Make responsible AI part of everyday operations

The right controls make useful AI easier to operate. People know what it may do, what evidence to expect, who is accountable, and what happens when something goes wrong. Those decisions belong in the workflow from the beginning.

This guide offers implementation questions, not a determination of regulatory compliance. Your utility’s applicable obligations and established operating procedures remain part of the design process.

Start with the consequence of a mistake

Assess each workflow by its possible effect and the authority it receives. Searching approved internal procedures, drafting a customer message, changing an asset record, and influencing physical operations call for different controls.

Consider who could be affected, how quickly an error could cause harm, whether it would be visible, and whether it could be reversed. Also assess the cost of an unavailable service. A tool can produce accurate outputs and still create an operational problem if staff cannot continue without it.

The NIST AI Risk Management Framework provides a voluntary approach to incorporating trustworthiness into AI development, use, and evaluation. Use it as a reference alongside the utility’s established risk processes. The practical controls below are an editorial starting point to adapt to the specific workflow.

Write an operating boundary

For every deployed use case, keep a short record of its purpose, owner, approved users, permitted sources, and allowed actions. Specify where the system must stop and ask for help.

Separate read access, drafting authority, and permission to commit changes. Give integrations only the access needed for the task. Recheck permissions when people change roles or new data sources are introduced.

Review the whole information path: retrieval, model processing, output storage, logs, support access, and retention. A deployment label such as “cloud” or “on premises” does not answer those questions by itself.

Make human review meaningful

Reviewers need more than an approve button. Show the source evidence, assumptions, unresolved conflicts, proposed changes, and the effect of accepting them. Allow correction, rejection, escalation, and a return to the existing process.

Assign reviewers who understand the task and have authority to decide. Make time for the review in staffing and value calculations. If the queue grows faster than people can assess it, the control is not operating as intended.

For an illustrative asset correction, show both old and proposed attributes and the supporting inspection. For a customer message, verify that account information belongs to the authenticated customer and that any operational status comes from an authorized source.

Evaluate the full workflow

Build a test collection from representative work, including exceptions. Keep sensitive evaluation material under appropriate access controls. Reserve cases for final evaluation rather than using every example to improve the system.

Test the situations most likely to expose a dangerous assumption:

• Relevant evidence is missing, contradictory, or outdated.
• Two customers or assets have similar identifiers.
• Retrieved material contains instructions that should be treated as untrusted content.
• A user requests information or an action beyond their permission.
• An integration times out or accepts only part of a change.
• The task falls outside the intended scope.

Measure supported answers, missed issues, incorrect actions, appropriate escalations, and the effort needed to reach an accepted result. Test abstention as a useful behavior when the evidence is insufficient.

Operate it like a maintained service

Keep a record of the sources and system versions behind consequential outputs. Monitor exceptions, user corrections, access failures, and unusual activity. Give users a clear route to report an incident without becoming technical investigators.

Choose who can pause the workflow and how unfinished work is recovered. Test the fallback before an outage makes it necessary. For changes that affect output or authority, repeat the relevant evaluation before release.

Review the operating boundary as the workflow expands. A successful drafting assistant does not automatically earn permission to publish, update customer accounts, or direct field activity. Each new action needs an explicit decision, supporting evidence, and an accountable owner.