A switching order that references Recloser 12A on Feeder 441 when the correct device is Recloser 12B on Feeder 443 is not a minor typo. It sends a crew to the wrong location, leaves an actual fault uncleared, and in a switching sequence can create a safety exposure that a crew cannot see coming. AI can draft work orders and switching narratives from telemetry in minutes. The hallucinated asset ID is the thing that will hurt you, and this lesson teaches you to catch it every time.

Why AI Work-Order Drafting Matters in Field Operations

A work order is an operational contract. It specifies exactly what work will be performed, on which equipment, under what safety clearance conditions, by whom, and in what sequence. In utility field operations, a poorly written work order is not just an administrative inconvenience. It is a potential safety event. The wrong equipment identifier sends a crew to energize a section that should be de-energized. A missing clearance requirement exposes a worker to live voltage. An incorrect restoration sequence creates an unintentional parallel path in a distribution system that has not been cleared for it.

Traditionally, work orders are drafted by an experienced planner or supervisor who knows the system, knows the field crew, and has been trained to recognize the specific safety checkpoints that a work order for this type of job on this type of equipment must contain. That knowledge is deep, procedural, and hard to transfer. It is also one of the categories of institutional knowledge most at risk in the Great Crew Change: an experienced switching supervisor who has written 3,000 switching orders over 20 years carries knowledge that cannot be fully captured in a procedure manual.

AI work-order drafting accelerates the production of that document from hours to minutes by reading telemetry data, prior work orders for the same equipment, and your system's standard operating procedures and generating a draft that already contains the required fields, the standard safety language, and the switching steps in the correct sequence. What AI cannot do is verify that the asset IDs it has used are correct, that the switching sequence it has proposed is safe for the current system configuration, or that the clearance requirements it has included are appropriate for the specific work scope. Those checks belong to a qualified supervisor.

AI drafts the order faster than any human typist. A qualified supervisor verifies every asset ID and every switching step before that order leaves the planning desk.

Telemetry Inputs to Work-Order Drafting

Modern utility work-order drafting workflows have access to telemetry data that previous generations did not. SCADA records show the current configuration of every device on the monitored network. Smart meters report the last known load on each service transformer. Distribution Automation devices report the status of automated switches, reclosers, and fault indicators. An EMS (Energy Management System) or ADMS records the operational state of the entire distribution topology in near-real-time.

When drafting a maintenance work order for a specific transformer or line section, this telemetry provides the factual basis for the work document. The SCADA record shows which devices need to be opened and closed to isolate the section. The meter data shows the number and approximate load of customers who will be de-energized. The device status history shows the last time each switching device was operated, which affects whether it needs to be tested before being relied upon in a clearance sequence.

The AI model uses this telemetry to populate the draft work order with the specific device identifiers, customer impact estimates, and switching steps appropriate for the job. The quality of the draft depends entirely on the quality of the telemetry data it reads. A SCADA record that has not been updated since a recent switching event will show a stale device status. A GIS record that has not been updated since a recent equipment replacement will reference a device that no longer exists. These data quality gaps are exactly the conditions under which an AI hallucination becomes indistinguishable from a data-sourced fact: both look like a confident assertion in the work order draft.

The Asset ID Hallucination Problem

Asset ID hallucination is the specific failure mode that makes AI work-order drafting dangerous without rigorous verification. A language model that has been trained on thousands of utility work orders develops a strong statistical sense of what a utility equipment identifier looks like. When it encounters a work request for "the recloser on Feeder 441 near intersection of Oak and Main," it will generate a device identifier that looks exactly like a real device ID. That identifier may or may not be the correct one for the specific device at that specific location.

The hazard is not random error. The hazard is a plausible, well-formatted device identifier that is wrong in a specific, consequential way. "RCL-441-12B" instead of "RCL-443-12B" will pass a visual inspection by someone who is not checking it against the authoritative device register. It will fail operationally when the crew arrives at the device location.

The control is simple but non-negotiable: every asset ID in an AI-drafted work order must be verified against the authoritative device register before the order is issued. In most utilities, the authoritative register is the SCADA database, the ADMS topology model, or the GIS system. The verification is a lookup, not a judgment. "Does device ID RCL-441-12B appear in the SCADA database as the recloser at this location?" If the answer is no, the work order draft contains a hallucination and must be corrected before the job proceeds.

Drafting Switching Narratives from Telemetry

A switching narrative is the sequential description of the switching steps required to safely isolate a section of the distribution or transmission system for maintenance, and the reverse sequence required for restoration. For anything more than a simple two-device isolation, switching narratives are complex documents that require careful attention to sequence, because switching steps performed out of order can create fault exposure, unintentional energization, or system configuration violations.

AI can draft a switching narrative from the current ADMS topology model faster than an experienced switchman can compose one from memory. The ADMS model contains the network topology, the device status, and the switching relationships that define which devices must be opened and in what order to achieve safe isolation. A model that has been given access to this topology data and has been trained on switching narrative formats can produce a first-draft narrative that already respects the isolation sequence, includes the required customer notification steps, and identifies the devices that need to be tested for absence of voltage.

The verification requirement for a switching narrative is more demanding than for a maintenance work order because the sequence matters as much as the device IDs. A switching narrative that contains the correct devices but proposes them in the wrong order can be more dangerous than one with an incorrect device ID, because an experienced switchman reviewing for device IDs might not catch a subtle sequence error that only an operator who knows the specific system configuration would recognize.

Switching narrative verification has three layers. First, every device ID must be verified against the authoritative topology model. Second, the sequence of steps must be reviewed by a qualified switching supervisor who traces through the proposed sequence against the current system one-line diagram to confirm that each step achieves the intended isolation without creating a hazard. Third, the restoration sequence must be reviewed with the same rigor as the isolation sequence, because restoration errors are at least as dangerous as isolation errors.

The Work-Order Verification Checklist

The verification process for an AI-drafted work order should be systematic, not ad hoc. A checklist approach ensures that critical verification steps are not omitted under time pressure. The following checklist applies to both maintenance work orders and switching narratives.

Device identity check: For every device identifier in the work order, look it up in the authoritative register (SCADA, ADMS, or GIS). Confirm that the ID matches the device at the specified location. Note any discrepancy and correct the draft before proceeding.

System configuration check: Confirm that the current system configuration (normal or post-switching) matches the configuration assumed in the work order. If the system is in a temporary configuration due to a prior switching event or outage, the standard switching sequence may not be appropriate. This check requires a current one-line or the ADMS display for the affected section.

Customer impact check: Verify the customer count and any special customers (hospitals, critical care facilities, emergency services, large industrials) that will be affected by the de-energization. Confirm that the required notification procedures have been initiated before field work begins.

Safety clearance check: Confirm that the clearance type specified in the work order matches the work scope. A clearance appropriate for a visual inspection task is not appropriate for a task requiring contact with energized components. This check requires the work scope description to be reviewed against the applicable safety standard.

Restoration sequence check: Review the restoration steps, not just the isolation steps. Restoration is where errors often occur because the sequence must be reversed and because field conditions may have changed during the maintenance period.

Authorization check: Confirm that the work order bears the signatures or electronic approvals of the required authorizing personnel before it is dispatched to field crews.

Worked Example: The Hallucinated Device Caught Before Dispatch

Here is a concrete example of how the verification discipline prevents an AI-generated error from becoming an operational incident. A planner is drafting a work order for a transformer replacement on Feeder 443 at Oak and Main. They enter the job details into an AI work-order drafting tool, which reads the ADMS topology for Feeder 443 and produces a draft isolation sequence: open sectionalizing switch SW-443-6, open recloser RCL-441-12B, test for absence of voltage at the work location, issue clearance.

The planner runs the device identity check. SW-443-6 appears in the SCADA database as the sectionalizing switch at the correct location. RCL-441-12B does not appear in SCADA as a device on Feeder 443. The correct device for that location is RCL-443-12B on Feeder 443. The AI model, trained on work orders that included both Feeders 441 and 443 with similar device naming conventions, produced a plausible-looking identifier that combined elements from both feeders.

The planner corrects the draft: RCL-441-12B becomes RCL-443-12B. The switching supervisor reviews the corrected order against the one-line, confirms the sequence is correct, signs the authorization, and dispatches the job. The crew performs the isolation, does the work, and restores the feeder without incident.

Without the device identity check, the draft would have been dispatched with the wrong recloser ID. The field crew, following the work order, would have attempted to open a device on Feeder 441 rather than Feeder 443. Feeder 443 would have remained energized at the work location. The test for absence of voltage step in the work order is the last line of defense before a crew contacts potentially energized equipment, but it is far better to prevent the wrong-device scenario from reaching the field.

Maintenance Narratives and Institutional Knowledge Capture

Beyond the formal switching sequence, a complete work order often includes a maintenance narrative that gives the field crew the contextual information they need to perform the job safely and effectively. This narrative typically covers the history of the equipment being worked on, any prior issues found during the last inspection, the access route to the work location, any non-standard conditions the crew should expect, and the customer notification status at the time of dispatch.

AI can draft this narrative efficiently from prior work order records, GIS notes, and inspection history. A planner who enters the equipment ID, the work type, and the planned date can receive a draft that already summarizes the equipment's maintenance history, pulls the access route from GIS, and includes a notation that the device had a marginal condition rating at the last inspection. That context improves the field crew's situational awareness before they arrive at the work location.

The verification discipline for the maintenance narrative is the same as for the technical content: every historical fact in the narrative must trace to a specific source record. If the narrative states "this transformer was last serviced on March 14, 2024, at which time oil levels were found normal," both the date and the finding must be confirmed against the maintenance history record in the work management system. AI models will occasionally insert plausible dates or findings that were not in the source record, particularly when the source records are incomplete or inconsistently formatted.

One of the highest-value applications of AI in work-order drafting is template refinement. An experienced switching supervisor who has issued 3,000 work orders over 20 years has developed a set of implicit standards for what a good work order contains. Some of these standards are captured in the utility's written procedures. Many are not: they live in the supervisor's judgment about which safety language to include for which type of job, which customer notification steps are appropriate for which circuit configuration, and which restoration steps need to be spelled out in detail versus which can be summarized.

By training an AI drafting tool on a corpus of that supervisor's approved work orders, the tool learns those implicit standards and applies them to new drafts. This is a form of institutional knowledge capture that has direct relevance to the Great Crew Change: when the experienced supervisor retires, the drafting tool continues to apply the standards they established to new work orders, giving less-experienced planners access to that accumulated judgment in the form of a draft that already reflects it.

The qualification is that institutional knowledge capture through AI is not perfect. The tool learns the patterns in the supervisor's approved work orders, which means it also learns any systematic errors or shortcuts that the supervisor may have established over time. A thorough training data review, combined with a check against the utility's current written procedures and safety standards, is required before deploying a work-order drafting tool that has been trained on historical orders. The goal is to capture the good practice, not to automate the legacy shortcuts.

ADMS integration also creates a specific data quality obligation. The ADMS topology model is the authoritative source of device identity and switching relationships for the AI drafting tool. Any ADMS model that carries stale device status from a prior switching event, or that has not been updated with a recent equipment replacement, is feeding the drafting tool outdated information that can produce work orders with wrong device IDs or incorrect starting conditions. Field operations teams who deploy AI work-order drafting must include ADMS model quality as a managed dependency: the model must be verified against field topology on a defined cycle, and any known discrepancies must be flagged in the work-order drafting tool so the reviewing supervisor is aware of them before sign-off.

Key Takeaways

• AI can draft maintenance work orders and switching narratives from telemetry and ADMS topology data in minutes, compressing a task that previously required hours of experienced planner time. The value is speed of draft production; the verification step cannot be compressed.
• Asset ID hallucination is the specific, predictable failure mode of AI work-order drafting. The model produces plausible-looking device identifiers that may be wrong in subtle, consequential ways. Every device ID in an AI-drafted work order must be verified against the authoritative device register (SCADA, ADMS, or GIS) before dispatch.
• Switching narrative verification requires three layers: device identity against the authoritative topology, sequence review by a qualified switching supervisor tracing through the proposed steps, and restoration sequence review with the same rigor as the isolation sequence.
• The quality of the AI draft is bounded by the quality of the underlying telemetry and GIS data. Stale SCADA records and outdated GIS equipment registers are the conditions that make AI hallucinations hardest to detect, because the wrong device ID looks exactly like a plausible, data-sourced device ID.
• The verification checklist ensures that critical checks (device identity, system configuration, customer impact, safety clearance type, restoration sequence) are performed systematically before every work order is dispatched, regardless of time pressure.
• The institutional knowledge value of AI work-order drafting is highest in the Great Crew Change context: it allows a less-experienced planner to produce a well-structured draft that an experienced switching supervisor can verify in minutes, rather than composing the entire order from institutional memory.
• ADMS model quality is a managed dependency for AI work-order drafting. A topology model with stale device status or unrecorded equipment replacements produces drafts with wrong device IDs; the reviewing supervisor must know which parts of the model have known discrepancies before signing off.
• The authorization step is non-negotiable. A qualified switching supervisor, not the AI tool, is the accountable party for every work order that goes to field crews. The cardinal rule applies here: the model drafted it, a qualified human authorized it, and that human is accountable for its correctness.