VakıfBank AI Financial Analysis Assistant

Founded in 1954 with the aim of utilizing foundation resources in the most effective way in line with the needs of economic development, VakıfBank has since contributed to the growth of Turkey's savings culture through modern banking practices and applications.

VakıfBank AI-Powered Financial Analysis Assistant: Redefining Corporate Lending with Hybrid GenAI

VakıfBank has pioneered a transformative shift in corporate lending by deploying an innovative, completely offline AI-powered financial analysis assistant. Integrated into the bank's local infrastructure, the system automates the evaluation of financial tables, sectoral data, and risk metrics. By transitioning from a manual, experience-dependent process to a hybrid decision support model, VakıfBank has achieved a radical leap in operational efficiency, reducing analysis time from 1 hour to just 9.25 seconds.

Before this transformation, the quality of commercial credit assessments relied heavily on the experience levels and time constraints of branch portfolio managers. Evaluating complex financial statements and ratios was a specialized, labor-intensive task prone to inconsistency, especially during peak periods. There was a critical need for standardization, speed, and objective consistency in credit decision-making.

The project stands out through its Hybrid AI Architecture and Privacy-by-Design approach:

• Local & Secure Infrastructure: To ensure data security and regulatory compliance, advanced LLMs (gpt-oss-120b and Llama 4 scout) were integrated to work entirely offline.
• Multi-Layered Processing: The system utilizes a sophisticated pipeline including PyMuPDF for text parsing and a validation layer to ensure document integrity before analysis.
• Prompt Orchestration: A custom multi-layered prompt architecture restricts the AI to a specialized financial role, ensuring standardized terminology and explainable outputs.
• Hybrid Accuracy: Recognizing the limitations of LLMs in complex math, the system employs a hybrid approach, using rule-based calculation services for liquidity and financial ratios to guarantee 100% mathematical accuracy.

Since its deployment in November 2025, the project has delivered extraordinary results:

• Velocity: The time required for preliminary financial review and ratio interpretation plummeted from an average of 60 minutes to 9.25 seconds.
• Operational Efficiency: The system has been actively used in 36,109 scoring processes by over 3,000 users.
• Labor Savings: Annually, the bank produces approximately 238,231 scores. While the manual process would require 238,231 man-hours (approx. 119 full-time experienced staff), this AI-driven automation has drastically lowered operational costs.
• Risk Mitigation: By standardizing analysis, the bank has enhanced its ability to detect early risk signals, leading to higher consistency in credit decisions and a potential reduction in Non-Performing Loans (NPL).
• vLLM-Powered Infrastructure: Using vLLM for high-performance inference, the system deploys gpt-oss-120b and Llama4-Scout models. This enables advanced batching and optimized KV-cache management, allowing concurrent analyses without latency
• Methodological Rigor (RAG): Using the latest embedding models, the system retrieves relevant sections from an internal guidebook. This grounds the AI's reasoning in established analytical standards, reducing hallucination risks.
• Hybrid Intelligence: To ensure 100% mathematical accuracy, complex calculations are handled by a rule-based service, providing the LLM with verified numerical data.
• Conversational Continuity: The assistant maintains contextual memory, allowing users to ask follow-up questions while preserving the integrity of the ongoing analysis

In the next phase, VakıfBank will conduct a fine-tuning process using 15,000 analyst-generated document-analysis pairs collected over the past year [Source Ek]. This will enable the LLM to internalize the organization's unique analytical style, prioritization logic, and reasoning patterns, ensuring that AI outputs converge perfectly with historical expert judgments.