The accurate prediction of Customer Lifetime Value (pCLV) is a cornerstone of modern marketing strategy, enabling businesses to optimize customer acquisition, retention, and engagement efforts. For decades, organizations have relied on probabilistic models and, more recently, standard machine learning techniques to forecast the future value of their customer base. However, these traditional approaches often struggle to capture the full complexity of the modern customer journey, which is characterized by interactions across numerous digital and physical touchpoints. The primary limitation lies in their difficulty processing the heterogeneous and often unstructured data generated by these interactions.
This paper introduces and validates a new paradigm: the application of Relational Foundation Models (RFMs) to pCLV. Our primary focus is a strategic analysis of the business and operational impact of RFMs compared to traditional models. We posit that RFMs, built on transformer architectures, offer a transformative approach by natively integrating complex, multi-modal data streams to model dynamic customer relationships with unprecedented accuracy. This analysis provides a conceptual framework for integration with enterprise systems and evaluates the strategic value proposition of this emerging technology.
The evolution of pCLV modeling reflects a continuous search for methods that can more accurately capture customer behavior. Early approaches were dominated by probabilistic models such as the Pareto/Negative Binomial Distribution (NBD) and the Beta Geometric/NBD (BG/NBD).9 These models are valued for their stability in non-contractual settings, where customer attrition is unobserved, by modeling transaction rates and the probability of inactivity.11 However, comparative analyses show that no single probabilistic model consistently outperforms others across all scenarios.9 The primary drawback of these models is their reliance on summarized transactional data—typically recency, frequency, and monetary (RFM) value—which ignores rich contextual information.
The advent of transformer architectures, adapted for tabular and relational data, marks a significant departure from prior methods. This has given rise to Relational Foundation Models (RFMs), a class of AI designed to predict outcomes from structured enterprise data.5,10 The core innovation involves serializing rows from relational databases into sequences of tokens, allowing the transformer’s self-attention mechanism to learn complex patterns within and across diverse data types.5 This approach enables the fusion of multi-modal data, where inputs like text, images, and tabular data are tokenized and projected into a common feature space for joint attention processing.3,4
Early work on relational enterprise data frequently emphasized explicit graph-based representations and Graph Neural Networks (GNNs) to model entity relationships within CRM and ERP systems. However, recent advances in Relational Foundation Models have shifted the field toward transformer-based architectures operating over serialized relational data. Large-scale tabular foundation models such as SAP-RPT-1, TabICL, and CASPR demonstrate that attention-based mechanisms can implicitly capture relational structure while offering superior scalability, deployment flexibility, and cross-task generalization compared to graph-centric approaches in enterprise environments.
For the temporal sequences inherent in pCLV, specialized architectures like the Temporal Fusion Transformer (TFT)15 and models incorporating time-aware attention mechanisms14 have been developed to explicitly handle time-decaying effects and preserve causal order. Models like ‘T-CLV’ have demonstrated superior predictive accuracy over traditional methods by leveraging self-attention to capture long-range dependencies across a customer’s entire interaction history.16
This study employs a strategic analysis framework to evaluate the impact of RFMs on pCLV calculation within enterprise systems. The methodology is based on a comprehensive synthesis of recent academic literature, technical documentation, and benchmark performance studies concerning transformer-based tabular data models. We compare the RFM paradigm against traditional pCLV approaches (probabilistic and standard machine learning) across three critical dimensions:
By analyzing RFMs through this tripartite lens, this paper provides a balanced assessment of their current capabilities, inherent challenges, and future potential as a dominant technology for pCLV in enterprise environments.
This section presents a structured evaluation of Relational Foundation Models by analyzing their predictive performance, architectural scalability, and strategic implications within enterprise predictive customer lifetime value systems.
A key advantage of RFMs is their ability to process and synthesize diverse, multi-modal data streams without extensive feature engineering. This holistic data ingestion allows for a more complete and dynamic customer profile, leading to superior performance in certain contexts. Transformer-based models designed for CLV, such as T-CLV, have demonstrated higher predictive accuracy than both probabilistic models and recurrent neural networks by capturing complex, long-range dependencies across a customer’s entire interaction history.16
However, benchmark studies reveal a nuanced performance landscape. Foundation-style tabular models show a distinct advantage on small to medium-sized datasets (<10,000 rows), particularly those with rich semantic content, where they consistently outperform gradient boosting models.1,2 For instance, on one benchmark, SAP-RPT-1 outperformed LightGBM by a 5.5% margin on a dataset with only 540 rows.2 This trend reverses as data volume increases. On datasets larger than 10,000 rows, gradient boosting models like LightGBM and CatBoost typically achieve superior results.1,2
The performance of RFMs on larger datasets is often surpassed by Auto-ML solutions like AutoGluon.18 For large, hybrid datasets, the performance of different model classes converges, suggesting that the choice is less clear-cut at scale.1
The transformative power of the self-attention mechanism comes at a significant computational cost. The original transformer architecture has a time and memory complexity of O(L²), where L is the sequence length.8 In the context of pCLV, where L can represent a customer’s entire event history, this quadratic complexity creates a severe scalability bottleneck for training and inference, posing challenges for deployment in resource-constrained environments.7 To address this, researchers have developed more efficient transformer variants. Models like the LogSparse Transformer, Informer, and Autoformer reduce complexity to O(L(logL)²) or O(L log L), making long-term forecasting more feasible.8
Despite these architectural improvements, real-world implementations face practical constraints. An analysis of SAP-RPT-1 revealed that the model was pre-trained on public web tables, not proprietary enterprise data, and operates with a restrictive context window of 2,073 rows by 50 columns, raising questions about its readiness for large-scale enterprise use cases.19 Furthermore, the computational overhead remains substantial; one study found SAP-RPT-1 to be, on average, 22 times more computationally expensive than LightGBM.2 Under heavy loads, such systems can experience significant increases in response time and resource utilization, indicating potential performance bottlenecks.7
Beyond predictive accuracy, the strategic value of RFMs lies in their potential to “horizontalize” AI systems by solving numerous narrow AI problems with a single, general-purpose model.5 Enterprise software vendors like SAP envision using a single RFM, such as RPT-1, to replace a multitude of existing narrow models for tasks like personalized recommendations, churn prediction, and sales order completion.5,10 This consolidation offers significant operational benefits, including reduced development and maintenance overhead associated with training and managing separate models for each predictive task.10 Models like SAP-RPT-1 are designed for in-context learning, generating predictions based on context examples provided at inference time, which further streamlines deployment by avoiding task-specific fine-tuning.10,17 Pre-trained Relational Foundation Models further enable in-context and zero-shot prediction by generalizing learned relational patterns across tasks and domains without requiring organization-specific retraining. This paradigm shifts the focus from building bespoke models to leveraging a powerful, pre-trained foundation for a wide array of business predictions.
The emergence of Relational Foundation Models presents a compelling, albeit complex, value proposition for predictive CLV. On one hand, the theoretical advancements are undeniable. The ability to fuse multi-modal data into a unified representation allows RFMs to model customer behavior with a fidelity that is opaque to traditional methods. The strategic vision of a single, horizontal model replacing dozens of narrow, task-specific models is a powerful driver for operational efficiency and could democratize access to advanced predictive capabilities within large enterprises.5
On the other hand, our analysis highlights a significant gap between this vision and the current reality of enterprise deployment. The primary obstacle is scalability. While RFMs excel on smaller datasets, their performance and cost-effectiveness diminish rapidly as data volume grows, ceding ground to well-established gradient boosting machines.1,2 The critique that a prominent enterprise RFM was pre-trained on public internet data rather than curated enterprise data, and operates with a limited context window, suggests that the technology may be closer to a research project than a revolutionary, enterprise-ready solution.19 This raises critical questions about the generalizability and robustness of models not pre-trained on the specific schemas and data distributions found within corporate ERP and CRM systems.
While Relational Foundation Models offer improved predictive capacity, their interpretability remains a critical challenge for enterprise adoption. Current practice relies on post-hoc explanation techniques such as SHAP and LIME, which provide localized feature attributions but lack end-to-end auditability. Future research should explore native interpretability mechanisms for RFMs, including attention-based attribution and traceable decision paths aligned with enterprise governance and compliance requirements.
Furthermore, the inherent properties of the transformer architecture are not universally beneficial for all predictive tasks. For some long-term time series forecasting scenarios, simple linear models have been shown to outperform complex transformers, whose permutation-invariant self-attention mechanism can fail to properly utilize temporal information.6 This suggests that for pCLV, which is fundamentally a time-series problem, a blind application of generic transformer architectures may be suboptimal. These limitations further reinforce the need for architectural and governance-aware advances in enterprise RFM design.20
Relational Foundation Models represent a significant paradigm shift for predictive Customer Lifetime Value, offering a path to overcome the data-modality limitations of traditional models. Their ability to learn from complex, heterogeneous data and the strategic benefit of model consolidation present a clear vision for the future of predictive analytics in enterprise systems. However, RFMs are not a panacea. The path to widespread adoption is contingent upon overcoming substantial hurdles related to computational scalability, cost-efficiency, and model validation.
Current benchmarks indicate that for large-scale enterprise data, established methods like gradient boosting often remain superior in both performance and cost. Future research must focus on developing more efficient transformer architectures, exploring pre-training strategies on genuine, large-scale enterprise datasets, and enhancing model interpretability. Until these challenges are addressed, RFMs will likely exist as a powerful but specialized tool rather than a universal replacement for existing pCLV methodologies.
