Start with a workflow, not a model
A good first generative AI use case has a recognisable user, a repeated task, accessible source material, and a way to check the answer. “An assistant for the business” is too broad. “Help maintenance planners find the current procedure and its source” gives a team something concrete to evaluate.
The examples below are candidate workflows, not claims of proven savings. Use them to build a shortlist with your operations, data, and risk teams. Language models can help people search, summarise, and draft. Their ability to produce fluent text does not establish that an operational decision is correct.
Six workflows worth exploring
1. Maintenance knowledge retrieval
User: a maintenance planner preparing a work package. Input: approved manuals, current procedures, and relevant maintenance history. Ask the assistant to find an answer with a page-level citation and document version. Measure whether it retrieves the right source and reduces search time after the planner checks the answer.
Boundary: it must not invent procedures or override the approved work-management process. An answer without reliable evidence should say that the source was not found.
2. Shift-handover drafting
User: an operations supervisor. Input: authorised logs, open actions, and agreed handover fields. Generate a draft that separates unresolved issues from completed work. Compare it with a supervisor-written baseline for missing facts, incorrect statements, and review time.
Boundary: the supervisor approves the record. A confident summary must not hide an uncertain or missing source.
3. Engineering document comparison
User: an engineer reviewing a revision. Input: two controlled versions of a specification. Ask for changed requirements with references to both versions. Measure important changes missed and false changes flagged, as well as total reviewer effort.
Boundary: generated comparisons support engineering review; they do not approve a design or replace change control.
4. Utility customer-service assistance
User: a service agent answering a customer query. Input: current, approved service policies and information the agent is authorised to see. Draft an answer with the relevant policy attached. Evaluate policy accuracy, escalation quality, and the agent’s total handling time.
Boundary: protect customer information and leave account changes behind the existing authorisation process.
5. Supplier and procurement review
User: a procurement analyst. Input: supplier responses and an approved evaluation rubric. Extract evidence into a comparison table, explicitly showing missing information. Check extraction accuracy and reviewer time on a representative sample.
Boundary: a model-generated score must not silently become an award decision. Keep the source evidence and accountable reviewer visible.
6. Renewable-asset reporting
User: an asset manager preparing a recurring report. Input: validated performance tables, maintenance notes, and reporting definitions. Draft a narrative that cites the source period and distinguishes observations from explanations.
Boundary: calculate numerical measures in validated systems. Check every generated number and avoid presenting a guessed cause as an established finding.
Choose the smallest useful pilot
For each candidate, write down the task frequency, current completion time, available documents, review requirements, and the cost of a wrong answer. Reject candidates whose required data cannot be used or whose critical outputs cannot be checked.
Then compare the remaining options on measurable value, data readiness, implementation effort, and risk. Do not let a high average score cancel out a serious access or safety problem. Prefer one workflow where a small team can establish a baseline and observe real use.
A useful pilot question: Can planners find an approved answer more quickly, without increasing incorrect or unsupported answers? This is more testable than “Can AI improve maintenance?”
Next, use the six-week pilot roadmap to define your evaluation and decision gate.
Further reading
For a broader view of energy AI, see the U.S. Department of Energy’s artificial intelligence resources. For a structured approach to identifying and managing AI risks, see the NIST AI Risk Management Framework. The workflow examples above are planning scenarios from AI Energy Leaders, not performance claims from these organisations.