A high-bill complaint can require a call center representative to review usage, an estimated meter read, a rate change, payment history, weather, and choice provider pricing changes. Policy frames the decision. Experience determines which data deserves attention first and when conflicting signals require another look.
When that reasoning stays with a small group of people, routine work becomes dependent on availability of high-bill subject matter experts. Expert judgment then acts as a capacity constraint. Calls wait for escalation, customer satisfaction declines, similar situations receive different treatment, and newer employees learn by interrupting the people already carrying the most complex work.
The practical test is the ability to carry approved reasoning into daily call responses: assemble the relevant data, show the evidence behind a recommendation, and return ambiguity to the person accountable for the outcome.
This article examines why embedding expert judgment is architectural, not technical, and why it requires governance and daily evidence before expansion becomes a capital-justified decision.
One person knows, everyone else waits
Utilities have extensive policy knowledge repositories, tariffs information, standard operating procedures, and training materials. Those assets establish the requirements for a decision. They rarely capture the full sequence an experienced employee uses to interpret a live call.
A billing analyst may know that an estimated read deserves attention before a usage explanation. A call center representative may recognize that a recent catch-up bill adjustment changes the outstanding balance in an abnormal manner. An experienced reviewer may see that an apparently eligible adjustment conflicts with a prior arrangement or a rate condition. Each judgment rests on a combination of data, payment sequence, and context.
The dependency is easy to miss. A team can appear adequately staffed while its daily throughput depends on a few employees who know where exceptions hide. The operating exposure becomes visible during absences, high-volume periods, turnover, and unfamiliar calls. Work either waits for a subject matter expert or proceeds with less of the reasoning that usually catches an error before it reaches a customer or creates an incorrect transaction.
This is an institutional problem, not an individual performance problem. The utility has the expertise; the AI workflow lacks a reliable way to carry it into each eligible decision.
Showing your work changes who can do it
AI can help only when it represents the path from evidence to action. A response that identifies a likely cause without showing the data, conditions, and policy logic behind it leaves the employee with another item to validate. Incomplete or unclear responses do not reduce dependence on the scarce subject matter expert who knows how to resolve the call.
The more useful role is narrower and more consequential. AI can assemble account, weather, usage, tariff, and billing information; test that information against approved conditions; identify contradictions and edge conditions. A routine call may receive a supported explanation. A call with incomplete answers may trigger an unnecessary truck roll for a meter inspection. A conflicting or difficult call may go directly to a high bill call queue.
The utility decides which evidence matters, which combinations permit a response, and which conditions require human review. The AI applies that decision logic consistently within the scope it has been given.
Consider the difference in a customer-contact AI workflow. An employee facing a high-bill complaint can receive a concise explanation tied to the usage, the weather, a recent meter read, and the applicable rate period. When those conditions conflict, the same AI workflow can preserve the evidence and route the call to a billing specialist rather than forcing the employee to reconstruct the escalation path from memory.
The answer remains accountable because the resolution approach remains visible.
What escalations actually tell you
The hardest reasoning is often a sequence, not a missing policy. An estimated meter read may deserve attention before a usage explanation. A service order may alter the billing period under review. A rate condition or prior arrangement may make an otherwise eligible adjustment inappropriate. When those conditions appear together, policy may not state which input takes precedence or whether the call should proceed, pause, or escalate.
Those gaps show up in daily work. A reviewer may change a recommendation because a service order changes the period under review. An adjustment may be held because meter, usage, and account data conflict. A reopened call may reveal a rate change that the explanation missed. These events identify the missing decision logic, data condition, or escalation rule that must be defined before the AI workflow expands.
A higher escalation rate early in deployment can be a useful finding rather than a failed result. It shows where approved reasoning ends and where subject matter experts must resolve the difference before it becomes institutional decision logic.
CIS doesn’t need replacement for improvement
Carrying expert judgment into execution does not require replacing the systems that own utility data and transactions. CIS, meter, service, and ERP platforms continue to govern data and transactions related to bills, usage, work orders, and financial activity. Their controls remain central to operations, service activity, customer history, and auditability.
The change occurs in the decision layer between those systems and the person performing the work. AI can retrieve authorized information, apply the relevant logic, and present a permitted next step inside an existing AI workflow for service or operations. The source system receives only the action that falls within the utility’s established business model.
Operations leaders own the decision path and its escalation criteria. System owners preserve data quality, controls, and transaction integrity. Reviewers remain responsible for the ambiguous exceptions that fall outside the approved path. AI connects those responsibilities to daily work; it does not absorb them.
Which decisions stay consistent?
Leadership does not need to infer success from an AI model demonstration. The evidence appears in the executional decisions that move through the day-to-day AI workflow.
Decision consistency across teams shows whether similar circumstances produce similar treatment. Reviewer changes to recommendations identify where the AI workflow lacks a condition or rule. Calls held because data is missing or conflicting show where inputs cannot yet support a permitted response. Reopened calls and repeat contacts reveal when an earlier explanation missed something material. Incorrect decisions, high time to resolution, and dependence on scarce subject matter experts show whether the new path reduces exposure or merely relocates it.
These measures should be interpreted together. A shorter handling time accompanied by more repeat calls would not justify an AI solution. A modest reduction in escalations with stable outcomes and clear reviewer acceptance may be more meaningful, especially in an AI workflow where the wrong decision creates customer, financial, or regulatory consequences.
Shared judgment reduces the bottleneck
Before expanding embedded expertise beyond a contained pilot, leadership should establish clear gatekeeping. Leadership should also define how approved reviewer decisions update the AI workflow, so recurring exceptions become governed logic rather than local workarounds. A named owner must be accountable for the reasoning path, escalation criteria, and performance against baseline. That owner may be a subject matter expert, a business leader, or a compliance officer (whoever carries responsibility to revise the logic and owns the outcome).
Define which decisions proceed without escalation, which require review, and which remain fully manual. Write those boundaries, test them, make them defensible. Specify not just the approved decisions but the exceptions that trigger escalation.
Measurement should precede expansion. Establish baseline performance over four to eight weeks before broader production use. That baseline defines what normal looks like. Expansion to adjacent AI workflows or wider scope should not proceed until performance remains stable against it.
Larger investments should follow operating evidence. A small deployment that reduces work, maintains consistency, and earns reviewer trust becomes justification for broader investment. A deployment requiring ongoing escalations, producing inconsistent outcomes, or extending decision time should not expand regardless of technical accuracy.
Who owns the reasoning path?
Utilities will continue to depend on experienced people for interpretation, exception handling, and decisions. The strategic opportunity lies in making their approved reasoning available more broadly, so capacity does not rise and fall with the availability of a few specialists.
AI can carry expert judgment into daily execution only when the utility preserves evidence, keeps AI boundaries clear, and treats escalation as part of the operating design. The strongest modernization cases come from observed performance: consistent decisions, credible review outcomes, and source systems that retain control of the transaction. That evidence justifies expansion with measurable confidence rather than an AI outcome being promised.
How can your utility prove that AI carries expert judgment into routine calls without adding rework? Subscribe to The Utility Stack for executive briefings on AI modernization in regulated utilities.