Systems and records
Utility administrative escalation dataset: which cases leave automation
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Take a utility case out of automation when it reports a possible hazard, touches a customer protection such as a medical designation with collection pending, rests on estimated or disputed readings or a mid-period tariff change, asks for protected infrastructure detail, or needs authority the automated step lacks. A utility administrative escalation dataset labels that stop decision, its evidence and when it became knowable, under a brief a domain reviewer signs.
Checklist
Utility administrative
Key takeaways
- Hazard reports in any channel leave the administrative workflow at once and go to the emergency process.
- Protected-customer designations matter most when a collection or disconnection step is pending.
- Estimated, edited or failed-check readings turn a billing dispute into a judgment call; keep the quality codes.
- Label when a trigger became knowable, not only when a person acted, so late escalations can be measured.
- Include near misses that correctly stay automated, or the model will escalate everything.
- Exclude infrastructure detail, emergency procedures and medical facts beyond a dated designation flag.
Which utility service cases should be removed from automated handling?
Remove a case from automation when it reports a possible hazard, touches a customer protection that changes what the utility may do, rests on a reading or rate that is estimated, corrected or in dispute, asks for protected infrastructure detail, or needs a decision outside the written authority of the automated step. Everything else, routine and well-evidenced administrative work, can stay automated with sampling.
Those five classes are not a matter of taste. Gas operators, for example, must have written emergency procedures that cover receiving, identifying and classifying notices of events that need an immediate response, including gas detected inside or near a building, under 49 CFR 192.615. A billing bot that receives an odor report in a chat about a payment plan has received exactly that kind of notice.
A utility administrative escalation dataset records those stop decisions so that a model can learn them: the trigger, the evidence that made it knowable, the moment it became knowable, where the case went and what happened. This guide gives the criteria table, the dataset brief a domain reviewer signs, a labeling schema and an illustrative case.
Why must the stop decision be labeled, not only written into a prompt?
Because a model trained on cases that automation finished teaches itself that every case can be finished. The escalations, the near misses and the cases that should have left earlier are the examples that carry the boundary, and they only exist if someone labels them.
The NIST AI Risk Management Framework asks organizations to define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems (GOVERN 3.2). A labeled escalation set is how that policy becomes testable: a reviewer can measure whether the system stopped where the policy says a person takes over.
There is also a regulatory edge to watch. The EU AI Act lists AI systems intended as safety components in the management and operation of the supply of water, gas, heating or electricity as high-risk in Annex III. A billing assistant is normally not a safety component, but a router that decides whether a hazard report reaches the emergency desk sits closer to that line. Counsel should classify the specific system; the dataset should make the hazard route visible so that classification is possible.
Which triggers move a utility case to a person, and on what evidence?
Each trigger below changes what the next step may be, and each depends on a fact with a date: when the report arrived, which designation was active, which tariff was in force, which reading quality code applied. The table is the core of the criteria.
| Trigger | Evidence field that shows it | Why automation stops | Route |
|---|---|---|---|
| Odor, leak, fire, downed line or flooding near equipment mentioned in any channel | Intake text, channel, timestamp | Possible emergency notice; gas operators must classify and respond under written procedures | Emergency desk immediately; the administrative task pauses |
| Medical or critical care designation on the account while a collection or disconnection step is pending | Designation flag with effective and expiry dates; collection stage | In Texas, for example, a critical care designation requires written notice to the customer and an emergency contact at least 21 days before disconnection | Credit and collections specialist |
| Bill built on estimated, edited or failed-validation readings and the customer disputes it | Reading quality codes per interval; bill version | The amount is not yet a fact; a rebill needs judgment | Billing analyst |
| Rate or tariff changed inside the billing period | Tariff id and effective date; service period dates | Proration and which rate applies need a dated rule check | Rates or billing analyst |
| Negotiated commercial contract term overrides the standard tariff | Contract id, clause, effective dates | The automated step only knows the standard tariff | Account manager |
| Request for maps, protection settings, network or access details | Request text; requester identity | May be protected infrastructure information | Security or CEII coordinator |
| Third party asks for usage or account data | Requester, authorization record | Authority to disclose is unverified | Privacy or customer authorization team |
| Field work order needs switching, isolation or site access decisions | Work order type and status | Field safety procedures are outside administrative scope | Operations dispatcher |
| Regulator complaint, legal notice or media inquiry | Channel, sender, keywords | Response has formal deadlines and owners | Regulatory affairs or legal |
What should the professionally reviewed escalation criteria dataset brief include?
The brief is the document a domain reviewer signs before any case is selected. It fixes scope, triggers, labels and exclusions so the dataset cannot drift toward whatever records were easiest to export. Use the table as the brief's outline.
At a glance
What should the professionally reviewed escalation criteria dataset brief include
| Brief section | What it fixes | Signs off | Failure if missing |
|---|---|---|---|
| Task boundary | Which administrative tasks are in scope, such as move-in, billing disputes, payment plans, outage status | Operations lead | Cases from field or control-room work slip in |
| Trigger taxonomy with version | The trigger list above, with definitions and examples | Operations, safety and customer protection leads | Labelers invent their own reasons |
| Jurisdiction and rule versions | Which state or national rules apply, with the version date used | Regulatory affairs | A rule change silently relabels history |
| Evidence fields per case | Linked records each case must carry | Data owner | Labels cannot be checked |
| Label schema | Decision, reason code, knowable-at time, route, outcome | Technical reviewer | Timing errors are invisible |
| Negative and near-miss quota | Share of cases that look risky but correctly stay automated | Technical reviewer | The model learns to escalate everything |
| Exclusions | Infrastructure detail, emergency procedures, medical facts beyond a designation flag | Security, safety and privacy leads | Protected content enters a dataset |
| Adjudication | Two reviewers per case, a tie-break owner, how disagreements are recorded | Domain reviewer | No measure of label reliability |
| Refresh triggers | Tariff, rule, designation program or scope changes that force a relabel | Data owner | A stale boundary is taught as current |
How should each escalation case be recorded?
Record each case as a linked bundle with one decision label, and keep the time a trigger became knowable separate from the time a person acted. That gap is what late-escalation labels measure.
Have two reviewers label every case independently and keep both labels, not only the agreed one. The share of cases where they disagreed, broken down by trigger, tells the dataset owner which criteria are still ambiguous; a trigger with frequent disagreement needs a sharper definition in the brief before more cases are labeled, not a third reviewer to outvote the other two.
Keep the automation trace even when it is embarrassing. A case where the assistant offered a payment plan to a customer who had just reported a smell of gas is among the most valuable records in the set, because it shows exactly the failure the dataset exists to prevent.
- case_id, jurisdiction, channel (phone, chat, email, portal) and a tokenized account reference.
- Redacted intake text and every later customer message, each with a timestamp.
- Asset and service point ids, meter ids and the reading series used for the bill, keeping each reading's quality code.
- Tariff id and effective date, bill version and any rebill, with the service period.
- Contract or special-agreement id and clause, if any.
- Customer protection flags as booleans with effective and expiry dates, never the underlying medical detail.
- Work order ids and types linked to the case, without switching or isolation steps.
- Automation trace: the last automated step reached and the action it proposed.
- Label: stay automated, escalate, or emergency route; reason code from the taxonomy; trigger knowable-at time; escalation time; route.
- Outcome and the reviewer pair's adjudication note, including disagreements.
Why do reading quality codes decide so many billing escalations?
Because an estimated reading and an actual one look identical once the code is dropped, and a dispute about an estimate is a different task from a dispute about a measured value. Keep the codes and let them drive the label.
The Green Button ESPI schema for sharing energy usage data defines reading quality codes that make the difference explicit: valid data that passed validation, values replaced or approved by a human, values replaced by a machine-computed estimate from historical data, values computed by linear interpolation, data that failed one or more checks, and raw data that has not gone through validation. A dispute about a bill whose intervals carry estimate or failed-check codes is a candidate for escalation; the same dispute on validated reads often is not.
The dataset should store which code each interval carried when the bill was issued and when the dispute arrived, since later validation can change it. Labeling against today's code teaches the model a fact it could not have known.
Illustrative case: a high-bill dispute that should have left automation on day two
This case is illustrative; the dates and counts are invented to show the method. A residential customer writes to the billing assistant that a bill is about three times the usual amount. The account carries an active critical care designation, and the account is two weeks into a past-due collection sequence.
Linked evidence shows the bill used 31 daily intervals, of which 12 carry the machine-estimate code. The automated assistant offered a payment plan on day one. On day two the customer replied that nobody had been able to read the meter. At that point two triggers were knowable: a dispute over estimated reads and a protected designation with collection pending.
The correct label is escalate, reason codes estimated-read dispute and protection-plus-collection, knowable at the day-two message, route to a billing analyst with a hold on the collection step. In the source history the case reached a person on day six. The dataset records both times, so a model can be scored on how early it would have stopped.
The near miss: another customer with the same designation asks only for the due date of a bill built on validated reads, with no collection step pending. That case stays automated, and including it is what keeps a model from routing every protected account to a person.
Where must the dataset stop: infrastructure detail, emergency procedures and medical facts?
Three kinds of content never belong in an administrative escalation dataset, even when the cases that mention them do. Exclude them by rule before records are selected, and narrow the scope if a scan keeps finding them.
Infrastructure detail comes first. Federal rules define critical energy infrastructure information as specific engineering, vulnerability or detailed design information about proposed or existing critical infrastructure that could be useful to a person planning an attack and that does more than give the general location (18 CFR 388.113). Keep a request for such detail as a labeled trigger, but drop any attachment or note that contains it.
Emergency and field procedures come next. The label can say a hazard report went to the emergency desk; the dataset should not contain the response steps, switching orders or site access instructions.
Medical facts are last. In Texas, for example, a critical care residential customer is one with a resident diagnosed by a physician as dependent on an electric-powered medical device to sustain life, and the designation does not relieve the customer of the obligation to pay. The dataset needs only the designation flag and its dates. When a reviewer cannot tell whether a field crosses one of these lines, stop and get the security, safety or privacy lead to decide before the case is kept.
Can a utility or energy retailer license its escalation histories?
Sometimes, after a rights and privacy review of the specific records. Customer terms, state rules on customer data, contracts with billing or CIS vendors and infrastructure protection duties all bear on it, and having access to the records does not settle any of them. De-identification helps but does not by itself create a right to license.
Where a utility or retailer decides to go ahead, SourceX runs the rights review, contracting, preparation and delivery for the selected records, with de-identification handled by SourceX through approved processing routes. Eligibility comes first, through a short company assessment that asks for no records, and the utility signs off on scope and terms before anything is delivered.
Frequently asked questions
What is a utility administrative escalation dataset?
It is a set of labeled utility service cases, such as billing disputes, payment plans and move requests, where each case records whether automation should continue, escalate to a person or route to emergency handling, why, and when that became knowable. It trains or tests the stop decision rather than the completed task.
Should a billing chatbot ever handle a gas odor report?
No. It should hand the report to the emergency process at once and pause the administrative task. Gas operators must have written procedures for receiving, identifying and classifying emergency notices, and an odor report is one of them.
Do medical designations always force escalation?
Not always. The designation matters when a collection, disconnection or service change is pending, because notice and contact duties can apply. A protected customer asking a routine question can stay in automation, and the dataset should include such cases.
Why keep near-miss cases in the dataset?
Without them a model learns that any risky-looking signal means escalate, which floods specialists and hides real emergencies. Near misses teach where the boundary is, and they let reviewers measure false escalations as well as missed ones.
Who should review the escalation criteria?
An energy operations expert with safety and confidentiality responsibility, plus the customer protection, privacy and security leads for their parts. Two reviewers should label each case, with a named owner to settle disagreements.
Sources
- 49 CFR 192.615 requires each gas operator to have written emergency procedures, including receiving, identifying and classifying notices of events that need immediate response, such as gas detected inside or near a building. Source
- 18 CFR 388.113 defines critical energy infrastructure information as specific engineering, vulnerability or detailed design information useful to a person planning an attack that does more than give general location. Source
- Texas PUC rule 25.497 defines a critical care residential customer by a physician's diagnosis of dependence on an electric-powered medical device and says the designation does not relieve the obligation to pay. Source
- The Green Button ESPI 4.0 schema defines reading quality codes for validated, human-replaced, machine-estimated, interpolated, failed-check and unvalidated data. Source
- Texas PUC rule 25.483 requires written notice of intended disconnection to a critical care residential customer and the emergency contact at least 21 days before the date, and utility contact before disconnecting. Source
- NIST AI RMF 1.0 GOVERN 3.2 calls for policies that define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems. Source
- The EU AI Act, Annex III point 2, lists AI systems intended as safety components in the management and operation of the supply of water, gas, heating or electricity as high-risk. Source
See if your company qualifies
A short company assessment. No data uploads are needed.