Automated ESG Reporting with Agentic AI for Enterprise Sustainability Compliance

gardenia_technologies · Consulting · 2025 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Sustainability compliance

Gardenia Technologies needed to reduce the administrative burden of ESG disclosure work, where enterprises must collect structured and unstructured data across many systems and complete large reporting frameworks accurately.

Constraints:

• ESG frameworks can contain hundreds or thousands of required data points
• data resides across databases, document stores, emails, prior reports, public web sources, and reporting questionnaires
• regulatory and organizational context require human expert validation
• usage is spiky around quarterly and annual reporting cycles

Architecture

Report GenAI is a serverless agentic ESG-reporting assistant on Amazon Bedrock. Users configure databases, document stores, public search, and a report specification, then an AWS Lambda-hosted LangChain ReAct executor uses Claude Sonnet/Haiku on Bedrock to iterate through disclosure questions. It calls a web-search Lambda, a text-to-SQL Lambda over Gardenia's emissions database, and a Lambda-hosted RAG tool backed by Faiss indexes persisted to S3. A Step Functions embedding pipeline uses Textract and Titan embeddings to build the vector store. Users review cited answers, calculation methods, and agent reasoning through a Streamlit UI on ECS Fargate before finalizing responses.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  streamlit-ui[Streamlit UI] --> amazon-cognito[Amazon Cognito]
  streamlit-ui --> agent-executor[Agent executor]
  agent-executor --> amazon-bedrock-models[Amazon Bedrock models]
  agent-executor --> report-specification[Report specification]
  agent-executor --> web-search-tool[Web search tool]
  agent-executor --> text-to-sql-tool[Text-to-SQL tool]
  text-to-sql-tool --> emissions-database[Emissions database]
  agent-executor --> rag-tool[RAG tool]
  rag-tool --> faiss-vector-index[Faiss vector index]
  embedding-pipeline[Embedding pipeline] --> amazon-textract[Amazon Textract]
  embedding-pipeline --> amazon-titan-text-embeddings[Amazon Titan Text Embeddings]
  embedding-pipeline --> faiss-vector-index
  llm-judge[LLM judge] --> amazon-bedrock-models
  llm-judge --> agent-executor

Components

Data flow

A user configures Report GenAI with authorized ESG databases, document stores, web search, and a report specification. The agent batch-fills disclosure questions by planning actions, calling the text-to-SQL tool for numerical database questions, the RAG tool for internal documents, and web search for public context. Retrieved evidence and computed values are turned into responses matching the required format. ESG experts review each cited answer and calculation method, can chat with the assistant or edit directly, and finalized reports are added back to the document store for future reporting cycles.

Why this architecture

The architecture reflects ESG data reality: no single retrieval method can answer all disclosure questions. A ReAct agent coordinates specialized tools for public search, structured emissions data, and internal documents while Pydantic ReportSpec keeps output aligned to reporting frameworks. Serverless Lambda, ECS Fargate, and on-demand Faiss loading fit intermittent reporting demand. Human review and LLM-judge layers are present because ESG reporting carries regulatory and business implications, and the team needs traceable sources and calculation methods rather than unverified generated prose.

Models used: Anthropic Claude Sonnet 3.5 · Anthropic Claude Haiku 3.5 · Amazon Titan Text Embeddings · Amazon Nova models

Scale

• 55% of sustainability leaders cite excessive administrative work, and 70% say reporting demands inhibit strategic initiatives.
  “55% of sustainability leaders cite excessive administrative work in report preparation, while 70% indicate that reporting demands inhibit their ability to execute strategic initiatives”article, stated
• 96% of the 250 largest companies report on sustainability progress.
  “Sustainability disclosures are now a standard part of corporate reporting, with [96% of the 250 largest companies reporting on their sustainability progress]”article, stated
• CSRD includes 1,200 individual data points, while CDP has about 150 questions.
  “The EU Corporate Sustainability Reporting Directive (CSRD) framework, for example, comprises of 1,200 individual data points”article, stated

  “Even voluntary disclosures like the CDP, which is approximately 150 questions, cover a wide range of questions related to climate risk and impact, water stewardship, land use, and energy consumption.”article, stated
• Omni Helicopters International cut CDP reporting time by 75%, from one month to one week.
  “Omni Helicopters International cut their CDP reporting time by 75% using Gardenia’s Report GenAI solution. In previous years, OHI’s CDP reporting required one month of dedicated effort from their sustainability team.”article, stated

  “In previous years, OHI’s CDP reporting required one month of dedicated effort from their sustainability team. Using Report GenAI, OHI tracked their GHG inventory and relevant KPIs in real time and then prepared their 2024 CDP submission in just one week.”article, stated

Alternatives considered (roads not taken)

Amazon Bedrock Knowledge Bases end-to-end retrieval

Rejected because: Gardenia wanted full control over extraction, embedding model choice, and vector database configuration, accepting higher management overhead.

Always-on vector database

Rejected because: ESG reporting has intermittent, spiky usage, so an in-memory Faiss index persisted to S3 and loaded on demand better matches cost and utilization.

Manual questionnaire completion

Rejected because: Manual reporting leaves sustainability teams managing data collection and questionnaires instead of strategic ESG work.

Patterns

tool-use-orchestration (agent-architecture)

The ReAct executor uses Bedrock models to choose among web search, text-to-SQL, RAG, report-specification, and review tools for each disclosure question.

rag-retrieval (knowledge-access)

The RAG tool retrieves relevant text from corporate document stores via a Faiss vector index persisted on S3.

structured-output-contract (output-control)

Pydantic ReportSpec definitions enforce expected response formats for reporting standards.

llm-as-judge (evaluation)

Bedrock models evaluate generated ESG answers for accuracy, completeness, consistency, framework alignment, and mathematical correctness.

human-in-the-loop-gate (governance)

ESG experts review cited responses, calculation methods, regulatory compliance, and discrepancies before finalizing disclosures.

Tradeoffs

Custom RAG pipeline

Chose Textract, Titan embeddings, Faiss, S3 persistence, and on-demand loading over Amazon Bedrock Knowledge Bases as an end-to-end managed retrieval stack.

Rationale: The custom path preserved control over text extraction, embeddings, and vector database configuration, while taking on more operational work.

On-demand in-memory vector store

Chose Faiss persisted on S3 and loaded when required over always-on vector infrastructure.

Rationale: ESG report generation is intermittent, so on-demand loading better matches spiky quarterly or annual usage.

Agentic batch-fill with expert review

Chose automated draft responses plus human validation and editing over fully manual questionnaire completion or fully autonomous submission.

Rationale: Automation removes undifferentiated data-gathering work, but experts remain responsible for factual accuracy, methodology, and regulatory compliance.

Gotchas (what bit them)

ESG answers need auditable source and calculation trails.

Root cause: Disclosure answers can affect regulatory compliance and business reporting, so unsupported generated text is not acceptable.

Resolution: Report GenAI includes cited sources, calculation methodology, human review, and visible agent reasoning.

LLMs are ill-equipped for analytical ESG database questions without tools.

Root cause: Many disclosure questions require processing numerical or tabular emissions data rather than generating prose from memory.

Resolution: A separate text-to-SQL tool generates, lints, corrects, and executes SQL against the emissions database.

Off-the-shelf retrieval would reduce control over document processing choices.

Root cause: The team wanted to choose extraction, embeddings, and vector-store design for their workload.

Resolution: They built a custom Step Functions/Textract/Titan/Faiss pipeline despite accepting higher management overhead.

Evaluation

Report GenAI uses dual-layer validation: ESG experts review and validate AI-generated responses, while Bedrock LLM judges assess accuracy, completeness, consistency, framework alignment, and mathematical correctness at response and sub-component levels.

Metrics: response accuracy relative to source documentation · completeness against question requirements · consistency with provided context · alignment with reporting framework guidelines · mathematical accuracy of calculations · confidence score and detailed assessment

Outcomes

• Omni Helicopters International cut CDP reporting time by 75%.
• OHI moved from one month of dedicated CDP reporting effort to one week for its 2024 submission.
• The system frees ESG teams from routine data gathering and analysis so they can focus on strategic sustainability initiatives.

The solution targets enterprise ESG reporting workloads with thousands of data points and quarterly/yearly spikes.