AI-Powered Segmentation and Prioritization of e-Mobility Presales Requests

AI-Powered Prioritization of E-Mobility Presales Requests: A CRISP-DM Case Study Amine Kazdar Polytechnique Montréal ([email protected]) Preprint — not peer-reviewed. Version 1.0 (2025-09-22) Abstract: The Tenders & Proposals team at an anonymized industrial partner in the e-mobility sector plays a key role in growing the business by answering presales requests. One of the main challenges they face is the absence of a structured way to prioritize those requests. Right now, decisions are mostly based on individual experience, which can result in spending time on low-impact opportunities while missing higher-value ones. To support the team, I created a decision-support tool powered by artificial intelligence to help prioritize quotes based on data rather than intuition. The model used is a Random Forest (RF), trained on past quotes and key variables like complexity level, customer segment, total value, price tier, and urgency. The features were cleaned, recategorized, and grouped when needed — for example, total value was grouped by business rule (0-$5M vs $5M-$25M). On top of that, I built a prioritization score using feature importance to assign a weight to each request. To improve the model’s performance, especially in identifying “Won (1)” opportunities, I tested several enhancements including a priority scoring matrix, Gradient Boosting (GB), and cross-validation. The result is a simple and effective tool that helps the team focus on the right requests, improve their chances of success, and manage their time better. In offline tests, the RF baseline achieved 90.0% accuracy (weighted F1 0.89), and a GB variant increased precision for ‘Won (1)’ precision to 50.0% (recall 13.3%). Keywords: RFQ prioritization; presales analytics; CRISP-DM; Random Forest; Gradient Boosting; emobility 1. INTRODUCTION 1.1 Context & Problem Presales teams at the company face daily overload when triaging RFQ/quote requests in an engineer-to-order context with strict due dates and limited capacity. In the absence of a reproducible, data-driven prioritization process, time can be spent on low-impact opportunities while higher-value ones are missed. 1.2 Objective This paper investigates how historical RFQ/quote data can be used to prioritize presales requests in order to maximize expected value while meeting service-level agreement (SLA) and capacity constraints. 1.3 Approach We build a priority scoring scheme that combines quote complexity, request type, price tier, urgency (Days Until Due = Due Date − Request Date), and total value into a single “cap = 5” score, and we train supervised models (RF and GB) on cleaned and recoded features, including value grouping by business rule (0–$5M vs $5M–$25M) . We evaluate with accuracy and class-aware metrics (precision/recall/F1), focusing on the “Won (1) ” class, and we report a lightweight deployment in Excel/VBA with a roadmap to Salesforce. 1.4 Contributions This work contributes: (1) an operational, business-aligned priority score (cap = 5) built from price tier, complexity, request type, urgency, and value; (2) an empirical evaluation with 5-fold cross-validation and class-aware metrics, including a precision-oriented GB variant; and (3) a deployed tool adopted by users, with planned Salesforce integration. 2. RELATED WORK We used Google Scholar and Compendex to identify prior work related to RFQ/quotation processes, bid/no-bid, and order acceptance/scheduling in ETO/MTO contexts, and standard process guidance (CRISP-DM) (Nederkoorn, 2017; Mattila, 2023). Prior studies on RFQ/quotation, bid/no-bid, and order acceptance in make-/engineer-to-order contexts discuss decision criteria and scheduling trade-offs for presales triage (Nederkoorn, 2017). Industrial ML work on prioritization and triage motivates the use of supervised models with class-aware metrics when positive outcomes (Won (1)) are rare (Saito and Rehmsmeier, 2015; He and Garcia, 2009). CRISP-DM provides a pragmatic process for problem framing, data preparation, modeling, evaluation, and deployment in industrial settings, which we follow in this study (Shearer, 2000). There is an absence of a structured way to prioritize and a lack of a reproducible ranking of quotes by expected value (probability of success × value) that fits operational constraints (SLA, capacity, complexity, customization). This study follows CRISP-DM and provides a single priority score and a lightweight deployment in Excel/VBA. These findings motivate a presales prioritization scheme that combines business-rule scoring and supervised models to steer effort toward higher-value, higher-probability RFQs (see also Nederkoorn, 2017; Mattila, 2023). 3. METHODOLOGY(CRISP-DM) This study follows the CRISP-DM (Cross-Industry Standard Process for Data Mining) framework, which structures data science work into six clear phases: Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment. This approach ensures a consistent process from problem definition to implementation. 3.1 Business Understanding (CRISP-DM Phase 1) The Tenders & Proposals team needs to improve how they allocate their time and effort across incoming requests. Without guidelines, high-potential opportunities might be overlooked, while simpler, lower-impact quotes consume time. A predictive tool can support better decisions by ranking quotes based on historical success patterns. 3.2 Data Understanding (CRISP-DM Phase 2) The dataset includes historical quote records from the team’s internal quote log with fields such as: Segment (renamed for confidentiality), Complexity level (based on estimated hours required to respond), Total value (USD), Price tier, Outcome (Won (1) , Not Won (0) , or No Outcome), and Due date and Request date (used to calculate urgency). Initial exploration found inconsistencies, missing values, and non-standard formatting. Constraints (business & operational): – Electronic products, accessories, bundles (product + accessory) – Many customized quotes (engineer-to-order) – RFQ → quote → order; production after Won (1) (ETO) – Limited capacity (engineering/presales hours) – Backlog (what to start first) – SLA commitments and due dates to meet Terminology. Segment (customer category), Complexity Level (effort tier from Standard to R&D), Priority Score (“cap = 5” rule combining variables), Outcome (binary: Won (1) / Not Won (0) ). 3.3 Data Preparation (CRISP-DM Phase 3) Cleaning: rows with missing outcomes were removed, and miswritten categories were grouped under standard labels. Feature engineering: Segment was considered at the beginning but excluded later based on team feedback. Complexity level was renamed into simplified categories based on estimated effort—Tier 1 (2–6 hrs) Standard; Tier 2 (8–24 hrs) Spec Review; Tier 3 (24–32 hrs) Service & Risk; Tier 4 (40–60 hrs) High Complexity; Tier 5 (60–140 hrs) R&D Involved. Total value: grouped into two categories by business rule: Lower (0–$5M) and Upper ($5M–$25M). We initially explored K-Means, which suggested a threshold around $7.13M; the team adopted the rounded $5M boundary for alignment and ease of use. N/A was kept as Unknown. Price Tier was regrouped and encoded with levels 0–3: Price List/MSRP/Price Book → Level 0; Gold/Gold ± → Level 1; Platinum/Platinum +3% → Level 2; FCM/FCM/Price List/FCM + % → Level 3. Urgency was defined as Days Until Due = Due Date − Request Date. Labeling: outcomes were encoded as binary values (1 = Won, 0 = Lost or No Outcome). Priority score (business rule): a point-based score combines price tier, complexity level, request type, days until due, and total value; the final score is capped at 5; the scoring matrix is shown in Table 1. Table 1. Scoring matrix for prioritizing quotes (final score capped at 5). Factor Description Points Price Tier Level Ordinal level 0 – 3 Complexity Level Tier 2/ Tier 3/ Tier 4/ Tier 5 0 / + 1 / + 2 / +3 Request type PO Submission; Quote/Tender & RFXs/Pricing +1 / 0 Days Until Due +1 If Days Until Due ≤ 3 +1 Total Value Lower (0–$5M) / Upper ($5M-$25M) 0 / +1 Final score Cap (max) 5 Exploratory check. We also ran K-Means on Total Value; the elbow/silhouette curve indicates an optimal k = 2 (Fig. 1) with centroids $430,556 and $13,823,022 (midpoint ≈ $7.13M). For reporting and adoption, we use the businessrule bands (0–$5M vs $5M–$25M). 3.4 Modeling (CRISP-DM Phase 4) We train a RF baseline using Price Tier Level, Complexity Level, Days Until Due, and Total Value category (Segment was analyzed initially and then excluded based on team feedback). We then test a GB variant. We evaluate: (i) RF baseline (without PriorityScore), (ii) RF + PriorityScore (score added as a feature), and (iii) GB + PriorityScore. We also report a 5-fold cross-validation variant. In initial experiments, 20% of the quotes were reserved as a fixed test set; evaluating 100% would inflate performance due to overfitting. Metrics reported are Accuracy, Precision, Recall, and F1score for both Won (1) and Not Won (0) classes, plus weighted and macro-averaged scores to account for class imbalance. 3.5 Evaluation Plan (CRISP-DM Phase 5) Unless otherwise noted, results are reported using the 5-fold cross-validation protocol described above; the 20% hold-out split is reported where explicitly indicated. Metrics follow the list in section 3.4, with a specific focus on the Won class due to class imbalance. 4. RESULTS 4.1 Per-Class Performance and Model Comparison Unless stated otherwise, the tables below report results on the 20% hold-out set; cross-validation averages are reported where indicated. Overall performance for the RF is 90.0% accuracy with weighted F1 = 0.89. However, due to class imbalance (most quotes are "Not Won (0) "), the precision and recall metrics provide deeper insight: Table 2. Per-class metrics (validation set): precision, recall and F1 for “Not Won (0)” vs “Won (1)” Metric Not Won (0) Won (1) Precision 93% 33% Recall 96% 20% F1 Score 95% 25% This shows that while the model is highly accurate for the dominant class ("Not Won (0) "), it struggles to correctly identify winning quotes, which are fewer and harder to detect. The Weighted F1 Score was 89%, confirming that the model performs well overall, but with limitations on the minority class. To address this, we introduced a business-rule priority scoring matrix (cap = 5). We then tested a GB model with and without cross-validation: Table 3. Model comparison: overall accuracy and positiveclass (“Won (1)”) metrics with effect of the priority matrix and cross-validation Model Accuracy Precision Recall F1 Note (RF)-Baseline 90.0% 33.0% 20.0% 25.0% No priority matrix RF + Priority Matrix (PM) 90.5% 37.5% 20.0% 26.0% Businessrule score added GB + PM 91.6% 50.0% 13.3% 21.0% Improved precision GB + Matrix + CrossValidation 90.5% 37.5% 20.0% 26.0% Balanced reliability 4.2 Feature Importance Based on the Gini impurity index, the most informative features in the final models were Price Tier, Complexity, Days Until Due, and Total Value. (Segment was analyzed initially but removed in the final configuration after user feedback.) 5. DISCUSSION In practice, a precision-oriented operating point can be acceptable when capacity and SLA constraints penalize false negatives (missed high-value opportunities) more than false positives. While the GB model improved precision for "Won (1) " predictions (up to 50%), it came with a lower recall, meaning it identified fewer winning opportunities. Still, this trade-off can be acceptable in scenarios where the cost of missing a high-priority quota is lower than acting on a false positive. Post-deployment feedback led to two practical adjustments: Segment (anonymized A–D) did not contribute to prioritization decisions according to the team; and Price Tier categories were simplified (internal update) from six to four levels, with the scoring logic adjusted accordingly. 6. DEPLOYMENT AND LIMITATIONS 6.1 Deployment and User Adoption The final model was deployed in Excel using VBA, making the tool accessible to the Tenders & Proposals team without technical prerequisites. Since quoting data is managed in Salesforce, the next step is to integrate the tool directly into Salesforce to automate data flow and enable online use. Phase 2 – Salesforce (in progress): a native field priority score computes the business-rule score (price tier, complexity, request type, urgency, value band) directly from Salesforce data 6.2 Limitations • Class Imbalance: Won (1) quotes are a small fraction of the dataset, affecting generalization and recall. • Data Quality: Manual entry and inconsistent categories reduce model performance. • Outcome Recording: Many quotes without recorded outcomes were removed; if not missing at random, this may bias results. • No Real-Time Feed: Current Excel workflow limits scalability vs. a native Salesforce integration. Data & risk constraints: – Heterogeneous data (non-standard categories, missing values, sometimes absent outcomes). – Internal Excel entry prior to Salesforce integration. – Technical/contractual risks that may impose bid/no-bid decisions. 6.3 Summary The tool successfully classifies and ranks presales quotes based on historical data, helping the organization Tenders & Proposals team to better focus their efforts. While recall for the “Won (1) ” class remains a challenge, the model provides a strong foundation for future improvements — including integration with Salesforce. 7. Conclusion This paper details the development of a data-driven prioritization tool aimed at supporting the Tenders & Proposals team within the company. As the volume and diversity of quote requests grow, the need for a structured, repeatable prioritization process becomes critical. The proposed tool leverages data science techniques to help team members evaluate quotes not just based on intuition or urgency, but using key features such as price tier, complexity level, urgency, and total value—transformed through feature engineering and model training. Throughout the process, the model evolved via iterative improvements: • Starting with a basic RF model tested on 20% of the dataset. • Introducing a business-rule priority score (cap = 5). • Testing GB to improve prediction for “Won (1)” quotes. • Finally, implementing 5-fold cross-validation to reduce overfitting and improve generalization. Despite class imbalance (with “Won (1)” quotes being underrepresented), the tool offers a practical, actionable way to triage incoming quotes. While predictive power for “Won (1)” remains modest, the model excels at filtering lowprobability opportunities—helping to focus time and resources. The tool was built and deployed in Excel using VBA, ensuring direct usability by the team without requiring additional platforms. A quote can now be entered, scored, and acted upon—within seconds. This is not the final version. With more data, team feedback, and integration into platforms like Salesforce, the model can become a central part of the team’s quoting process. But even in this first iteration, it already brings structure, clarity, and efficiency to an important business workflow. 8.ACKNOWLEDGEMENTS/AI Declaration During the preparation of this work the author used ChatGPT (OpenAI) for language editing and structuring. 9.REFERENCES He, H. and Garcia, E.A. (2009). Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering, 21(9), 1263–1284. Mattila, L. (2023). Improving purchasing request for quotation process. Master’s Thesis, Tampere University, Faculty of Management and Business. Nederkoorn, J.B.M. (2017). Innovating the quotation process at an ETO company: a design-science approach. Master’s Thesis, Eindhoven University of Technology, School of Industrial Engineering. Saito, T. and Rehmsmeier, M. (2015). The precision–recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLOS ONE, 10(3), e0118432. Shearer, C. (2000). The CRISP-DM model: The new blueprint for data mining. Journal of Data Warehousing, 5(4), 13–22. Appendix A. “Practitioner recipe: daily triage (Tendering POV)” Steps (from a Tender POV) • Make today’s list. List all quotes you must work on today + your available hours. • Give each quote a quick score(cap = 5) Base points from the scoring matrix o Price Tier Level: ordinal 0–3 o Complexity Level: Tier 2 +0 · Tier 3 +1 · Tier 4 +2 · Tier 5 +3 o Request type: PO Submission +1 · Quote 0 · Tender & RFXs 0 · Pricing 0 o Total Value (category): Lower (0–5M) +0 · Upper (5M– 25M) +1 (Cap the sum at 5 = “Score Cap / Final score weighted”.) • Sort and assign. Sort quotes by score (highest first). Starting from the top, assign work until today’s hours are used up. • Tiebreakers (if scores are close). 1. Higher value, 2) earlier due date, 3) strategic customer. • Repeat daily. Re-score each morning (or when a due date/effort changes). You’ll always start with the best impact items first. The recipe is validated in a case study using ranking metrics and business KPIs Appendix B. Search strings Google Scholar queries: "Request for quotation" AND " (prioritization OR scoring) " AND " (manufacturing OR ETO OR MTO) ": • MattilaLaura.pdf • Nederkoorn_2017.pdf Compendex queries : "Order acceptance" AND " scheduling" AND " make-toorder" AND " prioritization scores" AND " quote" " RFQ" AND " MACHINE LEARNING" AND " QUOTES" • Explainable AI in Request-for-Quote | Compendex • Transportation spends portfolio optimization using multi-level, multicriteria decision analysis | Compendex Appendix C. Exploratory value clustering (not used for reporting) Using K-Means on Total Value, we observed two clusters with centroids $430,556.18 and $13,823,021.58. The midpoint (~$7.13M) would split the data into Lower (~$106– $7.1M) and Upper (~$7.1M–$25M). For operational simplicity and internal alignment, the main analysis uses the business-rule bands (0–$5M, $5M–$25M). Figure 1. Optimal numbers of clusters for Total Value (elbow/silhouette) The knee at k = 2 yields centroids $430,556 and $13,823,022; the midpoint (~$7.13M) is the exploration split. The main paper uses business-rule bands.