Electricity Theft and Non-Technical Losses

Finding the electricity that is delivered but never billed, from meter tampering, bypasses, and illegal connections to faulty meters and billing errors, and deciding which premises a limited number of field inspectors should visit.

AI is suitable with conditions

Utilities that inspect at scale hold labelled history in the form of past inspection results, and the evidence that a model can use it is real: Endesa's model for its industrial and commercial customers was trained on about 57,000 field inspections and was put into use in an inspection campaign for large customers, and the team behind Naturgy's long-running system in Spain reports campaign precision well above random selection and above the company's earlier rules. It is conditional because those labels are biased by where inspectors were sent. A study of 3.6 million customers and 820,000 inspections at a Brazilian utility found that past inspections concentrated on certain neighbourhoods and customer classes, which makes a model's predictions unreliable for the rest. The researchers behind a system running for years at a Spanish utility warn that, without domain experts involved, dataset shift and bias are hard for the algorithm to identify and performance is at high risk of degrading. A World Bank lead energy specialist recommends that revenue protection start by metering the largest customers, which in one Brazilian utility meant 0.5% of customers producing 46% of revenue. A model belongs after metering and energy balance, ranking premises for inspectors, with part of each campaign spent on random inspections to correct its bias.

Problem Statement

A utility delivers electricity into a network and bills customers for what their meters record.

The difference between the two, after the technical losses of cables and transformers, is non-technical loss: energy that was consumed but never billed.

Much of it is theft, by tampering with or bypassing a meter or connecting directly to the network, but some of it is faulty meters and billing errors that look the same in the data.

Revenue protection teams have to find it with a small number of field inspectors and a very large number of premises.

Business Context

A survey of research on the problem notes that in some countries non-technical losses may reach 40% of the electricity distributed.

In India, the Ministry of Power reports that aggregate technical and commercial losses, which include unbilled energy and uncollected bills as well as technical losses, fell from 22.62% in 2013-14 to 15.04% in 2024-25, a period that included the national distribution scheme’s push for smart metering.

In Great Britain, Ofgem introduced a supply licence condition in 2014 because it judged that suppliers and network operators were not doing enough to detect theft, citing higher prices for honest customers, reduced safety, distorted competition between suppliers, and links to organised crime.

The problem therefore ranges from a matter of survival for utilities with very high losses to a regulatory obligation in markets where theft is a small share of supply.

Why This Is Hard

The first difficulty is that the labels come from inspections, and inspections are not random.

A study of a Brazilian utility’s data, covering 3.6 million customers and 820,000 inspection results, found that previous inspections had concentrated on certain neighbourhoods and customer classes, so a model trained on them is least reliable where inspectors rarely went.

The second difficulty is that a meter that has been bypassed does not record what it missed.

Theft shows up indirectly, as a drop against the premises’ own history, as consumption low for its kind of customer, or as a gap between the energy entering a transformer and the energy its customers are billed for.

The third difficulty is that alarms are noisy.

A revenue protection vendor quoted in 2008 said that, in its experience, as much as 95% of smart meter tamper flags were erroneous, triggered by vibration, outages, or contractors, and that most of the losses it helped find were not theft at all.

The fourth difficulty is that research accuracy is not the measure a utility needs.

This playbook’s judgement is that what matters is a method’s hit rate on real inspections against the targeting it replaces, and few published methods report it.

How The Solutions Fit Together

Tamper-resistant metering for the largest customers comes first, because a World Bank lead energy specialist found that in many utilities as few as 1 to 2% of customers produce as much as 50 to 60% of revenue.

Feeder and transformer energy balance runs beside it, showing which network areas lose the most energy, so inspectors work where the losses are rather than where they have always gone.

A model trained on inspection results supports both, ranking the premises within a campaign by their likelihood of an irregularity, once there are enough results to learn from.

Unsupervised screening of consumption patterns is a discovery tool where smart meters are new and inspection history is thin, and much of what it finds may be faulty meters and billing errors rather than theft, which one vendor put at about 80% of the losses its analysis found.

None of these proves anything.

Each sends an inspector to a premises, and the inspector’s finding is what counts.

Implementation Considerations

Record every inspection, including those that found nothing, with the reason the premises was chosen, because those records are both the labels and the measure of each targeting method.

Keep a share of every campaign as random inspections, which is this playbook’s recommendation rather than a sourced finding, so the hit rate of the targeted list can be compared with chance and the model sees premises it would not have chosen.

Separate theft from faulty meters and billing errors in the findings, because they need different responses and a model trained on all three as one label learns a mixture.

Check customer-to-transformer mapping before trusting an energy balance, since a customer recorded on the wrong transformer moves the loss with them.

The researchers behind a system that has run in production for years at a Spanish utility report that adding explanations of each prediction improved it and gave domain experts, data scientists, and business analysts common ground for judging it.

ROI Considerations

The return is energy and revenue recovered per inspection, net of the cost of the inspection and of disputes lost.

A World Bank knowledge note reports revenue protection metering of large customers that paid back in seven months in the Dominican Republic, and most such deployments paying back in 18 months to two years.

It also reports one utility whose commercial losses fell from nearly 40% to the low teens under that approach, while noting that most such projects were still being implemented.

A model’s return is the increase in hit rate over the targeting it replaces, measured on comparable campaigns, multiplied by the value found per successful inspection.

Security, Privacy, and Regulation

Detailed consumption data can reveal patterns of behaviour inside a home, as Ofgem’s 2018 impact assessment on half-hourly data notes, and Great Britain’s rules on which granularity of domestic data a supplier may use, and for what purpose, have changed since with the reform of half-hourly settlement.

A theft model has to be built on the granularity of data the utility is currently allowed to use for theft detection, checked against the rules in force, and the rules differ between countries and between domestic and business customers.

In Great Britain, supply licence condition 12A requires suppliers to detect, investigate, and prevent theft, and the Retail Energy Code sets out how they do it, including a Theft Detection Incentive Scheme, run in its 2025 to 2026 form, and a public tip-off service, according to the Retail Energy Code Company.

When Ofgem introduced the condition, it said that in cases of theft a customer’s vulnerability should be assessed in the light of what disconnection would mean for them, and the condition requires suppliers to report damage or interference to the network operator so safety concerns can be investigated and resolved.

Where The Evidence Disagrees

The research literature presents non-technical loss detection mainly as a machine learning problem, while the World Bank’s practitioner guidance starts with metering the largest customers and treats revenue protection as a test of a utility’s willingness to act, and warning that resistance to metering large customers may point to collusion between those customers and the utility’s own leadership.

Published models report high accuracy, but the Brazilian study of inspection bias argues that such figures are unreliable when the inspected customers do not represent the population, and the one production team here that reports real campaign results, at Naturgy, says they were much better than random selection but gives no random baseline figure and notes that bias and dataset shift sometimes made campaigns worse than expected.

The vendor view that most smart meter tamper flags are false dates from 2008, and no more recent independent measurement of alarm reliability was found, so this playbook treats alarms as one input among several rather than as worthless.

Symptoms

Root causes

Business impact

Traditional approaches

Solution approaches

Each of these is written for this problem specifically, not borrowed from a generic catalogue.

Recommendation

Protect the largest customers first with tamper-resistant metering, and use feeder and transformer energy balance to find the areas where losses are concentrated. Add a model that ranks premises for inspection only once there are enough inspection results to learn from, keep a share of every campaign as random inspections so the labels do not only reflect past choices, and keep a revenue protection analyst able to see why each premises was ranked. Use unsupervised screening of consumption patterns where smart meters are new and inspection history is thin, as a source of candidates rather than a verdict. A field inspector confirms every case on site, and revenue protection staff decide back-billing, disconnection, and prosecution, taking the customer's vulnerability and safety into account.

Data requirements

Meter readings and events high sensitivity
Interval consumption from smart meters, or periodic readings from conventional meters, with meter events and alarms such as cover open, reverse flow, and power down, kept per meter with timestamps.
Inspection results high sensitivity
Every field inspection, with the premises, date, why it was chosen, what the inspector found, including no irregularity, and whether the finding was theft, a faulty meter, or a billing error. The label for every model and the measure of every targeting method.
Network connectivity and feeder metering medium sensitivity
Which customers are supplied by which transformer and feeder, and the energy measured at each, needed for energy balance. Often out of date, and worth checking before relying on it.
Customer and contract records high sensitivity
Tariff, customer class, activity, contracted capacity, address, account history, and payment behaviour, used to compare a premises with similar ones.
Tip-offs and enforcement outcomes high sensitivity
Reports from the public, staff, and other agencies, and the outcome of back-billing, disconnection, and prosecution, so the team knows which findings were upheld.

Evaluation metrics

Human oversight

A field inspector confirms every irregularity on site and records the evidence, and nothing is presumed from data alone. Revenue protection staff decide back-billing, disconnection, and any referral to the police, taking into account the customer's vulnerability and making the installation safe. A model or alarm only ranks premises for inspection, and a revenue protection analyst reviews why a premises was ranked before it goes into a campaign.

Common pitfalls

When not to use AI

Sources

Practitioners

Research

Case studies

Regulation and guidance

Agent context

A distilled brief for an AI agent or an engineer scoping this work. It carries the constraints and pitfalls that change an implementation decision, not the whole page.

Get the agent context as Markdown

Objective

Reduce non-technical electricity losses by metering the largest customers, locating losses through feeder and transformer energy balance, and ranking premises for field inspection, so inspectors find more theft, faulty meters, and billing errors per visit without accusing customers on data alone.

Constraints

Inputs

Outputs

Success criteria

Last updated 2026-10-03