Introduction
So far in our Beyond ERP blogs (here, here and here), we discussed how enterprises struggle with obtaining actionable insights from ERP systems, how a solid data foundation can enable scalable insights and automation, and principles to adopt while doing visualization. The idea was to set the foundation for digitizing the finance function. Let’s face it – Finance and Accounting teams have always been on a journey of Mission Impossible – getting the right numbers to the right people within the right time. Finance and Accounting teams understand the importance of strategic and transactional numbers not just from financial aspects but also to report on the state of the business. But one challenge that most Finance teams always face is looking at the rearview mirror and being over-burdened with finding answers.
Enter Generative AI (GenAI) – the not-so-secret weapon that can help Finance and Accounting teams open new frontiers for automation, decision-making, and prediction. Instead of discussing how enterprises can approach GenAI systematically, we have picked up one use case – Propensity to Pay – and applied our learnings while deploying the GenAI concepts at few of our customers. We will also apply the concepts we discussed in our prior blogs. The goal is to predict which customers are going to pay, which customers will be “late,” and which customers had trouble receiving invoices. Let’s take a deep dive into how GenAI can transform a seemingly mundane task into something almost magical, with Propensity to Pay as our case study.
Understanding Propensity to Pay: Let’s Predict the Future (or at Least Your Cash Flow)
Imagine this: You’re working in the Order to Cash team of a multinational company, and you’re staring at a list of unpaid invoices. You know some customers will pay, but which ones? Can we use a “One-size fits all” strategy, or should we embrace personalization models to achieve higher ROI? That’s where the Propensity to Pay model steps in as a crystal ball for the finance team. With the aid of propensity modeling, one can focus resources on customers where engagement will generate a behavioral change, giving them the nudge to pay earlier. Traditionally, to build this model using data science methods, one would crunch numbers, look at past payment behavior, and make predictions for each customer. Data scientists would identify data sources, perform feature engineering, and constantly tweak the model. Let’s take this process and use GenAI to take it to the next level.
GenAI: Turning Your Finance Department into Jedi Masters of Prediction
Before we dive into the approach, let’s look at the challenge of working on the Propensity to Pay model. A company may be inclined to focus only on delinquent accounts with the highest propensity to pay. However, these accounts might simply miss a due date but will promptly pay as soon as they realize the oversight. Even though these customers may have a high propensity score, reaching out to them wastes valuable resources, as they are likely to pay on their own without collection efforts. Essentially, these customers can pay and will do so without intervention. Therefore, it is crucial to decide on the factors that can influence the propensity to pay model.
Some of the direct factors that influence the propensity to pay model are:
1. Customer-specific factors – Payment history, credit score, account tenure, and payment terms adherence
2. Invoice-specific factors – Invoice amount, invoice timing, invoice frequency, and payment method
3. External factors – Macroeconomic conditions, industry trends, and seasonality
4. Cash Flow & Financial Health – Liquidity ratios and working capital
5. Discounts, Incentives, and Penalties – Early payment discounts and late payment penalties
Though there are other indirect factors such as product delivery experience, competitive pressure, and behavioral data, we will not consider these factors for our use case here.
So, how Can We Leverage Generative AI to Create the Propensity to Pay Model? Let’s dive into the approach.
1. Data: The Gold Mine that Finance Teams are Probably Ignoring
First things first— we need to conduct a full data evaluation to determine which variables could be used and where the data comes from. For any AI models to work accuractely, we need to have proper high-quality data for it to be trained on else its garbage in means garbage out. If we want GenAI to look beyond obvious things such as who hasn’t paid in 90 days and dig deeper, we need to ensure data quality. We’re talking about:
– Identifying the factors and narrowing them down – Here, finance team members who work day in and day out will have all the information and context.
– Pulling data from various sources – These data elements can come from the ERP, CRM, payment systems, even random Excel files on a desktop.
– Feedback loop – This is often ignored but is essential to ensure that the GenAI model can be regularly trained.
– Standardized column names – Standardizing column names with business-context-driven labels instead of application-specific data field names helps GenAI understand it better.
While finance teams work hard to gather, handle, and contextualize data, these data pipelines need to be automated for GenAI to parse the datasets. Automated data pipelines can ensure that GenAI is constantly learning. So, instead of emailing that Excel file, the team member can upload it into a folder for your data platform to process. Only when these preliminary steps are fully executed do we recommend moving on with the development of the propensity-to-pay model; otherwise, the project will remain at the POC stage.
2. Integrating Structured and Unstructured Data Using RAG
For the propensity to pay model, we will leverage both structured and unstructured data, and this is where the RAG architecture shines. RAG combines structured and unstructured data to feed into GenAI, building contextually and comprehensively aware insights. For example, while structured data might indicate that a customer has delayed payments in the past, unstructured data—emails expressing dissatisfaction with a product—could reveal why those delays happened.
RAG (Retrieval-Augmented Generation) also helps mitigate hallucinations by ensuring the model isn’t simply generating answers based on its training but instead retrieves factual, up-to-date information from trusted sources. This is crucial when dealing with structured (e.g., ERP data) and unstructured data (e.g., customer emails). By anchoring GenAI’s outputs to real, validated data, you minimize the risk of hallucinations.
3. Feature Engineering: The Science Behind the Magic
Feature engineering is where the real next big effort that happens. GenAI can sift through millions of data points and automatically create features—aka important insights—that may give finance teams. Alternatively, a human in the loop can manage feature engineering. Some factors that GenAI helps bypass include model selection and constant experimentation involved in fine-tuning the model.
– Unseen behavior patterns: For example, irregular payments over time may indicate that such customers need to be prioritized.
– External factors: Including industry trends (unstructured data) as part of GenAI inputs.
When dealing with sensitive predictions like a customer’s likelihood to pay, it is important to establish boundaries on what the AI can and cannot generate. These guardrails can include limits on acceptable variability, the rejection of incomplete data, and cross-verification from multiple data sources to validate the final prediction.
4. Model Training: Teaching Your GenAI to Predict Who’ll Pay and whom we need to follow up with
Once the data is in, it’s time to train the model. We want to emphasize that high-quality data (such as payment histories, invoice data, and transaction trends) is crucial for GenAI to ensure accuracy. Here’s how this process can ensure a higher ROI and success:
– Real-time learning: One key to ensuring accurate predictions is through real-time or periodic retraining on fresh data. As new transactions, invoices, or customer behaviors emerge, this updated data must feed into the model.
– Cross-validation with external benchmarks: Continuously benchmark predictions against external experiences or industry benchmarks to gauge the model’s accuracy.
– Numerical calibration: GenAI can predict the likelihood of payment but may struggle with numerical accuracy. Calibrating for specific thresholds improves accuracy.
5. Debtor Segmentation Clustering
“One size fits all” may work for some organizations, but with diverse product ranges and operations in different countries, enterprises may want to follow a segmentation-based approach when applying the Propensity to Pay model. The output from the GenAI models should produce debtor segmentation clusters. The next best actions include collection strategies (e.g., revised payment terms), personalized targeting, and precise risk reporting.
Remember, the accuracy of the models may not be 100%, and adding a human-in-the-loop for critical use cases, especially in finance, ensures that significant decisions are reviewed before action is taken.
What if a traditional Propensity to Pay model already exist that a team is using?
The above steps to setup the model are not the end of the process. Building on top of this solution, Finance teams should strive to use GenAI to identify the next best action (NBA). Many enterprises would have already built a traditional ML or AI model for propensity to pay, consisting of a set of ML models (ensemble) that need to come together (Classification models, regression models and some others) and are based on features such as Customer Demographics, Payment History, Credit Score, Purchase history and Payment terms. Such teams can leverage GenAI on top of these models to build Next Best Action workflow.
The above are few of the approaches that can be leveraged to delivering this solution. However, there are other factors that need to be considered. Principles such as Ethics in AI, Data Privacy and Security, Bias and Fairness and Model Explainability and many more need to be considered while developing such solutions from a long term perspective. It is highly critical that constant evaluation of the model is not ignored.
The (Many) Benefits of GenAI: Flywheel Effect
We provided an approach to implement one use case. What are the tangible benefits of using GenAI for Propensity to Pay?
– Higher accuracy: GenAI helps avoid experimentation and provides responses quickly—it spots clues no one else can see.
– Less manual work: With automated data pipelines, GenAI automates mundane tasks, freeing up the team to focus on more important things (like figuring out how to spend the cash you’re collecting).
– Better cash flow: A better prediction model gives teams a heads-up to take action on late payments before they occur.
Future of GenAI in Finance & Accounting: Numerous use-cases but tread with caution
Finance function is all about reliability and accuracy. Propensity to Pay is just one of the use-cases and at the tip of the AI iceberg. But the experience in delivering this one use-case can act as a foundation for the onboarding other use-cases and also set the process in motion for other use cases such as:
– Automate audit trails by spotting irregularities faster.
– Predict expenses and budgets with real-time data that makes traditional forecasting look prehistoric.
Conclusion: Ready to Let GenAI Save Your Cash Flow?
Generative AI started as a transformative trend, and its inherent power can help reimagine the processes within Finance and Accounting teams. It’s not just about another chatbot or predictions but about developing a data-backed AI foundation to ensure teams focus on activities that lead to maximum ROI. With the optimal combination of data and AI, teams get solid, real-time insights that lead to better decisions and smoother cash flow. What’s not to love?
In our subsequent blogs we will talk about more use-cases and how can we build trust and transparency in the results of these GenAI models.