AI for Fraud Review in Utilities

A spreadsheet arrives from the metering team on a Monday morning. It lists every account on one low voltage feeder, ranked by the gap between this quarter's consumption and the same quarter a year earlier. Beside it sits the district loss figure, which says more energy left the substation than the billed accounts can explain.

Somewhere in those rows is the reason. Most of them are not it. They are a family away for the summer, a shop that shut, a communications module that stopped reporting, a meter swapped and never linked to the account.

A handful of investigators cannot visit all of them. They need the referrals that will hold, where the profile, the tenancy record and the asset history point the same way.

That selection problem, repeated across theft, tenancy fraud, application fraud and misdeclared sites, is where AI earns its keep. Not as a detector of fraud, but as the thing that assembles and ranks candidates before a person decides who gets a knock at the door.

Consumption that stops while the lights stay on

The cleanest pattern in the data is a collapse with no plausible reason. A profile that ran steadily for years falls to almost nothing and stays flat, while the account name has not changed, the property is recorded as occupied, and the water account at the same address still draws normally.

The judgement is not in spotting the drop. Exception reports have done that for decades. It is in deciding whether the drop has an innocent explanation already sitting in your records: a solar install, a tenant on a prepayment meter, a faulty communications link, a site that genuinely closed. An agent can check each one before anything reaches an investigator, and say what it ruled out and on what evidence.

Interval data says more than a monthly total

Where interval data exists, the shape matters more than the volume: no overnight base load, a profile going dark for identical periods daily, a register advancing while the load says nothing is drawing. A model can put those patterns into words for a case note.

Loss that sits between the substation and the bills

Non technical loss is a subtraction problem with bad inputs. What was measured into a district, minus what the billed accounts consumed, minus the technical losses you can estimate, leaves a residue. The residue is real but it is not an address.

Most of the gap is not theft. It is unbilled accounts, meters mapped to the wrong feeder, estimated reads carrying a stale factor, vacant properties on an unmetered supply, asset records that never caught up with a network reconfiguration. An agent working through the connectivity model, the meter inventory and the billing extract can say how much of the gap is data quality before anyone calls it theft. That alone changes which districts get a field programme.

Narrowing a feeder down to doors worth a visit

Once you trust the district figure, the work is geography. You have a feeder or a pressure zone with an unexplained gap and several hundred premises on it. Investigators want a short list.

This is assembly rather than inference. Pull every account on the feeder, its read history and read type, how long reads have been estimated, the asset record and its last verification, prior visits at that address and next door, and any disconnection never followed by a reconnection order. Rank by how many independent signals agree. An investigator walking a street with a few addresses flagged and a reason for each is a different proposition from one handed the feeder.

The serial on the account and the serial in the photograph

Field visits generate the evidence that wins a case: the seal number from the last exchange, the meter serial on the account against the serial in the photograph, the terminal cover condition, the reading on arrival. Those artefacts sit scattered across the asset system, the mobile workforce app, the meter data system and the case file.

An agent can gather them into one view and flag the disagreements: an account whose recorded serial is 'K04T 118293' while the last site photograph shows another, a seal reference appearing on two assets, a register that went backwards between visits, a reconnection with no order behind it.

One address, four names, and a balance that keeps moving

Tenancy fraud looks nothing like theft in the data. Consumption is normal. What cycles is the account holder. A name opens an account, runs up a balance, closes it as a move out, and a new name opens at the same address with the same consumption pattern and no gap in supply.

Often the names are related, share a date of birth, or differ by a transposed initial. Sometimes the address is written four ways, so the system never sees the repetition. Resolving premises and people across billing, credit and move records suits AI because the matching is fuzzy and the volume large. The second job is restraint: students, houses in multiple occupation and short lets produce this pattern honestly, and a landlord between tenancies is not a case.

The applicant with no trace at the address

New connections and new accounts are the one point where you can act before the loss. The signals are familiar from identity work: an applicant with no trace at the address, a name on the application that does not match the tenancy agreement, a mobile number or bank detail already attached to a closed account in arrears, documents reused across applications.

Schemes intended for people who need them deserve particular care. Social tariffs and priority services registers are built to be easy to join. Screening here should pass almost everything through, ask for evidence only where the application contradicts something you hold, and record why a case was queried. A wrong decline here does not cost margin. It is somebody who needed help and was told no.

A site declared as one thing and drawing like another

Commercial misdeclaration is quieter than theft and often larger. A unit declared as storage runs a kitchen. A domestic account supplies a workshop. A premises on an unmeasured water charge took on a process load. A site has grown past the threshold where it should be metered differently.

The evidence is comparative: consumption against declared class, load shape against the stated activity, the classification on the account against what the site visibly does. An agent can rank accounts by the size of the mismatch and draft the question the account manager should ask. Reclassification and any backdated charge stay a commercial decision with a human name against it.

Reads keyed from the kerb, visits logged from elsewhere

Not all leakage comes from customers. A visit recorded as attended but never made, a read keyed as actual when it was estimated from the kerb, a meter exchange logged without the old asset returned, an inspection closed with a photograph taken at another property.

The patterns are findable: completion times too short for the work, visits closed outside the access window, reads falling suspiciously close to the estimate the system had already generated, an operative whose outcomes differ markedly from colleagues on comparable rounds. This work usually belongs to audit rather than revenue protection, but the analysis is the same shape.

Safety changes the order of the queue

Interfered gas work, a bypassed service, an illegal reconnection, exposed conductors inside a domestic property. The moment the records suggest any of that, the case stops being about revenue.

An agent triaging referrals has to escalate those indicators ahead of value: reports of smell or heat at the address, prior condemned appliance notices, a safety disconnection followed by renewed consumption, a visit report describing work the operative was not qualified to touch. Ranking by recoverable amount alone buries the case that needed a response today.

What the investigator should be holding before the knock

A good referral pack is short, sourced and honest about what is missing. Address and account history, the profile with the anomaly described in words, the asset and seal position, what the last field visit found, any prior case at the premises or in the name, the tenancy position, and the vulnerability flags that constrain how the visit is run.

It should also state the counter explanation plainly. If the drop coincides with a meter exchange, say so. If the only read in the window was estimated, say so. Packs that hide the weak points waste a trip and, worse, justify visits that should never have happened.

Where AI fails here

Deciding to visit a home is a human decision. A referral score is an input, never the authority. Whoever signs off the visit carries the consequence of being wrong about somebody's home.

Anything touching supply stays in human hands. Disconnection, a safety isolation, removing a meter as evidence. These are irreversible, regulated and occasionally dangerous, and no automated recommendation should trigger them.

Vulnerability is a brake, not a feature to be weighed. A priority services registration, debt support, a known health dependency on supply: each changes how and whether the case proceeds. Treating it as one more ranking variable is how a revenue programme ends up in a newspaper.

Backbilling needs a person. An agent can produce the working, but the figure that goes on a bill rests on assumptions a customer may challenge, and it is owned by whoever can defend it.

The labels are thin and late. Confirmed cases are rare, outcomes arrive months after the referral, and the accounts you never visited are not negatives. Treating them as clean flatters the system badly. Measure on cases you have already dispositioned, and accept that the sample is biased towards what your old rules found.

And it cannot repair your asset data. Where meters are mapped to the wrong premises, it reasons confidently from the wrong record and refers an innocent household.

Where to start

Take one district where the loss figure has been stubborn, and have an agent do a single job: assemble every account on it into a ranked candidate list with the evidence and the counter evidence stated for each. Let it decide nothing. Then compare its ranking against what your investigators found on their last programme in a comparable area.

That comparison is the business case. If the top of the list holds the cases your team would have chosen anyway, you have saved the assembly time, which is most of the work. If it finds cases they would have skipped, you have found revenue. If it keeps surfacing data quality rather than theft, you have learned more than either.

Adopt AI Partners builds these agents inside the metering, billing and field systems you already run, so a ranked referral lands in the case queue rather than another spreadsheet. If you want to talk it through, describe how referrals reach your investigators to hello@adoptaipartners.com.