City Power’s R3.96 billion in non-technical electricity losses is the target of a tender for near-real-time monitoring of prepaid purchases, but experts warn analytics will flag suspects, not recover money.
According to City Power’s latest annual report, the 2024/25 figure was up from R3.45 billion the previous year. It accounted for 20.85% of electricity purchased and included theft, illegal connections, meter bypassing, faulty equipment and billing errors.
The report also shows City Power has already introduced automated meter management, targeted audits and customer database reconciliation to address non-technical losses.
City Power reported auditing 40 000 customers who had converted to prepaid electricity but were not purchasing tokens. At the end of June 2025, prepaid customers represented approximately 74% of its customer base of 396 461 but accounted for just 11% of revenue.
Tender mercies
According to June 2026 technical specifications for a tender advertised this month, City Power wants a platform to monitor prepaid electricity purchases in near-real-time.
The tender covers third-party prepaid vending, revenue enhancement, customer monitoring, meter audits and advanced metering infrastructure support, and successful bidders must integrate their technology into City Power’s existing systems.
The documents also require that the appointed providers maintain 99.5% system availability. All customer and transaction data will remain City Power’s property, the document states.
City Power says bidders must be able to identify revenue leakage, suspected theft, meter tampering, illegal connections, bypass conditions and other non-technical losses. Service providers must also identify customers with abnormal behaviour, such as low purchases, zero vending, declining consumption or irregular vending patterns, and recommend interventions.
Bidders may propose artificial intelligence (AI) for anomaly detection, theft prediction, revenue forecasting, predictive maintenance and customer risk scoring. However, City Power requires AI outputs to be transparent, explainable, auditable and compliant with the Protection of Personal Information Act, with human decision-making retained.
One-way traffic
Jacqui Muller, a researcher at Belgium Campus iTversity and a PhD candidate in computer science and information technology, warns that the proposed monitoring capabilities depend on the type of electricity meters installed.
“Standard prepaid meters are essentially one-way. You see the token sale, not the consumption,” Muller says. Near-real-time monitoring needs smart meters with reliable communications, a system that can take in interval readings, and clean links between the vending, billing, meter and GIS systems, she adds.
Muller notes that, without these, real-time “really means real-time sales data”.
A decline in electricity purchases could indicate theft, but could also reflect a faulty meter, an empty property, the installation of solar or backup power, or customers purchasing electricity in bulk, says Muller.
“Analytics is good at spotting that something is unusual. It’s poor at explaining why,” Muller says. “The model produces a list of suspects. The answer only comes from a physical meter audit and checking the customer’s records,” she adds.
Garbage in
Mark Walker, director and co-founder of T4i, also cautions that the effectiveness of automated detection depends on the quality, reliability and frequency of the data flowing from the meters.
AI could generate unreliable predictions when data is poor or incomplete, Walker says, while integrating new technology with existing electricity infrastructure presents risks involving outdated meters, incompatible systems, budget overruns and shortages of technical skills. “Garbage in, garbage out rules still apply and are paramount,” he says.
Muller also notes: “If past audits mostly targeted certain areas, the model will keep pointing back at those areas.” Bidders should be asked to report false-positive rates from field-verified outcomes, not just detection rates, she says.
Research published in August 2026 by Cape Peninsula University of Technology academics, who reviewed 90 studies published between 2015 and 2025, found that while machine learning shows promise in detecting electricity theft, evidence of its effectiveness in real-world deployments remains limited.
They identified challenges including insufficient verified theft data, false positives and difficulties applying models across different electricity networks.
Flags, not cash
Walker says the technology could theoretically cut non-technical losses, but illegal connections, incomplete data and municipal enforcement inertia mean the benefit would need to be weighed against the cost of the system.
Analytics could help City Power target inspections more effectively, says Muller. “Analytics doesn’t recover money. Inspections, prosecutions, reconnection policy and community trust do,” she says.
“Much non-technical loss is illegal connections in areas with no account to analyse,” Muller adds. She warns that detailed purchasing and consumption patterns can reveal household circumstances, including when properties are unoccupied, raising privacy concerns.
Safety catch
“A risk score should only ever trigger a physical inspection, never a disconnection.” She adds that customers must be able to challenge adverse findings.
Muller also warns that the models and the logic linking the systems could belong to up to five vendors, so City Power may not be able to explain or reproduce a decision once a contract ends.
City Power should insist on independent model audits and verified recovery figures and should not reward suppliers for generating alerts or pay the firm that scores the risk per disconnection or per rand recovered on the same case, says Muller. “Fix the data before buying the AI.”
Share
