Transforming Loan Underwriting with Agentic AI on Google Cloud
An Enterprise Architecture Proposal for German Banks (Part 1)Executive SummaryGerman banks have already digitized customer onboarding, KYC, SCHUFA integration, and rule-based underwriting. The remaining bottleneck lies in interpreting unstructured financial evidence, coordinating multiple…
An Enterprise Architecture Proposal for German Banks (Part 1)Executive SummaryGerman banks have already digitized customer onboarding, KYC, SCHUFA integration, and rule-based underwriting. The remaining bottleneck lies in interpreting unstructured financial evidence, coordinating multiple enterprise systems, and supporting underwriters with consistent, explainable recommendations. This proposal introduces Agentic AI orchestration layer built on Google Cloud Platform (GCP) that complements-not replaces-existing leading platforms. By deploying this architecture, banks can handle unstructured information, orchestrate enterprise services, and prepare explainable recommendations, enabling faster decisions while preserving regulatory compliance and human accountability.This three-part series outlines a target-state enterprise architecture designed specifically for the realities of the German banking sector. In Part 1, we dissect the current state of loan processing, acknowledging the significant automation already in place, and pinpoint the exact manual bottlenecks where Agentic AI delivers transformative value. Part 2 details the multi-agent ecosystems and the protocols (A2A and MCP) that allow these agents to collaborate with existing core banking systems. Part 3 provides a granular view of the GCP infrastructure, network security, and the rigorous governance required to satisfy BaFin, the EU AI Act, and the Digital Operational Resilience Act (DORA).The Reality of Current-State German Loan ProcessingWhen discussing AI in banking, there is common misconception that the current underwriting process is entirely manual — that a customer submits a paper application, a human requests a SCHUFA report, and someone manually calculates the decision. For a large German bank, this narrative is fundamentally incorrect. Today’s Tier-1 institutions already operate sophisticated, highly automated loan origination pipeline. To understand where Agentic AI adds value, we must first map the reality of the existing architecture honestly.The Existing Automated PipelineThe typical digital origination journey in Germany today is heavily orchestrated by Loan Origination System (LOS) and Business Rules Management Systems (BRMS). Most large German banks already have automated SCHUFA calls, automatic KYC, rule engines, pricing engines, automated affordability calculations, and workflow engines. The flow generally looks like this:Figure 1: Current-State Loan Processing Flow — Note the color coding: green indicates fully automated steps (KYC, AML/PEP, SCHUFA, BRMS, Pricing), orange indicates semi-automated steps (document upload, income verification), and red indicates the manual bottleneck (underwriter review for exception cases). Approximately 80% of straightforward applications pass through automatically; the remaining 15–20% require manual intervention.The standard origination journey generally follows these steps:Digital Intake: The borrower submits an application via a proprietary customer portal or through established broker platforms.Digital KYC & Screening: Identity is verified automatically via PostIdent or VideoIdent. The system performs automated Anti-Money Laundering (AML), Politically Exposed Person (PEP), and sanctions screening.Automated Bureau & Internal Checks: The system makes an automated API call to SCHUFA (and sometimes CRIF) to retrieve credit scores and existing liabilities. Simultaneously, it queries the core banking system (e.g., SAP, agree21, Avaloq, Temenos) to access internal customer risk, repayment history, previous defaults, relationship history, and existing exposure limits.Business Rules Engine (BRMS): The data is fed into the BRMS. The rules engine automatically calculates affordability, Debt Service Ratios, and Loan-to-Value (LTV).Pricing Engine: Based on the risk grade, LTV, and customer relationship, an automated pricing engine determines the applicable interest rate, margin, and product eligibility.Decision Point: If all rules pass, the application is auto-approved. If any rules fails or the case is complex, it is routed to a human underwriter.The True Bottleneck: Where Automation StopsIf the current system is so automated, why does mortgage underwriting still take 3 to 7 days? The answer is that the “happy path” describe above only works perfectly for the simplest, most standardized employed applicants with single-source of income. For a significant portion of applications-often 15% to 20%-the deterministic rules engine cannot reach a conclusive decision. When the rule failes, the file is routed to a human underwriter.This is where the automation stops and the cognitive bottlenecks begins. The human underwriting is suddenly faced with a mountain of unstructured data. The manual work is not about calculating afforability-the system already does that. The manual work is about understanding complex, messy, real-world financial evidence. Specifically, the underwriter must interpret:Multiple, inconsistent versions of salary slips (Gehaltsabrechnung)Complex self-employed tax assessments (Steuerbescheid) with variable incomeHandwritten documents or custom employer lettersRental income documentation and foreign income streamsMultiple applicants on a single file with different income typesException policies that require judgement beyond deterministic rulesCollateral validation and property valuation reportsCurrently, the underwriter spends 30 to 90 minutes per application acting as a data extractor and system integrator-reading PDFs, verifying authenticity, logging into 5 to 10 different internal systems, and manually typing extracted figures back into the LOS to force the rules engine to recalculate.Current Operational MetricsThe Real Transformation OpportunityThe true opportunity for Generative AI in German banking is not to replace the existing workflow, but to augment the cognitive gaps within it. The story is not “replace the rules engine with AI”. The story is far more pragmatic:The rule engine stays where it is. When the rules engine hits an exception, AI performs the cognitive work that currently requires 30–90 minutes of human data extraction. The human underwriter still makes the final decision.Today, when a rule fails, the process looks like this; the rules engine flags an exception, a human reads 100 pages of unstructured documents, manually synthesize the data across multiple systems, and eventually makes a decision. In the proposed future state, an Agentic AI layer performs that cognitive heavy lifting: it understands the complex documents, verifies the data against internal systems, calls external APIs, and prepares a synthesized evidence package. The human underwriter still makes the final, legally binding decision — but now they spend 10 to 15 minutes reviewing a pre-prepared, cited recommendations rather than 90 minutes acting as a data entry clerk.Agentic AI Sits Above Exiting Banking SystemsAgentic AI differs from standard Generative AI (like a chatbot) because it is goal-oriented and capable of tool use. In this proposed architecture, Agentic AI sits above existing banking systems as an intelligent orchestration and reasoning layer. It does not replace SAP, agree21, the bank’s established BRMS, or the pricing engine. It does not replace SCHUFA. Instead, it acts as a digital copilot that knows how to interact with all of these systems.When a complex application arrives, the AI orchestration layer:Reads the inconsisten, unstructured documents using Document AI.Queries the SCHUFA API, CRIF, and the core banking system to build a holistic profile.Invokes the existing BRMS to test the newly extracted data against the bank’s deterministic policies.Invokes the existing Pricing Engine to calculate rates based on the validated risk grade.Synthesizes the findings into a clear, cited recommendation memo for the human underwriter.Target State KPIsLooking Ahead to Part 2By correctly positioning Agentic AI as a cognitive orchestration layer rather than a rip-and-replace core system, banks can dramatically improve underwriter productivity and decision speed while preserving their massive investments in existing enterprise platforms — the LOS, the BRMS, the pricing engine, the core banking systems, and the established SCHUFA integration.In Part 2 of this series, we dive into the technical design of this orchestration layer. We will explore the specific multi-agent ecosystem — including the Document Agent, the Collateral Agent, the Rule Engine Integration Agent, and the Pricing Agent — and detail how the A2A and MCP protocols allow these AI agents to communicate securely with legacy banking infrastructure.References[1] SCHUFA Holding AG. “What role does SCHUFA play in the granting of loans.” SCHUFA Newsroom, Jan 2025. https://www.schufa.de/en/newsroom/finances/welche-rolle-spielt-die-schufa-bei-der-kreditvergabe/[2] BaFin. “Regulatory requirements for AI applications.” BaFin Publications. https://www.bankinghub.eu/finance-risk/bafins-regulatory-requirements-ai[3] European Banking Authority (EBA). “Guidelines on loan origination and monitoring.” EBA/GL/2020/06. https://www.eba.europa.eu/regulation-and-policy/credit-risk/guidelines-on-loan-origination-and-monitoringBefore you goPlease take a moment to like the post and follow the writer!Did you know that over 400,000 developers share what they’re building, learning, and discovering across our platforms every month? Learn how you can contribute hereThis story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!Transforming Loan Underwriting with Agentic AI on Google Cloud was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI