Turning a Decade of AER Regulatory History into a Queryable Submission Intelligence Layer

Ausgrid is a Distribution Network Service Provider (DNSP) serving greater Sydney and the Hunter Valley, with more than 1.7 million customers across one of Australia's most complex electricity distribution networks. As a regulated monopoly under the National Electricity Law, its allowed revenue is determined every five years by the Australian Energy Regulator (AER) through a formal, multi-stage process. Two complete cycles now exist in the public record: 2019–24 and 2024–29. The 2029–34 cycle is the next horizon. Each determination involves billions of dollars in capital expenditure and operational cost allowances, reviewed against a regulator that publishes its reasoning in forensic detail and references its own prior decisions explicitly.

Each AER regulatory cycle produces thousands of pages across a structured document hierarchy: Regulatory Proposals, AER Issues Papers, AER Draft Determinations, Ausgrid Responses, Expert Witness Reports, Alternative Control Services determinations, and Final Determinations. Two complete cycles, 2019–24 and 2024–29, already exist in full. They contain every objection the AER has raised, every cost category Ausgrid has defended, every accepted and rejected justification, and every negotiated outcome. The AER references this history explicitly when assessing new proposals. Ausgrid's submission teams were not.

• 18–24 months per regulatory cycle the most time-intensive and highest-stakes recurring workflow in the business
• Thousands of pages across two completed cycles no single analyst has read them in full, no tool has indexed them with cross-cycle linkage
• 12 distinct regulatory entity types (Revenue Allowances, Capex Programs, Opex Categories, RAB values, Community Outcomes, AER Objections, AG Responses, Expert Witness positions) scattered across unlinked PDFs
• Zero systematic institutional memory when senior regulatory affairs staff turn over between cycles, critical negotiating context walks out the door
• Pattern blindness: identifying which arguments the AER has consistently accepted or rejected requires weeks of manual comparison across thousands of pages
• 2–4 hours of manual document search to answer a single factual question about a prior cycle determination too slow for a live submission process

Semantica ingested the complete public corpus of AER and Ausgrid regulatory documents across both cycles, parsing PDFs with structure-aware chunking that preserves the regulatory document hierarchy rather than breaking at arbitrary character limits. The output is a temporal knowledge graph in Neo4j where every extracted entity is linked to its source document, part, section heading, and page number, and where cross-cycle evolution is tracked through typed evolves_to edges. A hybrid retrieval layer combining Pinecone semantic search and Neo4j graph traversal answers plain-English questions with exact source citations in under 3 seconds. A two-stage LLM pipeline using step-back query decomposition followed by grounded synthesis with 7 specialist regulatory agent tools surfaces patterns that would otherwise require days of manual review.

MATCH (e1:Expense {category: "Vegetation Management", cycle: "2019-24"})
      -[:EVOLVED_TO]->(e2:Expense {category: "Vegetation Management", cycle: "2024-29"})
MATCH (e1)-[:AER_OBJECTION]->(o1:Objection)
MATCH (e2)-[:AER_OBJECTION]->(o2:Objection)
RETURN
  e1.allowance            AS allowance_2019_24,
  o1.reason               AS aer_objection_2019,
  o1.source_page          AS objection_page_2019,
  e2.allowance            AS allowance_2024_29,
  o2.reason               AS aer_objection_2024,
  o2.source_page          AS objection_page_2024,
  e2.allowance - e1.allowance AS delta_allowance

-- Returns: position history, objection text, source page refs, and allowance delta
-- Execution time:  < 3 seconds
-- Previously:      2–4 hours of manual search across two cycles of documents

Ausgrid's regulatory submission process is one of the highest-stakes recurring workflows in Australian energy infrastructure. The team preparing the 2029–34 determination will include people who were not in the room when the 2019–24 objections were negotiated. Without a systematic intelligence layer, institutional knowledge depreciates with every staff change and accumulates nowhere. Semantica does not replace the regulatory affairs team. It gives them memory that does not walk out the door: a queryable, provenance-backed institutional record that makes every prior cycle finding accessible, citable, and usable in the preparation of the next. For any regulated DNSP approaching an AER cycle, the question is no longer whether that institutional memory matters. It is whether it will be there when you need it.