The company is a palm oil plantation group that manages water-related conditions across multiple work units to support plantation operations. Its Water Management team monitors various environmental parameters, including rainfall, water levels, groundwater conditions, land subsidence, and water-control structures. These activities are supported by weekly operational reports and standard operating procedures (SOPs).
However, the growing volume and variety of operational data and documentation made it increasingly challenging to consolidate information, access relevant insights, and retrieve procedures efficiently. The company therefore sought a solution to streamline access to its data and operational knowledge, enabling faster and more informed decision-making across the organization.
Challenges: Manual Reporting and Fragmented Access to Operational Knowledge
Before the project, the company did not yet have a generative AI initiative for its Water Management function. Weekly reports arrived from different work units in different file formats, and the team had to compile the information manually. Management spent considerable time reading and summarizing reports, while field operators often found it difficult to locate the relevant SOP for a particular situation. Analysis across work units also required manual effort.
The underlying data introduced additional complexity. During the initial assessment, Central Data Technology (CDT) identified inconsistent date and time-zone formats, quoted fields, decimal commas, and different schemas across similar water-level datasets. Without standardization and validation, these inconsistencies could affect the reliability of automated analysis. The company also needed to ensure that users only accessed information relevant to their assigned work unit and role.
Solutions: An AI-Powered Water Management Assistant on AWS
To address these challenges, CDT designed and implemented the Water Management Chatbot using Amazon Web Services (AWS), a generative AI assistant integrated into the company’s existing mobile, web, and portal applications. Users can ask questions in Bahasa Indonesia or English and receive answers based on validated operational data or relevant passages from the company’s documents.
The solution uses Amazon Bedrock as its generative AI foundation. For numerical questions, such as water-level trends, rainfall comparisons, or the number of water-control structures, the AI agent queries curated data through Amazon Athena. For procedural questions, it retrieves relevant content from an Amazon Bedrock Knowledge Base, supported by Amazon S3 Vectors and Cohere Embed v4 for multilingual document embeddings. Answers include their sources, helping users verify the information before taking action.
The data pipeline is built on Amazon S3 and AWS Glue. Data exported daily from the company’s SQL Server is organized into several layers. AWS Glue ETL jobs clean, transform, and aggregate the data, while Great Expectations validates data quality at each layer. The AWS Glue Data Catalog maintains table metadata for Amazon Athena.
The application layer runs a Strands Agents-based AI agent through FastAPI on Amazon EC2. OpenAI GPT-6 Luna on Amazon Bedrock acts as the primary model, while Amazon Nova Lite supports lighter tasks and serves as a fallback. Amazon Bedrock Guardrails helps keep responses within the Water Management scope, and Amazon Bedrock Prompt Management supports controlled prompt versioning. Langfuse records agent traces and feedback, while Amazon CloudWatch, Amazon EventBridge, and Amazon SNS support infrastructure and pipeline monitoring and alerting.
Security and access control are built into the architecture. The solution connects the company’s AWS accounts through a site-to-site VPN, keeps the backend in a private subnet, uses VPC endpoints and AWS KMS encryption, and applies least-privilege IAM. Requests carry the signed-in user’s identity, work unit, and role; access filters are applied to both data queries and document retrieval. The chatbot is advisory only: operational decisions, including water-gate operations and flood mitigation, remain with authorized personnel under the applicable SOPs.
Implementation: From Proof of Concept to Production
CDT delivered the project through its four-phase methodology (Assess, Design, Build, and Operate), working closely with the company’s Water Management and System Development teams.
The journey began with an Amazon Quick proof of concept (PoC) for management, designed to validate the platform’s capabilities in delivering business insights through natural-language interactions. The pilot explored various analytical features, including dashboards, automated workflows, and AI-powered chat responses grounded in company data. The PoC also helped identify data-quality gaps and gather key insights that informed the production design and implementation approach.
CDT then designed and built the custom Amazon Bedrock solution for integration into the company’s applications. The production data pipeline exports SQL Server data to Amazon S3 every day, processes it through AWS Glue, and validates each layer before the data is queried by Amazon Athena. In parallel, administrators upload SOPs and reports through a document uploader; metadata is applied and the files are synchronized with the Amazon Bedrock Knowledge Base.
The AI agent was integrated with the company’s applications through a private API. CDT and the company tested the solution with a set of reference questions covering data and document questions, complex questions, unavailable data, out-of-scope requests, and work-unit access boundaries. Functional integration tests were also completed.
To support adoption and long-term ownership, CDT delivered a two-day Amazon Quick knowledge-sharing session, a one-day Amazon Bedrock kickstart, and a two-day hands-on workshop using real Water Management data. CDT also provided sample code and a runbook so the company’s System Development team could understand, run, and modify the solution. After launch, the solution entered a one-month stabilization period with daily checks and weekly answer-quality reviews. It now runs in production on AWS in the Asia Pacific (Singapore) Region, inside the company’s applications.
Results: Faster Access to Information and a Foundation for Data-Driven Operations
The project has moved the company from manual compilation and fragmented document searches toward a centralized, conversational way to access Water Management information. Users can ask questions directly in the company’s applications, obtain numerical answers from curated datasets, and retrieve procedures from the knowledge base without manually searching across multiple reports and files.
The solution brings together around 15 monitoring datasets and a document base that started at around 200 documents and is designed to grow to approximately 800. The initial proof of concept demonstrated adoption with 10 users, and the company plans to expand the assistant to hundreds and later thousands of users through its applications.
Quality and efficiency are supported by measurable operating targets. The project set targets for response speed and for reducing the time needed to read and summarize weekly reports by at least 30%. The solution passed user acceptance testing, and answer quality, response time, user feedback, and safety interventions are monitored through Langfuse and Amazon CloudWatch.
Measured by the company’s internal team together with CDT on different sets of test questions, the time needed to read and summarize weekly reports went from days to about 15 minutes, well beyond the 30% target. This saves at least one working day per weekly report, about four working days per month. Answers are streamed, and the first part of the answer appears in under 3 seconds. The solution gives the company a scalable foundation for more consistent information access and better-supported Water Management decisions.
CDT and AWS: Enabling Practical Generative AI for Business Operations
Through the Water Management Chatbot, CDT combines AWS data services, generative AI, document retrieval, and security controls in a solution embedded in the customer’s existing applications. The result is a practical AI assistant that helps teams work with operational data and knowledge more efficiently, while keeping access governed and critical decisions in the hands of authorized personnel.
To explore how a generative AI assistant could work with your operational data, contact Central Data Technology here.
