From Predictive Risk Scoring to Intelligent Collections Management

Written By Jordy Bultman

Transforming Insurance Collections with AI in SAP FS-CD and FSCM-CR

Personal Introduction

Having been deeply involved in implementing SAP Creditworthiness and credit rating solutions at two distinct insurance companies, I encountered significant limitations with the traditional approach. The design was rigid; minor customer errors often led to undeserved rating drops, causing frustration among Insurance Debtor Management and Customer Relationship Management teams. Questions frequently arose: Is a customer consistently paying one day late truly less valuable than one always paying on time? Shouldn’t a long-term customer with multiple products be evaluated differently?

During my early Profitability and Performance Management training years ago, I recognized an opportunity to improve customer ratings by leveraging robust BW models or PaPM's calculation engine. Although our initial prototype didn't fully materialize, recent advancements in AI—particularly predictive analytics—now present a powerful tool to accurately predict early signs of premium payment default, allowing insurers to proactively manage customer relationships. This blog outlines a theoretical AI-driven setup for improving insurance collections management. I’ve chosen to integrate this in the excisting Collections Strategy functionality in SAP Fioneer S/4HANA FS-CD. This is one way to integrate. If you're interested in exploring this further, please reach out!

Brief overview of this blog

Current Struggles of Insurers in Collections Management

Today, insurers face increasing challenges with collections management due to evolving customer behaviors, economic volatility, and regulatory pressures. Traditional collection methods often fail to accommodate nuanced customer interactions or rapidly changing financial conditions, resulting in inefficient processes, reduced cash flow predictability, and strained customer relationships.

Importance of Smarter Collections Management in Insurance

Effective collections management in insurance is crucial not only for maintaining operational efficiency but also for ensuring financial stability, regulatory compliance, and strong customer relationships. Traditional methods relying on static, rule-based processes cannot adequately handle complex payment behaviors or evolving risk factors. A smarter, more proactive approach using AI-driven predictive insights is necessary to optimize collections outcomes and customer satisfaction.

Key Challenges in FS-CD and FSCM-CR Operations

When SAP Dunning by Collection Strategy (many moons ago😃) was released, it delivered a notable increase in flexibility. With BRF+, the dunning procedure could react to the actual status of a customer or contract rather than follow a fixed sequence of steps. When combined with SAP Credit Management’s rating capabilities, this formed a powerful toolset for collections teams. However, it was also a complex solution to manage, crossing multiple business domains and requiring strong alignment between Finance, IT, and Customer Operations. As a result, adoption remained limited and results inconsistent - at least in my experience.

This underscores the necessity for predictive, AI-powered solutions—designed for scalability, interpretability, and cross-functional utility—to enhance operational effectiveness and deliver actionable insight where traditional tools fall short.

Real-World Validation: What Other Industries Have Achieved

To reinforce the practical value of predictive AI in collections, we turn to cross-industry case studies. I’ve got a list of sources at the end.

  • WNS Utility Sector Case: A leading energy provider integrated a propensity-to-pay predictive model, boosting collections by 50% within 3 months, while reducing operational costs by 20%. Segment-specific strategies and tailored scripts drove efficiency and customer engagement.

  • QUALCO Financial Services Case: A loan servicing company deployed QUALCO's Data-Driven Decisions Engine (D3E), resulting in a 10% uplift in cash collections in the first month and better roll rates. Predictive segmentation enabled personalized treatment paths aligned with risk categories.

  • Alesco Debt Collection Agency Case: Using deep neural networks and third-party demographic data, Alesco improved collections by 15–18% over incumbent models, cut credit bureau costs, and automated prioritization. Their top 20% of accounts generated 84% of collections, emphasizing the Pareto principle in action.

These results demonstrate AI’s transformative potential. Insurers can expect similar benefits when predictive scoring is integrated tightly into SAP FS-CD and FSCM-CR frameworks. That is why I am looking into the following 2 use cases in this blog: 

  1. Predictive Payment Risk and Default Scoring

From the customer’s side, Predictive Payment Risk leads to more relevant and fair interactions. The Predictive Payment Risk and Default Scoring model uses real behaviour and context to assess likelihood of default—not just a fixed rule set. Customers showing early signs of trouble can be proactively offered flexible options, while those with strong patterns aren’t penalized for minor issues. It results in better communication, less unnecessary pressure, and more personalized service. Integration with FS-CD and FSCM-CR ensures this intelligence flows directly into dunning, credit review, or agent workflows—keeping everything aligned and efficient.

  1. Enhanced Dunning and Collections Strategies

These models bring adaptability to the dunning process. With BRF+ in FS-CD, scores can trigger differentiated treatments—an at-risk customer may be routed for early human contact, while low-risk accounts follow an automated path or are taken over by AI Agents/Chat bots. This lets collections feel more responsive, rather than reactive. You can still set legal dunning steps as fixed anchors is the process, a step that has to be completed like an official first reminder letter.

On the integration side, it's lightweight and in line with SAP best practices. Scores are pushed into FSCM-CR or FS-CD using existing enhancements like BRF+ expressions, BAdIs, or score fields. That means you get smart logic in place fast, without compromising the clean core or adding integration debt.

Where Is the Data Coming From?

Training data for AI scoring models comes from a variety of sources within and around the SAP for Insurance Landscape (including FS-CD):

  • Historical premium billing and payment behaviour

  • Dunning levels and write-offs

  • Customer contact records (via CRM or call center integrations)

  • Policy and product data (via policy admin systems)

  • Claims activity (frequency, status, pay-outs)

  • Agent channel metadata (broker vs. direct)

  • Optional: External signals such as credit bureau scores or economic indices

Beyond structured enterprise data, external contextual signals can significantly enhance prediction accuracy. For example:

  • News feeds indicating macroeconomic shifts (e.g., new tariffs, trade wars) that may impact regional financial stability

  • Weather-related disruptions such as wildfires or floods, where affected customers may delay payments due to emergency circumstances

  • Sector-level downturns that could affect regions or occupations with high insurance exposure (e.g., automotive layoffs in a manufacturing hub)

These contextual features can be consumed via public APIs or integrated data services and included in the feature set for model training.

​To prepare data for training AI models, various datasets are consolidated using tools like SAP Datasphere, Core Data Services (CDS) views, or SAP BW queries. These datasets are then exported for model training using platforms such as SAP AI Core, Python environments, or hyperscaler services. SAP Datasphere facilitates this process by providing robust data integration and modeling capabilities, enabling seamless extraction and preparation of data from diverse sources.

How Is the Model Trained and Called?

Once the data is prepared, the predictive model undergoes training using supervised machine learning techniques tailored to the complexity of the use case:

  • Gradient Boosted Trees (e.g., XGBoost): Ideal for structured datasets like FS-CD billing records and behavioral data, offering strong predictive accuracy with high interpretability.

  • Logistic Regression: Useful when probability scoring and model transparency are prioritized, such as in regulatory reporting scenarios.

  • Neural Networks: Applied where complex, multi-dimensional interactions exist—such as modeling behavior across bundled policies or product types.

Model training is typically conducted on SAP Business Technology Platform (BTP), using services such as SAP AI Core and AI Launchpad. SAP AI Core orchestrates the full machine learning lifecycle: training, deployment, and execution management.

During training, historical datasets are accessed from connected object stores on supported hyperscalers (e.g., AWS S3, Azure Blob), ensuring scalable data handling. Once trained and validated, the model is deployed as a RESTful scoring service hosted within SAP AI Core.

Integration into SAP is achieved by exposing the model via REST or OData endpoints. The model can be triggered:

  • On a schedule (e.g., weekly credit review)

  • In response to business events (e.g., just before executing a dunning run)

The resulting risk scores are written back into SAP systems for instance like: 

  • Calculated scores in FSCM-CR, using enhancement options like BAdIs or one of the specific BAPI’s that are available.

This tight integration enables dynamic, data-driven collections management that adjusts based on real-time predictive insights. Scores can be stored in one of several locations, depending on architecture:

  • FSCM-CR Score Field (UKM_SCORE) – If used to update the credit rating lifecycle

  • Custom BP Master Extension – For use in FS-CD-specific logic

  • FS-CD Contract Account or Insurance Object Attributes – If the score applies to a Policy or Contract.

Visualizing Risk: Predictive Risk Bands

Scores can be mapped to color-coded risk bands or rating charts that collections teams use to prioritize actions. To provide more nuance and foresight in these examples - I am proposing a dual-band scoring model. 

  • Green (Stable Low Risk) – Low risk with no significant predicted change in behavior

  • Green− (Emerging Risk) – Currently low risk, but prediction signals a potential negative shift (e.g., moving toward medium risk)

  • Green+ (Recovering Strong Payer) – Low risk, previously trending downward but now showing strong, stable patterns

This allows business users to differentiate between similar scores with different trajectories and treat customers accordingly.

Alternatively, insurers may opt to enrich standard FSCM-CR scoring with a complementary indicator at the Business Partner level. This could be implemented as a classification extension to the BP master data (e.g., a custom Z-field or a UKM extension) with values such as:

  • ++ for improving outlook

  • + for stable

  • - for declining

  • -- for high volatility or risk

BRF+ can then use these enhanced classifications to refine strategies. For example:

  • Offer a flexible payment plan for Yellow with ++ (likely to self-correct)

  • Trigger an agent call for Yellow with -- (early sign of deterioration)

  • Escalate Red -- for Debtor management to decide

This layered classification allows a smarter, more contextual approach to dunning and customer engagement, improving both operational accuracy and customer fairness.

Looping Back: Realizing the Use Cases in Practice

With the architecture, data, and integration foundations now in place, let's revisit our two core AI use cases and explore what their real-world implementation could look like—both in terms of system impact and daily benefits for Debtor Management teams and customers.

1. Predictive Payment Risk and Default Scoring

After following the implementation roadmap, a predictive model trained on historical FS-CD data (billing patterns, missed payments, claims context) is deployed and integrated into FSCM-CR. Each week or before major billing cycles, the model runs and updates a payment risk score for each customer.

In practice:

  • High-risk customers are flagged and routed to specialized agents

  • Medium-risk customers are offered automated payment/Installment plan options, potentially picked up by AI agents that engage the customer proactively through chat or email, presenting flexible payment terms and addressing concerns on the spot

  • Low-risk customers are handled via standard dunning logic, with AI agents also assisting in clarifying account status or responding to basic queries—freeing up debtor management specialists to focus on higher-risk cases

Benefits:

  • Debtor Management gains a clear, data-driven prioritization tool

  • Less time is spent checking accounts manually

  • Resources are focused where they have the most impact

  • Customers receive fairer treatment, especially those in temporary hardship, reducing unnecessary escalations

2. Enhanced Dunning and Collections Strategies

In this scenario, the predictive score is consumed by FS-CD’s BRF+ collection strategies. Dunning paths are no longer fixed but dynamically selected based on the customer's latest risk score. The predictive AI would use it’s knowledge on which dunning step works best in what scenario and assigns this step in FS-CD (trough BRF+)

For example:

  • A customer flagged as high-risk may trigger an immediate email followed by a personal call

  • Fixed points in the collection procedure are still executed; like a formal first dunning letter.

  • A long-time customer with a recent delay but overall low-risk may receive a friendly reminder and a payment link, potentially sent by an AI chatbot that can also provide clarification or offer instant rescheduling options

  • Those in affected regions (e.g., hit by natural disasters) may be automatically deferred for collections temporarily, with AI agents identifying and flagging such cases proactively—reducing strain on the Debtor Management team

  • Because the dunning still flows trough FS-CD, a full history is available.

Benefits:

  • Dunning becomes empathetic, strategic, and more successful

  • Customer complaints drop as interactions feel more personalized

  • Payment compliance improves thanks to relevant outreach timing and tone

Together, these two use cases transform both operational workflows and customer experience—enabling collections to become a partner in financial well-being rather than a reactive cost center.

Conclusion and Strategic Advice

AI-powered predictive analytics marks a turning point in insurance collections. Case studies confirm that industries using similar approaches have achieved:

  • 5–50% collection improvements

  • Reduced operational costs (10–20%)

  • Enhanced fairness and customer engagement

By embedding predictive models into FS-CD and FSCM-CR, insurers modernize collections into a proactive, customer-centric function. With SAP BTP and BRF+ logic, predictive scoring becomes operationalized—boosting performance and resilience in today’s volatile financial landscape.

Moreover, the use of transparent AI models ensures compliance with regulatory demands while fostering trust among stakeholders. Feature-rich, continuously learning models backed by lifecycle governance drive real strategic advantage.

Let collections become a driver of strategic value, not just a cost center.

Interested in Brainstorming Together?

If you're exploring how to bring AI-driven intelligence into your SAP collections landscape—or just want to spar ideas on the future of FSCM-CR, FS-CD, or credit risk—I'd love to connect. Whether you're early-stage or mid-implementation, there's value in a good exchange of ideas.

Reach out via LinkedIn or email to set up a conversation. Let's explore what's possible together.

Sources I used to set up this blog: 

SAP Documentation and Learning

Case Studies and Industry Applications

AI and Predictive Modeling Insights

Jordy Bultman
Next

Future-Proofing FS-CD: How to Transition from SAP ECC to S/4HANA (2/2)