Case Study: Automating Document Processing for a Riyadh Bank

When a major Riyadh-headquartered retail bank approached DEEP.SA in early 2024, their loan processing operation was at a breaking point. Growing application volumes, a complex bilingual document environment, and tightening SAMA compliance requirements had stretched a paper-dependent workflow to its limits. Twelve months later, AI automation had reduced average loan processing time from 11 days to under 3 days — a 74% reduction that transformed both customer experience and operational cost.

The Challenge

The bank's retail lending division processed approximately 3,800 loan applications monthly across home finance, personal finance, and SME products. Each application required extraction and validation of data from 15–40 documents including national ID cards, salary certificates, bank statements (Arabic and bilingual), employment letters, property valuation reports, and SIMAH credit bureau extracts.

The existing workflow relied on a team of 45 document processing staff who manually keyed extracted data into the core banking system. Error rates on manual keying ran at approximately 12%, requiring rework cycles that extended processing timelines and created compliance exposure when errors propagated to credit decisions. Staff turnover in the document processing team was running at 40% annually — a direct consequence of repetitive, high-pressure work that offered little career development.

The bank's compliance team had also flagged a structural risk: the manual workflow had no systematic audit trail correlating input documents to the specific data fields they contributed to credit decisions. In the event of a regulatory review or credit dispute, reconstructing the decision chain was an exercise that could take weeks of manual investigation.

The DEEP.SA Solution

Phase 1: Document Intelligence Pipeline

DEEP.SA deployed a document intelligence pipeline using DEEP NLP's Arabic OCR and document understanding capabilities. The system was trained on a corpus of 180,000 anonymized historical banking documents, with specific models developed for each document type: national IDs (Iqama and Saudi national ID in both standard and photographed formats), payslips from major Saudi employers including government agencies, and bank statements from all 30 Saudi licensed banks and major regional banks.

The extraction accuracy on training document types reached 98.3% for structured fields and 94.7% for semi-structured content. Critically, the system was designed to explicitly flag low-confidence extractions for human review rather than silently passing uncertain data downstream — a design principle that proved essential for maintaining compliance with SAMA's credit decision audit requirements.

Phase 2: Workflow Automation with DEEP Automate

Document extraction outputs were integrated with DEEP Automate to orchestrate the downstream processing workflow. Validated data flowed automatically to the bank's core banking system (Temenos T24) via a pre-built connector developed by DEEP.SA's integration team. The automation layer handled application completeness checks, document authenticity verification against ABSHER APIs for ID validation, and SIMAH credit bureau query triggering.

Incomplete applications triggered automated communications to applicants via SMS and email in Arabic, reducing the staff time previously spent on manual outreach. The time between application submission and first completeness review dropped from an average of 2.3 days to under 4 hours.

Phase 3: Compliance and Audit Architecture

Every extraction event, validation check, data write, and decision trigger was logged to an immutable audit trail in DEEP Secure's compliance logging module. For each credit decision, the system generates a complete provenance record linking the credit file to every source document and every data field extracted from it. This audit trail satisfies both PDPL data lineage requirements and SAMA's credit decision documentation standards.

Implementation Challenges and How They Were Resolved

The most significant technical challenge was the variation in document quality from photographed originals. Applicants submitting via the bank's mobile app frequently provided images with poor lighting, perspective distortion, or partial occlusion. DEEP NLP's preprocessing pipeline was enhanced with a document quality assessment module that identifies low-quality inputs before OCR processing and prompts resubmission through the mobile app — reducing failed extraction attempts by 67%.

The change management challenge was equally significant. Document processing staff faced understandable uncertainty about their roles in an automated workflow. The bank partnered with DEEP.SA to design a reskilling program that transitioned 28 of the 45 affected staff to quality assurance, exception handling, and customer service roles with enhanced responsibilities. The remaining 17 positions were absorbed through natural attrition over the 12-month implementation period, with no involuntary redundancies.

Results and Forward Roadmap

The operational metrics confirmed the business case. Average loan processing time fell from 11 days to 2.8 days. Manual data entry errors were reduced by 89%. Staff dedicated to document processing fell from 45 to 12 — all in higher-skill exception handling and oversight roles. Estimated annual cost savings of SR 4.2 million were achieved in the first year. Customer satisfaction scores for the loan application journey improved by 22 points on the bank's NPS measure.

The bank is now extending the document intelligence pipeline to trade finance, account opening, and mortgage origination workflows, with expected go-live for the expanded scope in Q2 2025. A real-time fraud detection layer using DEEP Analytics is also in scoping — using behavioral patterns in the document submission process to flag potentially fraudulent applications before they reach credit assessment.