AI uses for records
Utility billing adjustment dataset: rebuilding a bill after meter error
By SourceX Editorial · Reviewed by Noah Loul ·
8 sections
Short answer
Classify the error first, then fix the affected dates, recompute each billing period with corrected usage and the tariff in force, and only then apply the lookback limit that matches the correction's direction and the customer's class. A utility billing adjustment dataset packages that chain: meter and test facts, reading revisions, tariff versions, original bills, the approved adjustment and a check that ties them together.
Agent run
Utility billing adjustment
Key takeaways
- Classify the cause before calculating: meter error, billing error and tampering carry different rules.
- Recompute every affected period with the tariff in force at the time, splitting periods at rate changes.
- Apply the lookback last, because it depends on the direction of the correction and the customer's class.
- Grade the per-period figures, the window and the routing, not just the final adjustment.
- An agent working from records prepares the correction; a person approves it, and tampering always goes to a person.
What does it take for an agent to rebuild a bill after a meter correction?
It takes four decisions in a fixed order: classify the error, fix the affected period, recompute each billing period with the corrected usage and the tariff that was in force at the time, and then apply the lookback limit that matches the direction of the correction and the customer's class. Only after those steps does the adjustment amount mean anything, and the agent should hand the result to a person for approval rather than post it.
A utility billing adjustment dataset captures that sequence from real corrections. Each case links the meter facts, the reading revisions, the tariff versions, the customer class, the original bills and the approved adjustment, so a model can be trained or tested on the reasoning instead of on the final credit or charge alone.
The order matters because the rules depend on it. Under Pacific Gas and Electric's tariffs, meter error and billing error are defined separately, carry different examples and sit in different rules, so an agent that starts calculating before it has classified the cause can apply the wrong lookback to the whole case.
Meter error, billing error or tampering: which one is it, and why does that decide the lookback?
Decide the cause first, from the meter test and field records, because each cause carries its own adjustment window. PG&E Electric Rule 17 defines meter error as incorrect registration from a malfunctioning or defective meter and excludes billing error and unauthorized use. Rule 17.1 defines billing error as incorrect billing caused by the utility, the energy service provider or the customer, and lists examples such as incorrect meter reads, crossed meters, an incorrect meter multiplier and an inapplicable rate; field errors such as installing a meter incorrectly also count.
The windows differ by direction and class. The table below summarizes two published rule sets as examples. They are not a national standard: each state commission and each utility tariff sets its own, and the packet must cite the tariff sheet in force for the service territory and date.
| Situation | PG&E rule (California) | Texas PUC 16 TAC 25.126 | What the agent must check |
|---|---|---|---|
| Fast meter or billing error, customer overcharged | Refund up to three years, or only the months affected if the error period is known to be shorter | Back-billing that results in a credit is not limited when tampering is not involved | Whether the start of the error is known from a record |
| Slow residential meter, customer undercharged | Back-billing allowed for three months when the meter registers more than 25 percent slow | Additional charges limited to three months when no tampering is involved | Customer class and test result against the threshold |
| Slow nonresidential meter, customer undercharged | Three months for a Small Business Customer, three years for other nonresidential service, when more than 2 percent slow | Same three-month limit on additional charges absent tampering | Small business status as defined in the tariff, on the adjustment date |
| Billing error on a nonresidential account, undercharged | Three months for a Small Business Customer, three years otherwise | Covered where incorrect readings came from the distribution utility | Which party caused the error and when it began |
| Suspected tampering or unauthorized use | Excluded from meter error under Rule 17 | Additional charges up to six months where tampering affected the meter | Stop and route to the utility's revenue protection process |
What should the billing adjustment evidence and calculation task packet include?
It should hold every input the recalculation needs, with dates, plus the approved output and a check that compares them. The packet below is written so an energy operations team can adopt it as a case schema; replace the system names with your own customer information and meter data systems.
Two standard structures help. Green Button's ESPI model includes a UsageSummary for a billing period with fields such as billingPeriod, billLastPeriod, itemized cost line items, qualityOfReading and a required statusTimeStamp, and its interval readings can carry ReadingQuality entries. The Green Button Alliance describes UsageSummary as the place for billing statement information such as billing period dates, total consumption and itemized charges.
| Packet field | Contents | Typical source | Check before use |
|---|---|---|---|
| service_point and meter | Pseudonymous service point key, meter number, meter type, multiplier, install and removal dates | Meter data management and asset records | Multiplier and dates match the field work orders |
| test_result | Meter test date, method, registration error found and the tester's work order | Meter shop or field test record | The result crosses the tariff threshold that applies |
| error_class | Meter error, billing error or suspected tampering, with the evidence relied on | Test result, field notes, billing audit | Tampering cases leave the packet and go to a person |
| error_window | First and last affected dates, and whether the start is known or estimated | Work orders, read history, swap records | Known start dates cite a record |
| reading_history | Original reads, revised reads, estimates and each revision's timestamp and quality flag | Meter data management | Revisions are versions, not overwrites |
| tariff_versions | Rate schedule and tariff sheet in force for each affected period | Tariff library | Mid-period rate changes are split |
| customer_class | Residential, small business or other nonresidential, as defined in the tariff, with the date assessed | Customer information system | Class on the dates the rule tests |
| original_bills | Each affected bill's period, usage and line items as issued | Billing system or Green Button UsageSummary | Totals reconcile to the issued bills |
| recalculation | Corrected usage per period, rate applied, adjustment per period and the lookback cut applied | Analyst or agent output | Every number traces to a field above |
| approval_and_notice | Who approved, when, and the customer notice sent | Billing exception queue, correspondence | Approval precedes posting |
Which dated rule, asset fact or safety boundary changes the permissible next step?
Three kinds of fact change what the agent may do next: a date (when the error began, when a rate changed, when the customer's class changed), an asset fact (multiplier, meter swap, test outcome) and a boundary that takes the case out of the agent's hands entirely. Each should appear in the packet as a named constraint so a grader can see which one the agent ignored.
Dates cut the calculation. A known start date limits a refund to the affected months even when the tariff would allow a longer window, and a rate change in the middle of the window means two calculations, not one. A missing start date is not permission to use the maximum window; the agent should say the start is unknown and apply the tariff's rule for that situation.
Asset facts can turn a meter error into a billing error. A wrong multiplier entered at installation is listed as a billing error in PG&E Rule 17.1, not a meter error, so the same under-registration can carry a different lookback depending on which record explains it.
Some boundaries are absolute. An agent working from records never schedules field work, never tells a customer to open or touch metering equipment, and never decides a tampering question. It flags, routes and stops.
Checklist · 5 items
- Stop and route when field notes mention broken seals, bypasses or unauthorized use.
- Stop and route when the meter test record is missing or the test result does not reach the tariff threshold.
- Stop and route when the customer has disputed the bill or a complaint is open with the regulator.
- Split the calculation when a rate schedule, tariff sheet or customer class changed inside the window.
- Mark the start date as estimated, and say so in the output, when no record fixes it.
0 of 5 ticked. Tick each item as you go.
Illustrative case: a commercial meter found registering slow after a swap
This case is illustrative: the figures are invented to show the packet in use and are labeled as such, and the rule references assume a tariff written like PG&E's Rule 17. A small commercial account's meter is tested after a customer inquiry and found registering 4 percent slow. A work order shows the meter was installed 14 months earlier, which fixes the start of the error window. The account is classified as a Small Business Customer under the tariff, and the commercial rate schedule changed once, 5 months before the test.
Linked evidence produces the calculation. For each of the 14 billing periods the recorded usage is divided by 0.96 to give corrected usage; a period that recorded 9,600 kWh becomes 10,000 kWh, an under-registration of 400 kWh. Periods before the rate change are priced on the earlier schedule and periods after it on the later one. All of that is the full recalculation, but it is not the bill adjustment.
The lookback decides what may be charged. Because the customer was undercharged, the meter registered more than 2 percent slow and the account is a Small Business Customer, a tariff like Rule 17 allows back-billing for three months, so only the 3 most recent periods carry an adjustment. The recorded outcome shows a three-period back-bill on the later rate schedule, approval by a billing supervisor before posting, and a customer notice explaining the test, the periods and the basis.
The counterexample changes the direction. If the same meter had been registering 4 percent fast, the customer would have been overcharged, and the refund would cover all 14 affected months, because the error period is known and is shorter than three years. An agent that applies the three-period cut to both cases has learned a number, not the rule.
What proves the rebuilt bill is right, beyond an adjusted balance?
The posted adjustment, matched line by line to the recalculation and the approval, is the proof; an account status of adjusted or a balance that looks right is not. Grade the agent on the intermediate facts as well as the final figure, because a correct total can hide two compensating errors.
Pin each record to a time. ESPI's UsageSummary requires a statusTimeStamp, and its reading quality model assumes a reading is good unless a ReadingQuality entry says otherwise, so a missing flag does not prove a reading was validated. Keep the read revisions and their timestamps so the grader knows which version of the usage the agent was entitled to see.
| What is graded | Pass condition | Evidence that settles it | Common false pass |
|---|---|---|---|
| Error class | Matches the class supported by test and field records | Test record, work orders | Treating a multiplier mistake as a meter malfunction |
| Error window | Start and end dates match the record, or the start is marked estimated | Install or swap work order, read history | Using the maximum window when the start date is known |
| Per-period recalculation | Corrected usage and rate per period match the reviewer's figures | Reading history, tariff versions | Right total with wrong period split |
| Lookback applied | Window cut matches direction, class and tariff | Tariff sheet, customer class record | Same cut for credits and charges |
| Routing | Tampering, missing tests and disputes go to a person | Exception queue entry | A confident adjustment on a case that should have stopped |
How to keep customer usage and grid details out of a shared packet
Remove the customer, keep the arithmetic. Replace account numbers, service addresses, meter serial numbers and names with stable pseudonymous keys so periods and records still link, and keep the usage values, dates and tariff references that the task depends on. Interval usage can reveal occupancy and business hours, so share it at the billing-period level unless the task needs intervals.
Leave out infrastructure detail the task does not use: feeder and transformer identifiers, substation names and field crew assignments add little to a billing task and widen the confidentiality review. The energy operations reviewer confirms the case reflects the tariff in force, and a safety and confidentiality reviewer confirms nothing in the packet describes how to access or alter equipment.
Customer usage records may be covered by state privacy rules and commission orders as well as the utility's own customer agreements. Counsel should confirm the permitted use and what de-identification is required; de-identification on its own does not give the utility a right to license the records.
Offering corrected billing histories for AI through SourceX
Billing exception queues hold years of decided corrections with the evidence attached, which is what agent builders lack. For a utility that wants to license that history, SourceX runs the sourcing, rights review, contracting, preparation, delivery and payment that turn selected billing records into a licensed dataset for AI.
A utility or energy retailer can begin with a short company assessment that needs no data upload, and large interval archives can stay in the company's own storage or ship on encrypted drives rather than being hosted by SourceX. No billing record is delivered until the utility has approved scope and terms and signed.
Frequently asked questions
What is a utility billing adjustment dataset?
It is a collection of decided bill corrections, each linking the meter and test facts, the reading revisions, the tariffs in force, the customer class, the original bills and the approved adjustment. It is used to train or evaluate agents that prepare corrections for human approval.
Is a meter error the same as a billing error?
Not in the tariffs that define them. PG&E's rules treat a malfunctioning meter as meter error and a wrong multiplier, wrong rate or crossed meters as billing error, and the two can carry different adjustment windows.
Why can't the agent just use the maximum lookback?
Because the window depends on the direction of the correction, the customer class and whether the start of the error is known. A known start date limits a refund to the affected months, and back-billed charges are often capped far below the refund window.
Should the dataset include tampering cases?
Include them only as routing tasks, where the correct output is to stop and send the case to the utility's revenue protection team. Leave out any detail of how the tampering was done.
Can illustrative cases stand in for real corrections?
They help document the packet and test the grader, as long as they are labeled illustrative. Any training or evaluation claim should rest on approved real cases from the utility's own records.
Sources
- PG&E Electric Rule 17 defines meter error and sets refund and back-billing periods for fast and slow meters by customer class. Source
- PG&E Electric Rule 17.1 defines billing error, lists examples such as incorrect multiplier and crossed meters, and sets adjustment periods. Source
- Texas PUC rule 25.126 limits back-billed charges for non-compliant meters to three months absent tampering, six months with tampering, and does not limit credits. Source
- The Green Button ESPI UsageSummary summarizes usage for a billing period with billing period, bill amounts, cost line items, reading quality and a required status timestamp. Source
- ESPI interval readings can carry one or more ReadingQuality entries. Source
- ESPI treats a reading as good unless a ReadingQuality entry states otherwise. Source
- The Green Button Alliance describes UsageSummary as carrying billing statement information such as period dates, total consumption and itemized charges. Source
See if your company qualifies
A short company assessment. No data uploads are needed.