City Power: Issues Electricity Theft Monitoring Tender for R3.96 Billion Losses

• City Power is seeking a near-real-time monitoring platform to combat R3.96 billion in non-technical electricity losses for 2024/25.
• The tender requires bidders to identify various forms of revenue leakage and abnormal customer behavior, potentially using AI.
• Any proposed AI systems must be transparent, explainable, auditable, and compliant with POPIA, retaining human decision-making.
• Experts warn that monitoring effectiveness depends on smart meters and high-quality data, as standard prepaid meters only show sales.
• AI can identify unusual patterns but cannot explain their causes, necessitating physical audits to confirm suspected theft.

The explicit requirement for AI systems to be transparent, explainable, auditable, and POPIA-compliant, alongside the retention of human decision-making, sets a significant precedent for the ethical deployment of artificial intelligence in public services.

City Power has issued a tender seeking advanced solutions to combat its substantial non-technical electricity losses, which reached R3.96 billion in the 2024/25 financial year. This figure marks an increase from R3.45 billion in the preceding year and constitutes 20.85% of the total electricity purchased by the utility. These losses encompass a range of issues, including outright theft, illegal connections, meter bypassing, equipment malfunctions, and billing inaccuracies.

The utility aims to implement a platform capable of monitoring prepaid electricity purchases in near-real-time. This initiative builds upon existing measures City Power has already undertaken, such as automated meter management, targeted audits of customer accounts, and reconciliation of its customer database, all designed to mitigate these significant financial drains. Notably, an audit revealed 40,000 prepaid customers who had ceased purchasing tokens. Despite prepaid customers making up approximately 74% of City Power's 396,461 customer base by the end of June 2025, they contributed only 11% of the total revenue.

The recently advertised tender, detailed in technical specifications from June 2026, outlines comprehensive requirements for prospective service providers. The scope includes managing third-party prepaid vending, enhancing revenue streams, monitoring customer behavior, conducting meter audits, and supporting advanced metering infrastructure. Successful bidders will be mandated to integrate their proposed technologies seamlessly into City Power's existing operational systems and ensure a high system availability rate of 99.5%.

Crucially, the tender specifies that all customer and transaction data generated will remain the exclusive property of City Power. Bidders are expected to develop capabilities to pinpoint revenue leakage, identify suspected theft, detect meter tampering, uncover illegal connections, and recognize bypass conditions, alongside other forms of non-technical losses. Furthermore, providers must be able to flag customers exhibiting abnormal purchasing patterns, such as consistently low or zero vending, declining consumption, or irregular token acquisitions, and propose appropriate interventions. While artificial intelligence (AI) can be proposed for tasks like anomaly detection, theft prediction, revenue forecasting, predictive maintenance, and customer risk scoring, City Power has stipulated strict conditions: any AI outputs must be transparent, explainable, and auditable, ensuring full compliance with the Protection of Personal Information Act (POPIA), with human decision-making explicitly retained in the process.

Industry experts have raised significant concerns regarding the practical implementation and effectiveness of the proposed monitoring system, particularly concerning data quality and the limitations of AI. Jacqui Muller, a researcher at Belgium Campus iTversity and a PhD candidate, highlights that the efficacy of monitoring capabilities is intrinsically linked to the type of electricity meters currently installed. She explains that standard prepaid meters are "one-way" devices, only registering token sales rather than actual consumption data. True near-real-time monitoring, according to Muller, necessitates smart meters equipped with reliable communication capabilities, a system designed to process interval readings, and robust, clean data linkages across vending, billing, metering, and geographical information systems (GIS). Without these foundational elements, "real-time" monitoring would effectively be limited to sales data.

Muller further cautions that a decline in electricity purchases, while potentially indicative of theft, could also stem from various legitimate factors, including a faulty meter, an unoccupied property, the installation of solar or backup power, or customers opting to purchase electricity in larger quantities. She emphasizes that while analytics excels at identifying unusual patterns, it often falls short in providing the underlying reasons. Consequently, AI models would generate a "list of suspects," with definitive answers only obtainable through physical meter audits and thorough examination of customer records.

Mark Walker, director and co-founder of T4i, echoes these sentiments, stressing that the reliability of automated detection systems is directly proportional to the quality, consistency, and frequency of data transmitted from the meters. He warns that AI could produce unreliable predictions if the input data is poor or incomplete, and integrating new technology with existing infrastructure presents its own set of challenges.

The City Power electricity theft monitoring tender represents a critical effort to curb substantial financial losses, but it also underscores the complex interplay between technological ambition, data integrity, and regulatory compliance in South Africa. The explicit requirement for AI systems to be transparent, explainable, auditable, and POPIA-compliant, alongside the retention of human decision-making, sets a significant precedent for the ethical deployment of artificial intelligence in public services. This emphasis is particularly crucial given that AI-driven insights could lead to accusations of illegal activity against customers, necessitating robust safeguards against erroneous flagging.

For technology companies bidding on this tender, or any entity deploying AI for fraud detection within South Africa, the legal and ethical obligations are clear. Compliance with POPIA is paramount, demanding that AI outputs are not only accurate but also comprehensible and justifiable. The warnings from experts regarding data quality and the inherent limitations of AI in explaining anomalies highlight the potential for inaccurate "suspect" lists, which could lead to legal challenges and erode public trust. Therefore, ensuring high-quality data and maintaining human oversight are not merely operational best practices but essential components for navigating the legal landscape surrounding AI-driven fraud detection in a sensitive sector like utility management.

Lawyers advising tech companies bidding on City Power's tender, or any entity implementing AI for fraud detection in South Africa, must ensure the proposed systems comply with POPIA requirements for transparent, explainable, and auditable AI outputs, particularly when dealing with personal data and potential accusations of illegal activity. Compliance officers should note the emphasis on human decision-making and the need for robust data quality to avoid legal challenges arising from inaccurate AI-driven 'suspect' lists.

Source: Original reporting via industry analysis and tender documents