AI for Data Entry in Utilities

A fitter pulls an old meter out of a terraced house. He writes the removed index, the new serial, the new index and a note about a corroded tail on a paper job sheet, and drops it in the van. The sheet reaches the office in a bundle held with an elastic band. Somebody keys it. One digit of the removed index is ambiguous, so they choose the likelier reading. The note about the tail has nowhere to go on the screen, so it lands in a free text box that no report ever opens.

Months later the customer disputes a bill. Billing holds one index and the asset register a different serial. Nobody did anything wrong. The keying step was the weakest link in a chain that was otherwise sound. That step is everywhere here, because networks and suppliers run on records typed by one person reading what another wrote.

The application arrives as a site plan and a question

A new connection request rarely arrives as a clean form. It arrives as a part completed application, a site layout exported from a design package, a load schedule, and an email explaining which of three plot numbering schemes is current. Somebody reads all of it and types a job: location, load, phases, pressure tier, route length, plot references, consent position.

Deciding that the load table on sheet two supersedes the figure in the email is the judgement an administrator makes before any field gets filled. Read every attachment, propose the fields, and show where each came from, so the checker confirms rather than types.

What the crew wrote on the back of the job sheet

Meter exchanges, removals, resites and faults are recorded in the field and keyed in the office. Even where there is an app, crews fall back to paper when the signal goes or the form has no box for what happened.

Serials, indexes and seal numbers are the easy part. The hard part is the margin note, often the only record of something that matters: meter box rotten, tails undersized, no isolation valve, access refused, suspected interference. Reading those into structured flags, with the crop of the handwriting attached, means the office confirms an image of four characters instead of re-reading a sheet.

The read the customer typed and the photograph of the dial

Customers submit reads through an app, a web form, an automated line or an adviser, and they transpose digits, include the red decimal dials, read a dial meter backwards, send a photograph instead of a number, or give the read for the meter next door in a shared cupboard.

A rule that only checks whether consumption looks plausible rejects good reads. Reading the photograph against the submitted number is the stronger check. If the customer typed "04519" and the dials show "05419", the transposition is visible, and the right output is a flagged candidate, never a silent correction.

Estimates that become bills

An estimate is a calculation until it is billed, and then it is a figure a customer is asked to pay. Proposing it, and explaining which reads and which profile built it, is useful. Publishing it needs a person's name on it.

One wrong field and the whole message comes back

Registration and change of supplier flows are unforgiving by design. The message matches what the central systems hold or it is rejected, with a terse reason. A meter point reference keyed with a letter O for a zero. A postcode that disagrees with the address held centrally. A measurement class the previous supplier changed and nobody reflected locally. Each one sends the registration back.

The repeating failures are learnable. You hold a history of rejected messages, their reason codes, and the corrections that made them pass. That is the material for a check that runs before a message is sent and names the wrong field.

Who was responsible when the keys changed hands

Move in and move out details decide who pays for what, and they arrive late and contradicted. A tenant calls with a date and a read, the agent emails a different date, and the landlord says the property was empty for the gap. Somebody keys an occupancy change, a closing read and an opening read, and if any of the three is wrong the first bill is wrong for both parties.

Commercial supply adds contract data. Tariff details are keyed from a signed schedule: term dates, standing charge, unit rates by time band, capacity, pass through terms. A rate keyed into the wrong band bills quietly and wrongly for the whole term. Reading the record back against the signed document is cheap and almost never done.

Red pen on the as-built, and the map that never caught up

A job finishes and somebody marks up a drawing. The main went in further from the kerb than designed, the last stretch of service is a different material, the valve sits under new footway. That markup has to become an asset register update and a GIS change, keyed by a records team reading someone else's pen.

The backlog is invisible until a crew is sent to a location that does not match the ground. Reading markups into proposed geometry and attribute changes lets a records officer confirm rather than redraw. Nothing should enter the layer a crew will trust next winter without a person accepting it.

Taking a fault note while the next caller is holding

During a storm the problem inverts. There is no document. There is a person on the phone describing a leak, a bang, a smell, a street with no supply, and an adviser typing into a fault record while the queue builds behind them. Under that pressure notes compress into shorthand and the structured fields get whatever is quickest.

Transcribing the call and proposing the fields, address, symptom, hazard present, properties affected, helps because the adviser can listen instead of type. Anything carrying a safety flag then goes to a named person before it touches dispatch or network state. A smell of gas, a reported shock, exposed conductors, water near an electrical supply. The system drafts. A human owns.

The report that begins with a strike and ends in two registers

Damage and third party interference reports arrive as contractor emails, crew photographs and control room notes. One event becomes a fault record, an asset condition update and a damage report for recovery, each typed separately, and the versions drift apart. One read should produce all three as proposals sharing the same evidence, so the reference linking them exists from the start rather than being reconstructed later by someone searching on an address.

A note in the call log is not a register entry

Priority services information turns up in passing. A customer mentions dialysis at home during an outage call. A neighbour says the occupier is blind. Those mentions sit in free text, the register entry never gets made, and the next time supply is at risk the record does not say what somebody already knew.

A system can spot a possible priority services signal in a note and put it in front of someone who will speak to the customer. It should not write the entry. These records decide who is warned first when the supply goes, and consent to hold them is a conversation, not an inference.

Consents, wayleaves and the folder nobody inherits

Land rights are the quietest data entry problem in a network. Wayleaves, easements, access agreements, restrictions on building over an asset. The agreements exist as scanned deeds and letters, often held by whoever negotiated them, and the structured record is partial because keying legal text into fields is nobody's priority until access is refused.

Extracting the operative facts, parties, plan reference, annual payment, review date, access conditions, is tractable work with a clear output. Not a decision. A register that tells a planner, before a crew is booked, that an agreement exists and what it says about turning up.

When nobody can say which record is the meter

Eventually the systems disagree. Billing holds one serial, the asset register another, the field record from the last exchange a third, and the technical details held centrally something else again. Somebody reconciles that by hand across four screens.

The supply point that exists twice

Duplicates breed under time pressure. An operator cannot find the supply, so they create it, and two records start accumulating history for one physical meter, reads landing on one and the exchange on the other. Matching records across systems and presenting evidence for a proposed merge is comparison a system does well. Executing the merge is not, because a wrong merge is harder to unpick than a duplicate is to live with.

Where AI fails here

It misreads handwriting with confidence, and confident misreading is worse than a gap. A fitter's seven that is really a one posts cleanly and sits in the asset register unchallenged. Handwritten indexes and serials need a deliberately high threshold and a second look, tuned by region, because handwriting habits are local.

It cannot supply what was never recorded. If the job sheet does not state the removed index, no system can extract one. The honest output is a flagged gap and a question for the crew. A system that fills the gap with a plausible figure manufactures future disputes.

It degrades quietly when inputs change. A design package exports drawings differently, a code list is revised, a manufacturer changes a serial format, and accuracy falls before anyone notices. Field level accuracy has to be monitored in production, not proved once at launch.

Several things must not move without a named human. Anything with a safety consequence, including every gas escape, every reported shock and every hazard note from the field. Any entry on the vulnerable or priority services register. Any estimated read that becomes a bill. Closing an asset record a crew will rely on the next time they are sent to that street, because the cost of that error is paid by someone standing in a hole. Creating a new supply point record. Any adjustment with an irreversible financial effect.

Where to start

Pick the document type with the highest volume and the clearest downstream pain. For most network businesses that is field job sheets. For most suppliers it is customer submitted reads or rejected registrations. Count what it costs now before building anything: how many documents, how many fields each, how long per document, how many registrations rejected for data reasons.

Set the gate before you start. Accuracy per field type, the share that posts untouched, the share held for review, and a named owner for the held queue. Adopt AI Partners builds this inside your own environment, against your own billing, asset and GIS systems, owned by your people afterwards.

To talk through what this looks like on your own records, hello@adoptaipartners.com reaches us.