
Energy planning brings together documents that describe different parts of a system: demand projections, capacity proposals, maintenance needs, permitting conditions, and community priorities. AI can help organize those sources into a brief that makes the assumptions and unanswered questions easier to inspect.
Start with public documents and a defined planning question. The first deliverable should be an evidence table and questions for an engineer or planning team. A document assistant is a different system from an application connected to live grid operations.
On this page
First build: a planning evidence brief
Use three fictional documents: a facility demand summary, a generation proposal, and a public consultation note. The assistant extracts the quantities and assumptions each document actually states, identifies where comparisons are incomplete, and drafts a short review brief.
Prerequisites are an approved source list, document versions, a field schema, and a qualified reviewer. Keep nonpublic infrastructure details and live operational feeds outside this prototype. Establish a source owner before adding internal material later.
The workflow is: approved collection → extraction with citations → unit and assumption checks → comparison table → engineering questions → reviewed brief.
Build it step by step
- Write one planning question, such as “What evidence is missing before these two proposals can be compared?” Avoid giving the model an open-ended optimization objective.
- Register each source with title, date, version, geographic scope, and page references. Distinguish a proposal from an approved decision.
- Extract reported quantities with their units, periods, scenario names, and assumptions. Leave absent values empty.
- Validate units and arithmetic with ordinary code. Record conversions explicitly and preserve the original values. Do not invent a capacity factor, load shape, or operating period.
- Ask the model to summarize agreements, differences, and missing information. Require a source for every factual statement and label analyst questions separately.
- Have the reviewer inspect the cited pages and assess technical meaning. A numerically correct comparison can still involve incompatible system boundaries.
- Save the reviewed brief with source versions and outstanding actions. When a proposal changes, produce a new comparison rather than silently overwriting the earlier record.
A reusable instruction is:
Compare only the supplied fictional planning documents.
Preserve units, time periods, geography, and scenario assumptions.
Separate proposed, modeled, measured, and approved quantities.
Do not infer missing operating assumptions or system adequacy.
Return an evidence table and questions for an engineering reviewer.
A fictional comparison
Proposal P-A states a nominal capacity of 5 MW. Demand note D-B reports annual energy use of 18 GWh. The proposal does not provide expected annual output or a generation profile.
A useful output records both quantities and says they cannot establish adequacy by themselves. It asks for the expected output profile, demand timing, availability assumptions, and the relevant system boundary. It does not compare 5 with 18 as though they were the same measurement or claim that the proposal meets the facility’s needs.
A public consultation note can add concerns that require investigation, such as construction access. Preserve those as attributed concerns, rather than converting every comment into an engineering finding.
Keep the opportunity and evidence distinct
DOE’s April 2024 AI-for-energy report discusses opportunities across planning, permitting, operations, reliability, and resilience. It is useful background for exploring applications; it is not evidence that a particular assistant has been validated or a statement of today’s requirements for a project. DOE AI for Energy
If the project later reaches operational technology, involve the operations and security owners in a separate architecture review. NIST’s OT guidance addresses physical-system reliability and safety alongside security; apply that context to the actual environment. NIST SP 800-82
Test the comparison before trusting the brief
Include mixed units, inconsistent time periods, an outdated proposal, a missing appendix, a chart whose labels extract incorrectly, and two scenarios with similar names. Check that the system preserves uncertainty and refuses unsupported conversions. Include a document that tries to change the assistant’s instructions.
Measure evidence coverage, unit accuracy, valid citations, and reviewer correction time. Have the reviewer identify any material source the brief omitted. Do not infer improved energy performance from a document-processing pilot.
A small facility can start with manually selected public documents. A municipal utility or medium operator may add controlled planning files and team review. A large utility needs source ownership across engineering, operations, regulatory affairs, and security, plus separation from live control systems.
Budget for document quality, domain review, secure integration, and periodic reevaluation. The practical outcome is a more inspectable planning decision, with the assumptions visible to the people responsible for it.
Continue with workflow design, enterprise implementation, and AI governance.
Keep building
Next guide: Build an enterprise AI service people can depend on