Mind the Gap: How Hansab Fills ATMs Just Right

What are AS Hansab and the CASHX AI project?

Hansab AS, an Estonian technology firm, oversees the replenishment and emptying of ATMs, managing 360 machines with a team of five specialists. The current manual monitoring process is resource-intensive, leading to unnecessary transports and service fee discrepancies as it is difficult to predict ATM cash flow. The CASHX AI project, led by AS Hansab in collaboration with MindTitan, aims to enhance the efficiency of ATM cash availability management through AI-driven predictions.

The project’s goal is to create an AI solution that uses machine learning to more accurately forecast when ATMs will need more cash, ensuring that ATMs neither run out of money nor are prematurely refilled when there’s still enough cash on hand. By doing so, the project aims to increase the efficiency of cash management, reduce unnecessary transport, and lower operational costs, all while optimizing the fees charged for refilling the ATMs. The main measures of success will include ATM uptime, the cost of operations, and the refilling fees charged per machine.

This effort seeks to reduce the need for constant human checks, reduce the number of unnecessary trips to refill ATMs by at least 10%, and improve ATM service by maintaining a good mix of available cash denominations. Over time, this initiative is expected to not only save costs and boost revenues from fees but also offer this technology to other banks and ATM operators, thereby supporting environmental goals by reducing transport frequency.

“They provide comprehensive support, ensuring robust and secure implementations that are both scalable and competitively priced. Their proven track record and esteemed reputation further underscore their expertise and reliability,” he added.

The problem described

The primary problem the CASHX AI project aims to reduce is resource overuse, manifesting in operational effectiveness, service quality, and financial outcomes. The core issues associated with this problem include:

• Manual Monitoring Process: The current approach to monitoring ATM cash levels is labor-intensive, as it involves a team of five specialists overseeing the cash status of 360 machines. This process is not only resource-heavy but also prone to human error and underperformance.
• Inaccurate Cash Flow Predictions: Due to the manual and somewhat rudimentary monitoring tools and processes, predictions regarding when an ATM will run out of cash or become overfilled are often approximate. This leads to either cash shortages, where consumers cannot withdraw money, or overfill situations, where the ATM cannot accept any more deposits, resulting in poorer customer service.
• Unnecessary Cash Transports: The lack of precise prediction capabilities leads to suboptimal routing and scheduling of cash transport logistics, causing unnecessary trips to refill or empty ATMs. This can mean both Increased operational costs and environmental impacts due to the additional carbon emissions.
• Impact on Service Fees: The non-optimized ATM management directly affects the structure and amount of service fees. By better balancing cash inflows and outflows, we can increase revenue and enhance customer satisfaction
• Resource Misallocation: The substantial human and financial resources currently allocated to manually monitoring and managing ATM cash flows could be better utilized if the process were more automated and efficient. A predictive, AI-driven solution can streamline operations and free up valuable resources needed for other strategic initiatives.

The CASHX AI project addresses the critical challenge of transforming the ATM cash management process from a manual, resource-intensive, and error-prone system into a streamlined, predictive, and AI-enhanced operation. This change aims to optimize cash availability for consumers, reduce operational costs, improve service fee structures, and enhance overall service quality and sustainability.

AI-model described

The AI model for the CASHX AI project is designed to predict an ATM’s cash service requirements, ensuring optimal cash availability while minimizing operational inefficiencies. This involves creating a machine learning model that can accurately forecast when ATMs will need to be refilled or emptied, based on various factors.

Here’s an outline of the envisioned model:

Data Inputs

The model will use historical data from several sources, such as cash withdrawals and deposits, including amounts, times, and frequency; cash levels in ATMs over time; and even information on local events, holidays, and other seasonal patterns that might affect cash usage.

Model Development

The AI model for the CASHX project is set to begin with a Proof of Concept (POC) phase. This strategic choice facilitates a straightforward assessment of the solution’s feasibility and potential effectiveness without requiring extensive effort.

The initial phase encompasses such crucial steps as training, where historical data educates the model on predicting ATM cash requirements using the newly engineered features.

Next, following the training phase, validation tests the model on a distinct dataset not previously used during training, helping to evaluate the model’s predictive accuracy and its ability to generalize across different sets of data.

This methodical approach ensures that the foundational model is robust and reliable before advancing to more complex stages of development.

Deployment and Monitoring

Once the AI model has been developed and validated, the next step is deployment into production. This phase involves integrating the model into the existing ATM monitoring systems, so it begins to make automated predictions about ATM cash requirements and generating alerts.

Key to this process is ensuring that the integration is seamless, allowing the model to function smoothly alongside current operations. Additionally, ongoing monitoring of the model’s performance is crucial.

As patterns in cash usage change over time, continuous oversight helps to quickly identify any decrease in performance or shifts in predictive accuracy. These steps are essential to maintain the reliability and effectiveness of the AI system in a live environment.

Continuous Improvement

After deployment, the AI model will embark on a cycle of continuous improvement, consistently integrating new data to stay current with evolving consumer behaviors and market conditions.

Regular model retraining and refining will ensure it adapts effectively to new patterns and changes.

The goal of this AI model extends beyond merely predicting ATM cash requirements; it aims to achieve such predictions with high accuracy and efficiency. This level of precision is expected to substantially reduce operational costs and enhance customer satisfaction by maintaining optimal cash levels in ATMs.

Ultimately, the model contributes to better financial performance and environmental outcomes for AS Hansab, marking a significant step forward in their operational capabilities.

AI model development

The creation of the AI model for the CASHX AI project is envisioned to progress through a structured, phased approach, ensuring meticulous development, testing, and integration into existing systems. This process includes several critical steps designed to validate feasibility, develop the solution, prepare it for production, and enhance its reliability and data collection capabilities.

Here’s an overview of how the model should be created:

Phase 1: Proof of Concept (POC)

Objective: Validate the feasibility and business applicability of the AI solution for predicting ATM cash requirements.
The initial phase kicks off by setting up a machine learning development environment to begin experimenting with historical data. The process starts with minimal data processing, allowing for the training of a basic model to identify potential patterns quickly.

Phase 2: Initial end-to-end setup

Objective: Develop a comprehensive machine learning solution that can be tested in a controlled environment, alongside the necessary ecosystem for its continuous operation.
In this phase, the project focuses on constructing robust data pipelines that ensure a continuous flow of data and on deploying the first machine learning solution API. This stage also includes enhancing the machine learning model for greater accuracy and complexity. The upgraded model is then tested in real-world scenarios to ensure it integrates well with actual business processes, with adjustments made based on the feedback received

Phase 3: Production-ready MVP

Objective: Finalize all parts of the solution to be production-ready, involving key stakeholders more closely in the development process to ensure compatibility with business processes.
The third phase aims to finalize all solution components, making them ready for production. This involves refining and perfecting the machine learning models and support systems with the insights gathered from the previous phase. The solution is integrated into the existing business processes to ensure seamless operation and wide user acceptance. Extensive testing leads to the necessary adjustments that enhance the solution’s reliability and effectiveness in live environments.

Phase 4: Increased Reliability and Improved Data Collection

Objective: Enhance the solution’s structure for better support and data analysis, ensuring future scalability and automatic calibration.
The final phase enhances the solution’s infrastructure to support better data collection and analysis, preparing it for future scalability and automatic adjustments. Advanced monitoring systems are developed and implemented to collect data precisely, aiding continuous improvements. The system also undergoes enhancements to boost its reliability and complete preparations for a full-scale production launch. Usage and feedback are closely monitored to refine the machine learning model further and ensure its optimal performance.

“The most difficult part of the CASHX AI project AI model development was ensuring the model’s integration into existing business processes during the production-ready MVP phase. This required refining the machine learning models, ensuring compatibility with business operations, and conducting extensive testing to enhance reliability and effectiveness in live environments.”
(Kaarel Ajaots, Business Development Manager, Hansab)

Technical Considerations

For the CASHX AI project, developing a robust AI model capable of optimizing ATM cash management required several pivotal technical considerations. The model’s architecture starts with a benchmark model for initial validation, which allows for quick assessment and sets the stage for potentially more complex architectures depending on the project’s needs and progression.

Managing the data effectively is also critical. This includes setting up efficient data pipelines and preprocessing routines that ensure the data used is of high quality and relevance, which is essential for machine learning model accuracy.

Integration and testing form another crucial aspect. The project requires continuous testing both in controlled environments and in real-world settings. Supported by developing APIs and seamless integration with existing monitoring systems, this process ensures the solution works well within the current infrastructure.

Additionally, a feedback loop–crucial for refining the model over time–uses real-world performance data and ongoing data analysis to adjust the model. This structured approach not only aims to develop a model that can accurately predict ATM cash needs but also strives to reduce operational inefficiencies and improve customer satisfaction by making cash management processes more efficient.

Hurdles and solutions

1. Data Quality and Availability: Access to high-quality, relevant, and comprehensive data is crucial for training accurate machine learning models. Inadequate or poor-quality data can significantly impair the model’s performance. As a solution, we implemented robust data cleaning and preprocessing techniques. We also established partnerships with financial institutions for richer data access and, when faced with gaps, consider synthetic data generation.
2. Integration with Existing Systems: The AI solution must integrate seamlessly with Hansab’s existing ATM monitoring and management systems, which may use outdated technology or lack API support. As a solution, our team developed middleware or used API gateways to facilitate communication between new and old systems. Via close collaboration between the AI team and Hansab’s IT department, we plan for gradual integration.
3. Model Accuracy and Reliability: Ensuring the model accurately predicts ATM cash requirements under varying conditions is challenging, especially given unpredictable consumer behavior and external factors like holidays or emergencies. To overcome this hurdle, we used ensemble learning techniques to improve prediction reliability. Moreover, continuous updates provide the model with new data and feedback to adapt to changing patterns.
4. Regulatory Compliance and Data Privacy: Financial data is sensitive, and projects must comply with regulations such as the EU’s GDPR, which governs how personal data is collected, stored, and processed. Thus, we designed the project with privacy in mind, using techniques like data anonymization and encryption. We also stay informed about regulatory requirements and incorporate compliance checks throughout the project lifecycle.
5. User Adoption and Trust: Ensuring that both bank employees and customers trust and accept the AI-driven processes is crucial for the project’s success. The answer to this challenge is engaging stakeholders early in the development process through demonstrations and trials. We offer training and support to ease the transition and highlight the benefits of the new system.
6. Scalability and Maintenance: As the solution expands to manage more ATMs or adapt to new requirements, maintaining performance and scalability can become challenging. We designed the system with scalability in mind, using cloud-based services and microservices architecture. Planned regular maintenance and updates to the software and models ensures a seamless user experience.

“Working with MindTitan to tackle these hurdles was a collaborative effort, requiring technical expertise, strategic planning, and effective communication,” Karel Ajaots said.