Credit Stream AI

Unlike in retail banking, reviewing a loan application of medium and large corporates requires a credit analyst to review a large amount of credit documentation to prepare the conclusion on the request. This may include financial statements, business plans, charter documents, collateral documents, etc... The volumes may be up to 1000 pages and therefore this requires extensive manual efforts. Nowadays it has low level of automation and the review process can take weeks. Usage of GenAI agents to automate this process not only significantly reduces the costs but also speeds up the decision-making process, providing a competitive advantage in the market. CrediStreamAI solution that analyzes large volumes of corporate documents up to 1000 pages of documents and automatically generates pre-populated reports for credit assessments, legal due diligence, and collateral assessments. The system significantly reduces the manual workload for analysts (by ~ 4 man/days)

2026

The solution is scalable across all ~600 branches in Kazakhstan and is used for corporate lending. The project covers several critical bank functions: credit, legal, and collateral analysis. Since just within first 3 months after project launch- 780 conclusions and reports have been prepared by AI agents for client applications each of the limits with > 1 bln tenge. For every conclusion ~4 man/days of credit analyst’s efforts were saved, i.e. total cost savings is about 3100 man/days or 12 FTEs equivalent of credit analysts.

In-house built solution, integrated with Banks loan origination systems. Documents are read not through classic OCR, but by multimodal models: they perceive the document as a whole, like an analyst — table structure is preserved, and the system is resilient to stamps, seals, and complex layoutsQuen LLM is intsalled on-premise.

The project does not use solutions from the Global CIO catalog

The implementation is built around agent orchestration and AI workflows: extraction, validation, discrepancy resolution, and opinion generation stages are executed as a managed graph of steps, where the output of each stage is verified by the next — and only then is the pre-filled opinion generated. What distinguishes the solution from conventional GenAI assistants is its built-in internal verification of results. The verification embeds banking specifics: mathematical validation of financial statements, comparison against industry benchmarks, and detection of contradictions in charter documents.

Credit Stream AI is Halyk Bank's proprietary AI platform for automating the preparation of expert opinions in lending to mid-sized and corporate clients. Unlike retail, where a loan decision is made within minutes, a credit expert manually reviews financial statements, business plans, charter and collateral documents — up to 1,000 pages per application, which takes weeks of man-hours.

The system is built as a multi-agent pipeline that digitizes large volumes of documents and generates pre-filled opinions for four expert functions: credit analysis, legal review, collateral review, and risk assessment. Documents are read not through classic OCR, but by multimodal models: they perceive the document as a whole, like an analyst — table structure is preserved, and the system is resilient to stamps, seals, and complex layouts. A single workflow — document classification, locating the required forms, field extraction, and mapping them to strict structured schemas — processes financial statements, breakdowns, business plans, charter and collateral documents.

The implementation is built around agent orchestration and AI workflows: extraction, validation, discrepancy resolution, and opinion generation stages are executed as a managed graph of steps, where the output of each stage is verified by the next — and only then is the pre-filled opinion generated. What distinguishes the solution from conventional GenAI assistants is its built-in internal verification of results. The verification embeds banking specifics: mathematical validation of financial statements, comparison against industry benchmarks, and detection of contradictions in charter documents.

All models are trained and deployed on the bank's own GPU capacity, which provides a practically absolute level of cybersecurity: client data never leaves the bank's perimeter. This is a fundamental requirement for Halyk when working with sensitive data. Security is reinforced by multi-level filtering of model requests and responses, and strict output data schemas; GPU infrastructure and SLA are under continuous monitoring.

At the same time, AI does not make the final call: the expert receives a transparent result with flags for inconsistencies and questions for clarification, reviews it, and makes the decision personally — the result is fully traceable and verifiable.

Kazkhstan