AI-Powered Enterprise Knowledge Layer

The Challenge

A large financial organisation operated across multiple countries, systems and regulatory environments. Over time, its knowledge landscape became fragmented:

• thousands of pages of internal documentation (HTML-based internal portals),
  
• process and policy descriptions in PDFs,
  
• Excel-based data dictionaries owned by different teams,
  
• enterprise databases and data warehouses with inconsistent schemas and naming.
  

Understanding how products, systems, data fields and processes were connected required manual document reading and expert knowledge. Impact analysis for changes (regulatory updates, system modifications, new products) was slow, risky and heavily dependent on a few key people.

The organisation needed a way to turn scattered documentation and data into a structured, AI-usable knowledge model.

Our Solution

We designed and implemented an AI-powered knowledge and modelling layer that automatically builds and maintains a machine-readable representation of the organisation’s business and data landscape.

The solution included:

1. Large-scale ingestion and preprocessing

We built data pipelines (Databricks, Apache Spark, Airflow) to ingest:

• internal HTML documentation,
  
• PDFs,
  
• Excel files,
  
• enterprise database metadata.
  

Content was cleaned, normalised and split into semantically meaningful chunks, forming the foundation of an AI-ready corpus.

2. Vectorisation and RAG foundation

Chunks were embedded and stored in vector databases, enabling retrieval-augmented generation (RAG). This allowed language models to answer questions using the institution’s own knowledge instead of generic training data.

3. Specialised AI agents for knowledge modelling

We implemented a set of AI agents that:

• extracted business entities (products, customers, systems, fields, processes),
  
• classified them into business domains (lending, payments, reporting, etc.),
  
• identified relationships (which product uses which system, which field belongs to which table, which process triggers which event).
  

The outputs were aggregated into a structured knowledge model / knowledge graph.

The Result

The institution gained:

• a living, machine-readable knowledge model of its business and data landscape,
  
• the ability to ask questions in natural language about internal systems and data,
  
• significantly faster impact analysis when systems, products or regulations changed,
  
• reduced dependency on manual document parsing and individual expert knowledge.
  

This knowledge layer became a foundation for further AI and analytics initiatives.