured fields available for modeling. Conversely, the algorithm could identify statistical patterns across historical offers that individual professionals might not recognize from their own experience.
These differences are worth examining because they do not necessarily indicate that either perspective is incorrect. A model is limited by the information it receives and the relationships it learns, while commercial judgment depends on experience, context and assumptions that may not always be supported by the broader historical evidence.
The experts generally considered the explanations understandable and potentially useful, rating a tool that combined predictions with explanations more favorably than one providing predictions alone, with average scores of 4.2 and 3.6 out of 5, respectively.
The evaluation was nevertheless limited to five participants and four selected cases, and the researchers did not observe a sales process in which professionals used the explanations to make decisions and subsequently measured whether their results improved.
What the findings provide is an initial indication that explainability can make predictive information more meaningful to commercial professionals, particularly when the model’s assessment invites them to reconsider aspects of an offer they might otherwise have overlooked.
Why a Pricing Signal Is Not Pricing Advice
Pricing deserves particular attention when interpreting win/loss explanations because it is one of the areas where a statistical relationship can easily be mistaken for a recommendation.
Suppose a model associates offers containing expensive specialized components with a higher likelihood of success. It might be tempting to conclude that adding similar features to another quotation would improve its chances of being accepted, even though the association could reflect the types of customers requesting those components or the requirements of the projects involved.
In the research by Sormunen and colleagues, the pricing-related features included the cost of specialized product options, the total cost of the equipment and a ratio describing the degree of specialized configuration. These variables contain useful commercial information, but they should not be treated as interchangeable with the final quoted selling price or a discount that can be adjusted independently.
A specialized configuration may be associated with historical success because it responds to a particular customer need, while changing that configuration for another buyer could increase costs without improving the attractiveness of the offer.
Discounts and other commercial conditions present similar difficulties, since prices are often shaped by competition, negotiation history, purchasing requirements and the strategic importance of individual accounts.
SHAP can show how a variable contributes to a model’s prediction, but it does not establish the causal effect of changing that variable. Determining whether a different price or configuration would improve conversion requires additional evidence, particularly when the adjustment could also affect profitability or the technical suitability of the proposed equipment.
For sales teams, the explanation is therefore a reason to investigate the commercial context more closely, rather than an instruction to change an offer simply because a feature has a positive or negative contribution.
Bringing Explanations into Commercial Reviews
The usefulness of explainable AI ultimately depends on how commercial professionals work with the information, especially in organizations where evaluating opportunities and preparing technical quotations require considerable time and coordination.
A sales manager reviewing an open opportunity could examine its win/loss prediction alongside the variables that influenced the model’s assessment, comparing those signals with the account team’s understanding of the customer’s requirements and previous interactions.
Where the two perspectives differ, the disagreement could prompt a more focused discussion. The model may be relying on a pattern that no longer reflects current market conditions, or the salesperson may possess relevant customer knowledge that was never recorded in the data used for training.
This type of review also has implications for deciding where commercial resources should be directed. Win probability can help identify opportunities worth investigating, but it does not capture every consideration involved in prioritization, including expected deal value, margin, strategic relevance and the effort needed to prepare or revise a quotation.
Similarly, explanations associated with product characteristics may offer useful context when discussing configurations, while pricing-related contributions can help clarify the factors behind a prediction without establishing an optimal selling price or discount. For these applications to be dependable, organizations need to assess the model’s performance on representative commercial data, examine whether estimated probabilities are sufficiently calibrated for their intended use and ensure that the input variables reflect information available at the time the prediction is made.
Changes in product portfolios, customer requirements or commercial policies can also affect the relevance of historical relationships, making periodic evaluation important when a model is used to support ongoing decisions.
The study offers evidence that experienced professionals can find SHAP-based explanations understandable and useful, but establishing their effect on actual sales decisions would require a further evaluation of how the information is used and whether it improves outcomes compared with existing practices.
Making Predictive Evidence More Useful to Sales Teams
In B2B sales, a prediction can provide an additional perspective on an opportunity, particularly when historical offers contain patterns that are difficult to recognize through individual experience. Understanding the information behind that prediction, however, is essential if commercial teams are to interpret it within the circumstances of a particular negotiation.
The research by Sormunen and colleagues illustrates how explainability can support that interpretation, while the differences between SHAP contributions and expert assessments show why statistical evidence and commercial experience may highlight different aspects of the same offer.
Rather than treating the model as an authority or relying entirely on individual judgment, organizations can use these differences to examine assumptions, identify missing context and understand which characteristics influenced the predictive assessment.
The contribution of explainable AI is not to tell sales teams exactly how to win a deal, but to give them a clearer basis for evaluating an opportunity and deciding which questions deserve further investigation.