Credit analysis for small and medium enterprises is one of the most document-intensive workflows in banking. An analyst might review twelve to twenty financial documents — annual accounts, management accounts, bank statements, tax returns — before producing a credit assessment. At a typical bank processing hundreds of SME applications per month, that volume makes the process a significant operational bottleneck. We built a platform to address it.
The Problem We Set Out to Solve
Most bank credit tools are either legacy systems requiring significant manual data entry, or generic AI tools that were not designed with banking data in mind. The result is fragmented analysis: an analyst extracts figures from documents in one system, builds a financial model in a spreadsheet, and writes a narrative in a word processor. Each step is manual, there is no structured audit trail, and the quality of the output depends heavily on individual skill and time pressure.
The goal we set for this platform was specific: reduce the time from document submission to completed credit report to under thirty minutes, while producing outputs that are consistent in structure, compliant in format, and auditable end-to-end.
Platform Architecture
- Document upload — PDF and Excel files submitted through a secure interface and stored in Azure Blob Storage with access controls and an immutable audit log
- OCR and parsing — documents processed through Azure Document Intelligence for layout-aware extraction, with table parsing for financial statements and key-value detection for summary fields
- Chunking and indexing — extracted content segmented into semantically coherent chunks and indexed in Qdrant (vector database) alongside structured metadata
- Evidence extraction — targeted AI queries retrieve specific financial metrics (revenue, EBITDA, debt ratios, cash position) from the indexed content using vector similarity search
- Modular credit analysis — independent AI modules generate each section of the credit report in parallel
- Structured report output — the assembled report is returned in a defined format with source document references per extracted finding
Why Modular AI Sections — Not One Big Prompt
Early AI-based credit analysis used a single large prompt: provide all the documents, request a credit report. This approach has fundamental problems. Context windows fill up with documents, models lose track of specific figures, and output consistency is unpredictable — thorough in one section, absent in another. The quality degrades with document volume, which is exactly when you need it most.
Our platform runs each report section as an independent module with its own targeted prompt and dedicated evidence retrieval step:
- Business analysis — company profile, market position, ownership structure
- Financial summary — revenue trends, profitability, working capital, debt structure
- Risk assessment — key risks, mitigants, regulatory considerations
- Monitoring insights — recommended covenants and data points to track post-approval
Each module retrieves only the evidence it needs from the vector index, processes a focused prompt, and returns a structured section. Modules run in parallel, meaning report generation time is determined by the slowest module — not the sum of all modules. This is what makes sub-thirty-minute report generation achievable even with large document sets.
Outcomes
- Three times faster credit processing compared to manual analysis workflows
- Consistent report structure across analysts, applications, and time periods
- Audit-ready outputs with source document references per extracted figure
- Multi-document processing without context window limitations
This platform demonstrates how AI can be applied to complex banking workflows without sacrificing the rigour that regulated environments require. It is part of TechZiel’s work in AI-powered decision and credit analysis systems. If you are assessing how AI could accelerate your credit operations while maintaining audit and compliance standards, contact us to discuss your specific context.