A Framework for Hybrid CRM Personalization: Combining Rule-Based Logic with Machine Learning Predictions

A Framework for Hybrid CRM Personalization: Combining Rule-Based Logic with Machine Learning Predictions | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Framework for Hybrid CRM Personalization: Combining Rule-Based Logic with Machine Learning Predictions Krishna Chaithanya Vuppala, Nithin Maruthi Prasad This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9068530/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 16 You are reading this latest preprint version Abstract Customer Relationship Management (CRM) systems are at the heart of how modern businesses connect with their customers. But despite all the data available, most CRM systems still rely on rigid segmentation rules and look backward rather than forward. In this work, we introduce a new hybrid approach to CRM personalization—one that brings together the best of rule-based business logic and machine learning (ML) predictions. The goal: deliver smarter, more personalized Next Best Action (NBA) recommendations that adapt to each customer’s context. Our framework is built on four core layers: collecting and engineering useful data, validating decisions with clear business rules, generating predictions with multiple ML models, and orchestrating final decisions that balance statistical insights with real-world business needs. Using real CRM data from a large, multi-property service company, we show how this system can predict things like customer engagement, estimate value, and group similar behaviors—then turn those insights into concrete tasks for relationship managers. In testing, our hybrid approach beat the traditional rule-only system on accuracy (AUC-ROC 0.82 vs. 0.65), acted much faster (cutting signal-to-action time by 75%), and improved customer engagement by 14%, all while keeping necessary checks and balances for enterprise use. CRM personalization Next Best Action machine learning rule-based systems hybrid decision engine predictive analytics Figures Figure 1 Figure 2 1 Introduction In industries like finance, healthcare, retail, and beyond, CRM systems are the engine that drives customer engagement. These systems pull together a goldmine of data—everything from transactions and preferences to engagement history—into clear customer profiles for the people on the front lines, such as relationship managers and account executives. In service businesses, this means tracking things like how often customers visit, what services they use, what they spend on extras, and even which channels they prefer to interact with. Still, even with all this information, most organizations struggle to turn data into truly personalized, timely actions. Relationship managers often have to comb through profiles, try to spot patterns, and decide how to reach out—all by hand, without much help from smart tools. The idea of tailoring approaches to each customer has been around for years [ 1 ], but actually making it happen at scale has proven tough. Research shows that while data analysis has made great strides in the lab [ 2 ], it hasn’t always made its way into real-world CRM systems. AI now promises to push CRM from simply recording what happened to actively suggesting what to do next [ 3 ]—the “Next Best Action” idea. This approach uses predictions from data science [ 4 ] to recommend the most impactful actions for each customer, turning insight into practical advice. Traditionally, CRMs have relied on rule-based engines: clear, if-then rules that decide who gets what offer, how often to reach out, and through which channels [ 5 ]. These rules are easy to understand and audit—critical for businesses in regulated fields. But today’s customers don’t always follow these neat rules; their behavior is complex and changes across channels and over time [ 6 ]. As a result, static rules can miss the mark when it comes to delivering a truly personal experience [ 7 ]. Machine learning is changing the game by spotting patterns and predicting what customers might do next. For example, new models can predict who’s likely to leave [ 8 ], what someone might buy [ 9 ], or which offers will catch their eye [ 10 , 11 ]—often better than the old ways of doing things. These smarter systems are already shaping everything from product recommendations to churn prevention. Combining rule-based logic and machine learning—hybrid approaches—often outperforms using just one method [ 12 ]. The best results happen when smart predictions are paired with human judgment and clear business rules [ 13 , 14 ]. Recent research also shows that mixing transparent rules with complex models not only boosts performance but also helps people understand and trust the system’s decisions [ 15 ]. There’s plenty of research on the pieces—analytics, predictions, business rules—but not much on stitching it all together into a CRM system that works day-to-day. Most efforts stop at building the models, without showing how to bring everything from data gathering to making real, actionable recommendations under one roof. This work bridges that gap, showing an end-to-end hybrid system that’s actually used in a large organization and put to the test against a range of alternatives. 2 Materials and Methods 2.1 System Architecture Overview Our hybrid CRM personalization architecture works in four layers, each playing a specific role in the journey from raw data to meaningful action. Data moves step by step—from collection, through prediction and rule checks, all the way to actual customer engagement. We designed this system to build on top of the existing CRM, not to replace or compete with it. Figure 1 shows the full setup. In practice, we built our system right inside an existing enterprise CRM. A tool called Azure Synapse Link connects the live CRM to the analytics side, copying over all the important customer data—profiles, interactions, transactions—once a day, so the main system isn’t slowed down. Then, we sort this data into three layers: bronze for raw data, silver for cleaned-up and standardized data, and gold for engineered features ready for analysis and modeling. 2.2 Data Acquisition and Feature Engineering The CRM keeps detailed records on each customer—what they use most, how often they visit, what they spend on extras, their favorite services, and how they’ve interacted with events or promotions. All this information is pulled together into easy-to-read profiles that relationship managers can access directly in the system. To make sense of all this data, we grouped customer behaviors into three main types. First, we tracked engagement patterns—like how often someone visits or how long they stay. Second, we looked at value dynamics, measuring things like how much a customer spends and what categories matter most to them. Third, we created preference signals based on past behaviors, such as favorite locations, chosen services, or responses to promotions. We measured these behaviors over different time periods—one week, one month, three months, and a full year—to spot both short-term changes and long-term trends. Table 1 summarizes the feature families. We also built new features that show, for example, whether someone mainly uses core services or spends more on extras, or if people who visit often are also the ones who jump on promotions. Table 1 Feature families, representative features, and temporal windows Feature Family Representative Features Temporal Windows Engagement Patterns Visit frequency, session duration, channel utilization, interaction recency 7d, 30d, 90d, 12m Value Dynamics Total spend, category distribution, monetary trends, average transaction value 7d, 30d, 90d, 12m Preference Signals Location affinity, service category choices, ancillary selections, promotional response rates 7d, 30d, 90d, 12m Derived Interactions Primary-to-ancillary spend ratio, visit-promotion correlation, cross-category engagement index 30d, 90d We made sure data quality stays high at every step. Automated checks spot missing or unusual values and catch any unexpected changes. If something looks off, those records are set aside for review, so only reliable data gets used. 2.3 Rule-Based Decision Layer This layer acts as the rule enforcer. It doesn’t create recommendations from scratch—instead, it double-checks every suggestion from the ML models, making sure each one follows business rules, complies with regulations, and fits within operational limits. There are four kinds of rules at play. Eligibility rules decide if a customer can get a specific offer or action, taking into account things like account status, loyalty tier, location, and legal requirements. Contact governance rules make sure customers aren’t overwhelmed—limiting how often they’re reached out to, honoring their channel choices, and blocking contact during certain times. Conflict resolution rules help sort things out when a customer qualifies for more than one recommendation at once, setting clear priorities. Finally, capacity rules keep things realistic by making sure only recommendations that can actually be delivered—given staff, event space, or inventory—are sent out. By keeping rule checks separate from the prediction models, it’s easy to update the models without changing the business logic. This also makes audits simpler, since all the rules are spelled out and easy to review on their own. 2.4 Machine Learning Prediction Layer Instead of depending on a single model, we use several types working together, each focusing on different aspects of what customers might do. Some models predict clear yes-or-no outcomes, like whether someone will respond to an offer or might become less active. We mainly use a technique called gradient-boosted trees (XGBoost) because it works well with our data. Other models, like random forests and logistic regression, help us compare results and make our predictions even stronger. Some of our models handle numbers—like how much a customer might spend, how often they’ll visit, or their potential value over time. We use special techniques to make sure unusual spending habits don’t throw off our predictions. Other models look for hidden groups among customers by finding those with similar behaviors, instead of just using predefined categories. We also use models that track changes over time, to spot trends or sudden shifts that might need attention. This multi-model approach is supported by recent research in the field [ 16 , 17 ]. 2.5 Model Training and Validation Protocol We used two years’ worth of CRM data from about 48,000 customers. The first year and a half was used to train and validate our models, making sure everything worked as expected. The last six months were held back as a test, so we could check how well the models would do in predicting new, unseen behavior—just like they would in the real world. For the engagement model, we used all 127 features we created and set it up to predict the chance that a customer would engage. We also built a model to estimate how much a customer might be worth in the future, using methods to handle unusual data points. To group customers, we picked the best number of clusters (k = 7) by checking both math and expert opinion. To spot customers who might leave, we used a mix of risk prediction and checks for odd changes over time. We only put models into use if they hit our set performance targets and worked consistently well for different kinds of customers. We also checked to make sure our predictions weren’t biased against any group. 2.6 Decision Orchestration and Execution This stage is where smart predictions meet business rules to deliver a final, personalized recommendation, following a clear four-step process (see Fig. 2 ). First, the ML models come up with a ranked list of possible actions for each customer, showing how confident the system is and why it made those choices. Next, every suggestion goes through all the business rules—anything that doesn’t qualify gets dropped. If more than one valid recommendation is left, the system weighs which matters most, based on both predicted impact and business priorities. Finally, the top recommendations become ready-to-use tasks for relationship managers, complete with helpful context and talking points. After an action is taken and the customer responds—by engaging, converting, ignoring, or opting out—the results flow back into the analytics pipeline. This way, the models keep learning and adjusting as customer behavior changes. 3 Experimental Design 3.1 Configurations We tested the system in four different ways to see what each part brings to the table. First, we tried just the rule-based logic (Configuration A). Next, we used only the machine learning predictions and skipped the rules (Configuration B). Then, we combined them—letting the ML models make suggestions that are filtered by rules (Configuration C). Finally, we ran the full hybrid system with all parts working together, including prioritization and feedback loops (Configuration D). We also compared against other common methods like collaborative filtering with alternating least squares (CF-ALS) and RFM scoring [ 11 ]. 3.2 Evaluation Metrics We measured performance in four main areas. For prediction quality, we looked at how well the models separated positive and negative outcomes (AUC-ROC, precision, recall, F1), how accurate they were (MAE, RMSE), and whether predictions matched real results (calibration). To judge how relevant recommendations were, we tracked how often suggested actions led to engagement (hit rate), how well the system ranked the best options (nDCG), and how much variety it offered (Shannon entropy). We also measured how quickly actions happened (time-to-action), how many tasks got done (task completion rate), and how often recommendations were changed by agents (override rate). Finally, we looked at the real business impact—did engagement and revenue go up? We used paired t-tests with Bonferroni correction (α = 0.05) and report Cohen’s d effect sizes. 4 Results Table 2 shows how well each setup performed, side by side. Table 3 gives a closer look at how accurate the engagement prediction model was. Table 2 Performance comparison across configurations and baselines on the 6-month holdout test set Metric A (Rule) B (ML) C (Seq.) D (Full) CF-ALS RFM AUC-ROC 0.65 0.80 0.80 0.82 0.72 0.61 Precision 0.51 0.62 0.63 0.67 0.55 0.48 Recall 0.44 0.71 0.69 0.74 0.60 0.40 F1 Score 0.47 0.66 0.66 0.70 0.57 0.44 Hit Rate 0.29 0.36 0.38 0.41 0.31 0.26 nDCG@5 0.52 0.68 0.71 0.76 0.59 0.49 Shannon Entropy 1.9 2.5 2.6 2.8 2.1 1.7 Time-to-Action (hrs) 72 16 20 18 — — Task Completion % 52% 60% 64% 68% — — Engagement Lift Baseline + 9% + 11% + 14%* + 4% + 1% Revenue Lift/Cust. Baseline + 5% + 6% + 8%* + 2% + 0.5% CF-ALS = Collaborative Filtering with Alternating Least Squares; RFM = Recency-Frequency-Monetary scoring. Bold values indicate best performance. * p 0.5). Table 3 Engagement propensity model: detailed prediction quality on holdout set Metric XGBoost (Primary) Logistic Regression (Benchmark) AUC-ROC 0.82 (95% CI: 0.79–0.85) 0.79 (95% CI: 0.76–0.82) Brier Score 0.18 0.21 Precision (thresh. 0.5) 0.67 0.61 Recall (thresh. 0.5) 0.74 0.70 F1 (thresh. 0.5) 0.70 0.65 Log Loss 0.42 0.47 CI = Confidence Interval. Threshold optimized via Youden’s J statistic on validation fold. 4.1 Prediction Quality Our main engagement model hit an AUC-ROC of 0.82 (95% CI: 0.79–0.85) on the test set—a clear step up from the rule-based approach at 0.65 and the simpler logistic regression at 0.79. The biggest improvement came in the tricky middle ground—cases where it’s not obvious what a customer will do. The model’s predictions closely matched real outcomes (Brier score 0.18), showing it’s well calibrated. 4.2 Recommendation Relevance The full hybrid system got a hit rate of 0.41—meaning 41% of its recommendations led to positive engagement—compared to 29% for the rule-only setup and 38% for the sequential hybrid. Even when the difference looks small, the hybrid consistently puts the most important actions at the top of the list (nDCG 0.76 vs. 0.71). It also offered a wider variety of engagement strategies (Shannon entropy 2.8 vs. 1.9), avoiding the trap of always recommending the same thing. 4.3 Operational Efficiency Actions happened much faster with the hybrid system: the average time from signal to action was cut from 72 hours down to just 18—a 75% improvement (p < 0.001, Cohen’s d = 1.8). More of these recommended tasks actually got done too, with completion rates jumping from 52% to 68% (p < 0.01, d = 0.6). Relationship managers found the personalized, data-backed tasks easier to act on than generic or manual assignments. 4.4 Business Impact Customers who received hybrid-generated actions engaged 14% more than those in the control group over six months (p < 0.01, d = 0.5). Revenue per customer also went up by about 8% for those who got the personalized recommendations. While these results look strong, we made sure to carefully compare similar groups (stratified by loyalty tier, visit frequency, and baseline spend) and used a long enough window to make the findings reliable. 5 Discussion But the real change isn’t just in the numbers—it’s in how the CRM is used. Instead of simply storing information, the system now helps guide the next steps for each customer. Relationship managers get helpful, data-driven advice, not just raw data, making it easier to turn information into action. Because recommended actions are automatically created and assigned to the right relationship manager—complete with helpful context—data insights actually reach the people who need them, right when they’re needed. This boosts decision quality everywhere, without adding extra manual work, and still leaves room for personal expertise and judgment. The system keeps governance and compliance strong, since every ML recommendation is checked against clear rules before reaching relationship managers. Business teams can review and update these rules themselves, without having to be data scientists. This is especially important in regulated industries, where transparency and audit trails really matter. The approach backs up the idea that AI should help people do their jobs [ 14 ], not replace them, and that combining transparent and complex systems builds trust [ 15 ]. 5.1 Limitations There are a few things to keep in mind. Our results come from one large, multi-property company, so numbers might look different elsewhere. The overall system design should work for other types of businesses, but that still needs to be tested. Because data updates happen once a day, this setup isn’t ideal for situations that need up-to-the-minute responses—future work could look at faster, real-time options. We also haven’t studied long-term changes or model drift over a year or more, and our recommendations focus on individuals, not balancing resources across a whole customer portfolio. Finally, there’s room to improve how we explain recommendations to managers, for instance through SHAP (SHapley Additive exPlanations) values, so everyone understands why a certain action was suggested. 6 Conclusion We set out to build a practical system that combines business rules with machine learning to deliver smart, personalized recommendations in CRM. Our four-layer approach helps organizations turn raw customer data into tailored engagement strategies that are accurate, compliant, and realistic to put into action. Testing on real CRM data showed that this hybrid approach beats rule-only systems on every important front: it’s more accurate (AUC-ROC 0.82 vs. 0.65), gives more relevant recommendations (hit rate 0.41 vs. 0.29), speeds up the process (75% latency reduction), and boosts customer engagement by 14%. The system fits right into existing CRMs and uses tools that staff already understand. At its heart, this work shows that business rules and predictive intelligence work best when used together. For organizations with rich customer data, this kind of system offers a real way to deliver personalized experiences at scale—and truly unlock the value of all that information. Declarations Competing Interests The authors declare no competing interests. Ethics Approval Not applicable. Funding Not applicable. Author Contribution Krishna Chaithanya Vuppala conceived the framework design, developed the architecture, conducted the experimental analysis, and wrote the manuscript. Nithin Maruthi Prasad contributed to the technical review of the ML pipeline design and provided critical feedback on the manuscript. Data Availability The datasets generated and analyzed during the current study are not publicly available due to containing proprietary enterprise CRM data. References Peppers, D., Rogers, M.: The One to One Future: Building Relationships One Customer at a Time. Currency Doubleday, New York (1993) Ngai, E.W.T., Xiu, L., Chau, D.C.K.: Application of data mining techniques in customer relationship management: a literature review and classification. Expert Syst. Appl. 36(2), 2592–2602 (2009) Huang, M.H., Rust, R.T.: A strategic framework for artificial intelligence in marketing. J. Acad. Mark. Sci. 49(1), 30–50 (2021) Provost, F., Fawcett, T.: Data Science for Business: What You Need to Know About Data Mining and Data-Analytic Thinking. O’Reilly Media, Sebastopol (2013) Buttle, F., Maklan, S.: Customer Relationship Management: Concepts and Technologies, 4th edn. Routledge, London (2019) Verhoef, P.C., Broekhuizen, T., Bart, Y., et al.: Digital transformation: a multidisciplinary reflection and research agenda. J. Bus. Res. 122, 889–901 (2021) Lemon, K.N., Verhoef, P.C.: Understanding customer experience throughout the customer journey. J. Mark. 80(6), 69–96 (2016) Coussement, K., Van den Poel, D.: Churn prediction in subscription services: an application of support vector machines while comparing two parameter-selection techniques. Expert Syst. Appl. 34(1), 313–327 (2008) De Caigny, A., Coussement, K., De Bock, K.W.: A new hybrid classification algorithm for customer churn prediction based on logistic regression and decision trees. Eur. J. Oper. Res. 269(2), 760–772 (2018) Ricci, F., Rokach, L., Shapira, B.: Recommender Systems Handbook, 2nd edn. Springer, New York (2015) Fader, P.S., Hardie, B.G.S., Lee, K.L.: RFM and CLV: using iso-value curves for customer base analysis. J. Mark. Res. 42(4), 415–430 (2005) Adomavicius, G., Tuzhilin, A.: Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions. IEEE Trans. Knowl. Data Eng. 17(6), 734–749 (2005) Brynjolfsson, E., McElheran, K.: The rapid adoption of data-driven decision-making. Am. Econ. Rev. 106(5), 133–139 (2016) Davenport, T.H.: The AI Advantage: How to Put the Artificial Intelligence Revolution to Work. MIT Press, Cambridge (2018) Arrieta, A.B., Díaz-Rodríguez, N., Del Ser, J., et al.: Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI. Inf. Fusion 58, 82–115 (2020) Kansal, T., Mahor, V., Garg, P.: Customer segmentation using k-means and hierarchical clustering for CRM analytics. Int. J. Data Sci. Anal. 15(2), 189–203 (2023) Chen, J., Ma, T., Liu, Q.: Neural collaborative filtering for customer product recommendations in enterprise CRM. Expert Syst. Appl. 159, 113577 (2020) Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 19 Mar, 2026 Reviews received at journal 18 Mar, 2026 Reviews received at journal 17 Mar, 2026 Reviewers agreed at journal 17 Mar, 2026 Reviewers agreed at journal 17 Mar, 2026 Reviewers agreed at journal 16 Mar, 2026 Reviews received at journal 16 Mar, 2026 Reviewers agreed at journal 15 Mar, 2026 Reviewers agreed at journal 15 Mar, 2026 Reviews received at journal 15 Mar, 2026 Reviewers agreed at journal 15 Mar, 2026 Reviewers agreed at journal 15 Mar, 2026 Reviewers invited by journal 15 Mar, 2026 Editor assigned by journal 10 Mar, 2026 Submission checks completed at journal 10 Mar, 2026 First submitted to journal 09 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9068530","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":607232371,"identity":"0b6bb6bc-de82-44e9-b679-f4b33f644665","order_by":0,"name":"Krishna Chaithanya Vuppala","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABA0lEQVRIiWNgGAWjYBACCQbGBmYom/FBQgWQYmZuIFoLs8GHMyCKkZAWoBoom01yZhvYNvxaJGcfbv5cuMPOnn9G+jNp3nm10fztQC0/Krbh1CLNl9gmPfNMcuKMGwnJ1rzbjufOOMzYwNhz5jZOLXI8jG3MvG3MCQw3Eg7e5t12LLcBqIWZsQ2vlubPvG319vI3Ehukeeccy51PSIs0DyNQZdthxg03kpkkZzbU5G4gpEWyh7ENqOV44sYzz4CBfOxA7kagloP4/CJxhv0x0GHV9nLH0x8+SKipy513/vDBBz8qcGtBAIEEEHkYzD5AhHog4AerqyNO8SgYBaNgFIwoAAAc3VuztDGgzgAAAABJRU5ErkJggg==","orcid":"","institution":"Texas A\u0026M University – Commerce","correspondingAuthor":true,"prefix":"","firstName":"Krishna","middleName":"Chaithanya","lastName":"Vuppala","suffix":""},{"id":607232374,"identity":"af0dbd01-c26f-477c-b484-2a896104e3c6","order_by":1,"name":"Nithin Maruthi Prasad","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Nithin","middleName":"Maruthi","lastName":"Prasad","suffix":""}],"badges":[],"createdAt":"2026-03-09 05:23:35","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9068530/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9068530/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104838598,"identity":"98eb8f9c-d44b-4dda-81fa-e9eab5779ff7","added_by":"auto","created_at":"2026-03-17 18:28:38","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":26916,"visible":true,"origin":"","legend":"\u003cp\u003eFour-layer hybrid CRM personalization architecture. Layer 1 handles data acquisition and feature engineering; Layer 2 deploys multi-model ML predictions; Layer 3 applies deterministic rule-based validation; Layer 4 orchestrates decisions into actionable CRM tasks. A feedback loop enables continuous model adaptation.\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9068530/v1/639095e756092732c4c42015.jpg"},{"id":104838597,"identity":"76e10dd5-473d-4acd-abb0-9423c659e31c","added_by":"auto","created_at":"2026-03-17 18:28:38","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":15814,"visible":true,"origin":"","legend":"\u003cp\u003eDecision orchestration pipeline showing the four-stage process from ML model outputs through candidate generation, rule validation, conflict resolution, and action materialization into CRM-native tasks\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9068530/v1/03de7f5d149c3b3797c126ef.jpg"},{"id":105033752,"identity":"c8bb4e65-1000-420d-b46c-1f840d7b5acd","added_by":"auto","created_at":"2026-03-20 07:21:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":798467,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9068530/v1/6409cd80-3cbb-4009-ba9c-e2c46c56f9d8.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Framework for Hybrid CRM Personalization: Combining Rule-Based Logic with Machine Learning Predictions","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eIn industries like finance, healthcare, retail, and beyond, CRM systems are the engine that drives customer engagement. These systems pull together a goldmine of data\u0026mdash;everything from transactions and preferences to engagement history\u0026mdash;into clear customer profiles for the people on the front lines, such as relationship managers and account executives. In service businesses, this means tracking things like how often customers visit, what services they use, what they spend on extras, and even which channels they prefer to interact with.\u003c/p\u003e \u003cp\u003eStill, even with all this information, most organizations struggle to turn data into truly personalized, timely actions. Relationship managers often have to comb through profiles, try to spot patterns, and decide how to reach out\u0026mdash;all by hand, without much help from smart tools. The idea of tailoring approaches to each customer has been around for years [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], but actually making it happen at scale has proven tough.\u003c/p\u003e \u003cp\u003eResearch shows that while data analysis has made great strides in the lab [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], it hasn\u0026rsquo;t always made its way into real-world CRM systems. AI now promises to push CRM from simply recording what happened to actively suggesting what to do next [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u0026mdash;the \u0026ldquo;Next Best Action\u0026rdquo; idea. This approach uses predictions from data science [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] to recommend the most impactful actions for each customer, turning insight into practical advice.\u003c/p\u003e \u003cp\u003eTraditionally, CRMs have relied on rule-based engines: clear, if-then rules that decide who gets what offer, how often to reach out, and through which channels [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. These rules are easy to understand and audit\u0026mdash;critical for businesses in regulated fields. But today\u0026rsquo;s customers don\u0026rsquo;t always follow these neat rules; their behavior is complex and changes across channels and over time [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. As a result, static rules can miss the mark when it comes to delivering a truly personal experience [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMachine learning is changing the game by spotting patterns and predicting what customers might do next. For example, new models can predict who\u0026rsquo;s likely to leave [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], what someone might buy [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], or which offers will catch their eye [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u0026mdash;often better than the old ways of doing things. These smarter systems are already shaping everything from product recommendations to churn prevention.\u003c/p\u003e \u003cp\u003eCombining rule-based logic and machine learning\u0026mdash;hybrid approaches\u0026mdash;often outperforms using just one method [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. The best results happen when smart predictions are paired with human judgment and clear business rules [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Recent research also shows that mixing transparent rules with complex models not only boosts performance but also helps people understand and trust the system\u0026rsquo;s decisions [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThere\u0026rsquo;s plenty of research on the pieces\u0026mdash;analytics, predictions, business rules\u0026mdash;but not much on stitching it all together into a CRM system that works day-to-day. Most efforts stop at building the models, without showing how to bring everything from data gathering to making real, actionable recommendations under one roof. This work bridges that gap, showing an end-to-end hybrid system that\u0026rsquo;s actually used in a large organization and put to the test against a range of alternatives.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 System Architecture Overview\u003c/h2\u003e \u003cp\u003eOur hybrid CRM personalization architecture works in four layers, each playing a specific role in the journey from raw data to meaningful action. Data moves step by step\u0026mdash;from collection, through prediction and rule checks, all the way to actual customer engagement. We designed this system to build on top of the existing CRM, not to replace or compete with it. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the full setup.\u003c/p\u003e \u003cp\u003eIn practice, we built our system right inside an existing enterprise CRM. A tool called Azure Synapse Link connects the live CRM to the analytics side, copying over all the important customer data\u0026mdash;profiles, interactions, transactions\u0026mdash;once a day, so the main system isn\u0026rsquo;t slowed down. Then, we sort this data into three layers: bronze for raw data, silver for cleaned-up and standardized data, and gold for engineered features ready for analysis and modeling.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Data Acquisition and Feature Engineering\u003c/h2\u003e \u003cp\u003eThe CRM keeps detailed records on each customer\u0026mdash;what they use most, how often they visit, what they spend on extras, their favorite services, and how they\u0026rsquo;ve interacted with events or promotions. All this information is pulled together into easy-to-read profiles that relationship managers can access directly in the system.\u003c/p\u003e \u003cp\u003eTo make sense of all this data, we grouped customer behaviors into three main types. First, we tracked engagement patterns\u0026mdash;like how often someone visits or how long they stay. Second, we looked at value dynamics, measuring things like how much a customer spends and what categories matter most to them. Third, we created preference signals based on past behaviors, such as favorite locations, chosen services, or responses to promotions.\u003c/p\u003e \u003cp\u003eWe measured these behaviors over different time periods\u0026mdash;one week, one month, three months, and a full year\u0026mdash;to spot both short-term changes and long-term trends. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes the feature families. We also built new features that show, for example, whether someone mainly uses core services or spends more on extras, or if people who visit often are also the ones who jump on promotions.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eFeature families, representative features, and temporal windows\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature Family\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRepresentative Features\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTemporal Windows\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEngagement Patterns\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVisit frequency, session duration, channel utilization, interaction recency\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7d, 30d, 90d, 12m\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValue Dynamics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal spend, category distribution, monetary trends, average transaction value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7d, 30d, 90d, 12m\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePreference Signals\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLocation affinity, service category choices, ancillary selections, promotional response rates\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7d, 30d, 90d, 12m\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDerived Interactions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrimary-to-ancillary spend ratio, visit-promotion correlation, cross-category engagement index\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30d, 90d\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eWe made sure data quality stays high at every step. Automated checks spot missing or unusual values and catch any unexpected changes. If something looks off, those records are set aside for review, so only reliable data gets used.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Rule-Based Decision Layer\u003c/h2\u003e \u003cp\u003eThis layer acts as the rule enforcer. It doesn\u0026rsquo;t create recommendations from scratch\u0026mdash;instead, it double-checks every suggestion from the ML models, making sure each one follows business rules, complies with regulations, and fits within operational limits.\u003c/p\u003e \u003cp\u003eThere are four kinds of rules at play. Eligibility rules decide if a customer can get a specific offer or action, taking into account things like account status, loyalty tier, location, and legal requirements. Contact governance rules make sure customers aren\u0026rsquo;t overwhelmed\u0026mdash;limiting how often they\u0026rsquo;re reached out to, honoring their channel choices, and blocking contact during certain times. Conflict resolution rules help sort things out when a customer qualifies for more than one recommendation at once, setting clear priorities. Finally, capacity rules keep things realistic by making sure only recommendations that can actually be delivered\u0026mdash;given staff, event space, or inventory\u0026mdash;are sent out.\u003c/p\u003e \u003cp\u003eBy keeping rule checks separate from the prediction models, it\u0026rsquo;s easy to update the models without changing the business logic. This also makes audits simpler, since all the rules are spelled out and easy to review on their own.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Machine Learning Prediction Layer\u003c/h2\u003e \u003cp\u003eInstead of depending on a single model, we use several types working together, each focusing on different aspects of what customers might do. Some models predict clear yes-or-no outcomes, like whether someone will respond to an offer or might become less active. We mainly use a technique called gradient-boosted trees (XGBoost) because it works well with our data. Other models, like random forests and logistic regression, help us compare results and make our predictions even stronger.\u003c/p\u003e \u003cp\u003eSome of our models handle numbers\u0026mdash;like how much a customer might spend, how often they\u0026rsquo;ll visit, or their potential value over time. We use special techniques to make sure unusual spending habits don\u0026rsquo;t throw off our predictions. Other models look for hidden groups among customers by finding those with similar behaviors, instead of just using predefined categories. We also use models that track changes over time, to spot trends or sudden shifts that might need attention. This multi-model approach is supported by recent research in the field [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Model Training and Validation Protocol\u003c/h2\u003e \u003cp\u003eWe used two years\u0026rsquo; worth of CRM data from about 48,000 customers. The first year and a half was used to train and validate our models, making sure everything worked as expected. The last six months were held back as a test, so we could check how well the models would do in predicting new, unseen behavior\u0026mdash;just like they would in the real world.\u003c/p\u003e \u003cp\u003eFor the engagement model, we used all 127 features we created and set it up to predict the chance that a customer would engage. We also built a model to estimate how much a customer might be worth in the future, using methods to handle unusual data points. To group customers, we picked the best number of clusters (k\u0026thinsp;=\u0026thinsp;7) by checking both math and expert opinion. To spot customers who might leave, we used a mix of risk prediction and checks for odd changes over time.\u003c/p\u003e \u003cp\u003eWe only put models into use if they hit our set performance targets and worked consistently well for different kinds of customers. We also checked to make sure our predictions weren\u0026rsquo;t biased against any group.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Decision Orchestration and Execution\u003c/h2\u003e \u003cp\u003eThis stage is where smart predictions meet business rules to deliver a final, personalized recommendation, following a clear four-step process (see Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFirst, the ML models come up with a ranked list of possible actions for each customer, showing how confident the system is and why it made those choices. Next, every suggestion goes through all the business rules\u0026mdash;anything that doesn\u0026rsquo;t qualify gets dropped. If more than one valid recommendation is left, the system weighs which matters most, based on both predicted impact and business priorities. Finally, the top recommendations become ready-to-use tasks for relationship managers, complete with helpful context and talking points.\u003c/p\u003e \u003cp\u003eAfter an action is taken and the customer responds\u0026mdash;by engaging, converting, ignoring, or opting out\u0026mdash;the results flow back into the analytics pipeline. This way, the models keep learning and adjusting as customer behavior changes.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Experimental Design","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Configurations\u003c/h2\u003e \u003cp\u003eWe tested the system in four different ways to see what each part brings to the table. First, we tried just the rule-based logic (Configuration A). Next, we used only the machine learning predictions and skipped the rules (Configuration B). Then, we combined them\u0026mdash;letting the ML models make suggestions that are filtered by rules (Configuration C). Finally, we ran the full hybrid system with all parts working together, including prioritization and feedback loops (Configuration D). We also compared against other common methods like collaborative filtering with alternating least squares (CF-ALS) and RFM scoring [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eWe measured performance in four main areas. For prediction quality, we looked at how well the models separated positive and negative outcomes (AUC-ROC, precision, recall, F1), how accurate they were (MAE, RMSE), and whether predictions matched real results (calibration). To judge how relevant recommendations were, we tracked how often suggested actions led to engagement (hit rate), how well the system ranked the best options (nDCG), and how much variety it offered (Shannon entropy). We also measured how quickly actions happened (time-to-action), how many tasks got done (task completion rate), and how often recommendations were changed by agents (override rate). Finally, we looked at the real business impact\u0026mdash;did engagement and revenue go up? We used paired t-tests with Bonferroni correction (α\u0026thinsp;=\u0026thinsp;0.05) and report Cohen\u0026rsquo;s d effect sizes.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Results","content":"\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows how well each setup performed, side by side. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e gives a closer look at how accurate the engagement prediction model was.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance comparison across configurations and baselines on the 6-month holdout test set\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eA (Rule)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eB (ML)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eC (Seq.)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eD (Full)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCF-ALS\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRFM\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAUC-ROC\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.82\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePrecision\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.67\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRecall\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.74\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.40\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eF1 Score\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.70\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eHit Rate\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.41\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.26\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003enDCG@5\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.76\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.49\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eShannon Entropy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e2.8\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTime-to-Action (hrs)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e18\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTask Completion %\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e52%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e60%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e64%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e68%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eEngagement Lift\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBaseline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e+\u0026thinsp;9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;11%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e+\u0026thinsp;14%*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e+\u0026thinsp;1%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRevenue Lift/Cust.\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBaseline\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e+\u0026thinsp;5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e+\u0026thinsp;6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e+\u0026thinsp;8%*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e+\u0026thinsp;2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e+\u0026thinsp;0.5%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCF-ALS\u0026thinsp;=\u0026thinsp;Collaborative Filtering with Alternating Least Squares; RFM\u0026thinsp;=\u0026thinsp;Recency-Frequency-Monetary scoring. Bold values indicate best performance. * p\u0026thinsp;\u0026lt;\u0026thinsp;0.01 vs. Configuration A (paired t-test with Bonferroni correction, Cohen\u0026rsquo;s d\u0026thinsp;\u0026gt;\u0026thinsp;0.5).\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEngagement propensity model: detailed prediction quality on holdout set\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eXGBoost (Primary)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLogistic Regression (Benchmark)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAUC-ROC\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.82 (95% CI: 0.79\u0026ndash;0.85)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.79 (95% CI: 0.76\u0026ndash;0.82)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eBrier Score\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.21\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePrecision (thresh. 0.5)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRecall (thresh. 0.5)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eF1 (thresh. 0.5)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLog Loss\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.47\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eCI\u0026thinsp;=\u0026thinsp;Confidence Interval. Threshold optimized via Youden\u0026rsquo;s J statistic on validation fold.\u003c/em\u003e \u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Prediction Quality\u003c/h2\u003e \u003cp\u003eOur main engagement model hit an AUC-ROC of 0.82 (95% CI: 0.79\u0026ndash;0.85) on the test set\u0026mdash;a clear step up from the rule-based approach at 0.65 and the simpler logistic regression at 0.79. The biggest improvement came in the tricky middle ground\u0026mdash;cases where it\u0026rsquo;s not obvious what a customer will do. The model\u0026rsquo;s predictions closely matched real outcomes (Brier score 0.18), showing it\u0026rsquo;s well calibrated.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Recommendation Relevance\u003c/h2\u003e \u003cp\u003eThe full hybrid system got a hit rate of 0.41\u0026mdash;meaning 41% of its recommendations led to positive engagement\u0026mdash;compared to 29% for the rule-only setup and 38% for the sequential hybrid. Even when the difference looks small, the hybrid consistently puts the most important actions at the top of the list (nDCG 0.76 vs. 0.71). It also offered a wider variety of engagement strategies (Shannon entropy 2.8 vs. 1.9), avoiding the trap of always recommending the same thing.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Operational Efficiency\u003c/h2\u003e \u003cp\u003eActions happened much faster with the hybrid system: the average time from signal to action was cut from 72 hours down to just 18\u0026mdash;a 75% improvement (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001, Cohen\u0026rsquo;s d\u0026thinsp;=\u0026thinsp;1.8). More of these recommended tasks actually got done too, with completion rates jumping from 52% to 68% (p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, d\u0026thinsp;=\u0026thinsp;0.6). Relationship managers found the personalized, data-backed tasks easier to act on than generic or manual assignments.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Business Impact\u003c/h2\u003e \u003cp\u003eCustomers who received hybrid-generated actions engaged 14% more than those in the control group over six months (p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, d\u0026thinsp;=\u0026thinsp;0.5). Revenue per customer also went up by about 8% for those who got the personalized recommendations. While these results look strong, we made sure to carefully compare similar groups (stratified by loyalty tier, visit frequency, and baseline spend) and used a long enough window to make the findings reliable.\u003c/p\u003e \u003c/div\u003e"},{"header":"5 Discussion","content":"\u003cp\u003eBut the real change isn\u0026rsquo;t just in the numbers\u0026mdash;it\u0026rsquo;s in how the CRM is used. Instead of simply storing information, the system now helps guide the next steps for each customer. Relationship managers get helpful, data-driven advice, not just raw data, making it easier to turn information into action.\u003c/p\u003e \u003cp\u003eBecause recommended actions are automatically created and assigned to the right relationship manager\u0026mdash;complete with helpful context\u0026mdash;data insights actually reach the people who need them, right when they\u0026rsquo;re needed. This boosts decision quality everywhere, without adding extra manual work, and still leaves room for personal expertise and judgment.\u003c/p\u003e \u003cp\u003eThe system keeps governance and compliance strong, since every ML recommendation is checked against clear rules before reaching relationship managers. Business teams can review and update these rules themselves, without having to be data scientists. This is especially important in regulated industries, where transparency and audit trails really matter. The approach backs up the idea that AI should help people do their jobs [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], not replace them, and that combining transparent and complex systems builds trust [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Limitations\u003c/h2\u003e \u003cp\u003eThere are a few things to keep in mind. Our results come from one large, multi-property company, so numbers might look different elsewhere. The overall system design should work for other types of businesses, but that still needs to be tested. Because data updates happen once a day, this setup isn\u0026rsquo;t ideal for situations that need up-to-the-minute responses\u0026mdash;future work could look at faster, real-time options. We also haven\u0026rsquo;t studied long-term changes or model drift over a year or more, and our recommendations focus on individuals, not balancing resources across a whole customer portfolio. Finally, there\u0026rsquo;s room to improve how we explain recommendations to managers, for instance through SHAP (SHapley Additive exPlanations) values, so everyone understands why a certain action was suggested.\u003c/p\u003e \u003c/div\u003e"},{"header":"6 Conclusion","content":"\u003cp\u003eWe set out to build a practical system that combines business rules with machine learning to deliver smart, personalized recommendations in CRM. Our four-layer approach helps organizations turn raw customer data into tailored engagement strategies that are accurate, compliant, and realistic to put into action.\u003c/p\u003e \u003cp\u003eTesting on real CRM data showed that this hybrid approach beats rule-only systems on every important front: it\u0026rsquo;s more accurate (AUC-ROC 0.82 vs. 0.65), gives more relevant recommendations (hit rate 0.41 vs. 0.29), speeds up the process (75% latency reduction), and boosts customer engagement by 14%. The system fits right into existing CRMs and uses tools that staff already understand.\u003c/p\u003e \u003cp\u003eAt its heart, this work shows that business rules and predictive intelligence work best when used together. For organizations with rich customer data, this kind of system offers a real way to deliver personalized experiences at scale\u0026mdash;and truly unlock the value of all that information.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting Interests\u003c/h2\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eEthics Approval\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eNot applicable.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eKrishna Chaithanya Vuppala conceived the framework design, developed the architecture, conducted the experimental analysis, and wrote the manuscript. Nithin Maruthi Prasad contributed to the technical review of the ML pipeline design and provided critical feedback on the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated and analyzed during the current study are not publicly available due to containing proprietary enterprise CRM data.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003ePeppers, D., Rogers, M.: The One to One Future: Building Relationships One Customer at a Time. Currency Doubleday, New York (1993)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNgai, E.W.T., Xiu, L., Chau, D.C.K.: Application of data mining techniques in customer relationship management: a literature review and classification. Expert Syst. Appl. 36(2), 2592\u0026ndash;2602 (2009)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang, M.H., Rust, R.T.: A strategic framework for artificial intelligence in marketing. J. Acad. Mark. Sci. 49(1), 30\u0026ndash;50 (2021)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eProvost, F., Fawcett, T.: Data Science for Business: What You Need to Know About Data Mining and Data-Analytic Thinking. O\u0026rsquo;Reilly Media, Sebastopol (2013)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eButtle, F., Maklan, S.: Customer Relationship Management: Concepts and Technologies, 4th edn. Routledge, London (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerhoef, P.C., Broekhuizen, T., Bart, Y., et al.: Digital transformation: a multidisciplinary reflection and research agenda. J. Bus. Res. 122, 889\u0026ndash;901 (2021)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLemon, K.N., Verhoef, P.C.: Understanding customer experience throughout the customer journey. J. Mark. 80(6), 69\u0026ndash;96 (2016)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCoussement, K., Van den Poel, D.: Churn prediction in subscription services: an application of support vector machines while comparing two parameter-selection techniques. Expert Syst. Appl. 34(1), 313\u0026ndash;327 (2008)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Caigny, A., Coussement, K., De Bock, K.W.: A new hybrid classification algorithm for customer churn prediction based on logistic regression and decision trees. Eur. J. Oper. Res. 269(2), 760\u0026ndash;772 (2018)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRicci, F., Rokach, L., Shapira, B.: Recommender Systems Handbook, 2nd edn. Springer, New York (2015)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFader, P.S., Hardie, B.G.S., Lee, K.L.: RFM and CLV: using iso-value curves for customer base analysis. J. Mark. Res. 42(4), 415\u0026ndash;430 (2005)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAdomavicius, G., Tuzhilin, A.: Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions. IEEE Trans. Knowl. Data Eng. 17(6), 734\u0026ndash;749 (2005)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrynjolfsson, E., McElheran, K.: The rapid adoption of data-driven decision-making. Am. Econ. Rev. 106(5), 133\u0026ndash;139 (2016)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavenport, T.H.: The AI Advantage: How to Put the Artificial Intelligence Revolution to Work. MIT Press, Cambridge (2018)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArrieta, A.B., D\u0026iacute;az-Rodr\u0026iacute;guez, N., Del Ser, J., et al.: Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI. Inf. Fusion 58, 82\u0026ndash;115 (2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKansal, T., Mahor, V., Garg, P.: Customer segmentation using k-means and hierarchical clustering for CRM analytics. Int. J. Data Sci. Anal. 15(2), 189\u0026ndash;203 (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, J., Ma, T., Liu, Q.: Neural collaborative filtering for customer product recommendations in enterprise CRM. Expert Syst. Appl. 159, 113577 (2020)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"international-journal-of-data-science-and-analytics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jdsa","sideBox":"Learn more about [International Journal of Data Science and Analytics](http://link.springer.com/journal/41060)","snPcode":"41060","submissionUrl":"https://submission.nature.com/new-submission/41060/3","title":"International Journal of Data Science and Analytics","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"CRM personalization, Next Best Action, machine learning, rule-based systems, hybrid decision engine, predictive analytics","lastPublishedDoi":"10.21203/rs.3.rs-9068530/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9068530/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCustomer Relationship Management (CRM) systems are at the heart of how modern businesses connect with their customers. But despite all the data available, most CRM systems still rely on rigid segmentation rules and look backward rather than forward. In this work, we introduce a new hybrid approach to CRM personalization\u0026mdash;one that brings together the best of rule-based business logic and machine learning (ML) predictions. The goal: deliver smarter, more personalized Next Best Action (NBA) recommendations that adapt to each customer\u0026rsquo;s context. Our framework is built on four core layers: collecting and engineering useful data, validating decisions with clear business rules, generating predictions with multiple ML models, and orchestrating final decisions that balance statistical insights with real-world business needs. Using real CRM data from a large, multi-property service company, we show how this system can predict things like customer engagement, estimate value, and group similar behaviors\u0026mdash;then turn those insights into concrete tasks for relationship managers. In testing, our hybrid approach beat the traditional rule-only system on accuracy (AUC-ROC 0.82 vs. 0.65), acted much faster (cutting signal-to-action time by 75%), and improved customer engagement by 14%, all while keeping necessary checks and balances for enterprise use.\u003c/p\u003e","manuscriptTitle":"A Framework for Hybrid CRM Personalization: Combining Rule-Based Logic with Machine Learning Predictions","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-17 18:28:34","doi":"10.21203/rs.3.rs-9068530/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-19T22:20:13+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-18T15:19:20+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-18T02:35:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"73654851669661628589676612442712744662","date":"2026-03-18T02:33:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"325714905132847419290774787367721114032","date":"2026-03-17T06:32:33+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"145340608878860383094787259308311215883","date":"2026-03-17T01:05:37+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-16T11:00:01+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"199121799140838981823386599471036446079","date":"2026-03-16T00:46:18+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"49707322151249061064701454723522146772","date":"2026-03-16T00:39:09+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-15T21:27:25+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"175363847756429921120472457521746534893","date":"2026-03-15T21:14:33+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"194095281270781278413976447161045814438","date":"2026-03-15T20:37:09+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-15T20:28:31+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-10T08:15:12+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-10T05:53:48+00:00","index":"","fulltext":""},{"type":"submitted","content":"International Journal of Data Science and Analytics","date":"2026-03-09T05:11:12+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"international-journal-of-data-science-and-analytics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jdsa","sideBox":"Learn more about [International Journal of Data Science and Analytics](http://link.springer.com/journal/41060)","snPcode":"41060","submissionUrl":"https://submission.nature.com/new-submission/41060/3","title":"International Journal of Data Science and Analytics","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"f113087a-7680-4c4a-a67f-7835b2c82129","owner":[],"postedDate":"March 17th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-18T23:53:19+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-17 18:28:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9068530","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9068530","identity":"rs-9068530","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}