Transforming Loan Underwriting with Agentic AI on Google Cloud
An Enterprise Architecture Proposal for German Banks (Part 3)In Part 1 and Part 2 of this proposal, we established how an Agentic AI orchestration layer can augment existing German loan origination pipelines. We detailed the multi-agent ecosystem — Document Agent, Borrower Profile Agent, Collateral…
An Enterprise Architecture Proposal for German Banks (Part 3)In Part 1 and Part 2 of this proposal, we established how an Agentic AI orchestration layer can augment existing German loan origination pipelines. We detailed the multi-agent ecosystem — Document Agent, Borrower Profile Agent, Collateral Agent, Rule Engine Integration Agent, Pricing Agent, Fraud Agent, and Underwriter Copilot — and explained how the A2A and MCP protocols enable these agents to collaborate securely with existing core banking systems, BRMS, and external APIs like SCHUFA.However, the true test of any architecture proposal in the German banking sector is not its conceptural elegance, but its operational resilience and regulatory compliance. Deploying Generative AI at scale requires a cloud infrastructure capable of satisfying the stringent demands of BaFin (MaRisk, BAIT), the European Banking Authority (EBA), the Digital Operational Resilience Act (DORA), and the EU AI Act.In this final part, we detail the target-state Enterprise Architecture on Google Cloud Platform (GCP). We map the specific GCP services to each layer, outline the secure network topology, and explicitly address the governance controls required for production deployment in a Tier-1 German financial institution.Enterprise Architecture on Google Cloud PlatformTo meet the demands of a large German bank, the architecture must be modular, scalable, and secure by design. We propose a highly segmented architecture on GCP that clearly separates concerns: intake, orchestration, AI processing, data management, and security. The following diagram provides the complete, layered view.Figure 5: High-Level Enterprise Architecture — shows all layers from Channel Intake through to the Human-in-the-Loop decision gate. Note the explicit inclusion of Existing Banking Systems (green, on-premises) alongside the new Agentic AI Layer (purple, GCP). A2A protocol governs communication between the Orchestration layer and the Agent layer. MCP protocols connects agents to both GCP services and on-premises systems. The Security & Governance plane (pink) enforces policies across all layers.Granular GCP Service MappingThe architecture relies on the following specific GCP components, mapped to each layer:Detailed Component ArchitectureThe following diagram provide a more granular view of how the individual GCP services connect and interact within the underwriting workflow, showing the specific data flows between the orchestration layer, the agent microservices, the MCP server layer, and both GCP and on-premises systems.Figure 6: Detailed Component Architecture — Shows the complete service interaction map. Cloud Workflows (Orchestrator) delegates A2A Tasks to the seven sepcialist agents running on Cloud Run. Each agent makes MCP Calls to its corresponding MCP Server, which in turn connects to the actual system (SCHUFA API, Core Banking, BRMS, Pricing Engine, Document AI, BigQuery, or Valuation API). The Security & Governance plane enforces policies across the entire stack.GCP Infrastructure and Network DesignTo visualize how these components communicate securely at the infrastructure level, the following diagram details the underlying network topology. This is the level of detail that enterprise architects and platform teams expect when evaluating a cloud deployment for a regulated financial institution.Figure 7: GCP Infrastructure & Network Architecture — The complete network topology showing the Custom VPC in europe-west3 (Frankfurt), the DMZ subnet with Cloud NAT and Cloud Router, the private compute subnet hosting Cloud Run and GKE, the private data subnet protected by VPC Service Controls, Private Service Connect for internal API access, Cloud Armor WAF at the edge, Apigee API Gateway, Cloud Interconnect to on-premises core banking, and secure mTLS outbound to SCHUFA via statis NAT IP.Key Network Design DecisionsVirtual Private Cloud (VPC) and Data Residency: The entire solution resides within a custom VPC strictly constrained to the europe-west3 (Frankfurt) region. This ensures data residency and compliance with German data protection laws. No customer data leaves the Frankfurt region.Private Subnets with No Internet Access: All compute resource (Cloud Run, GKE) and data stores (AlloyDB, BigQuery) are deployed in private subnets with no direct inbound internet access. This eliminates an enture class of attack vectors.Private Service Connect (PSC): Communication between the Cloud Run agent microservices and Google managed APIs (Vertex AI, BigQuery, Document AI) occurs entirely over the private Google network using PSC. Sensitive financial data never traverse the public internet, even when accessing GCP services.Secure Outbound Integration to SCHUFA: When the Borrower Profile Agent needs to query SCHUFA via its MCP Server, the request is routed through a Cloud NAT to provide a static, allowlisted outbound IP address. The connection is secured via mutal TLS (mTLS) managed by Apigee. SCHUFA’s firewall allowlist only the bank’s static NAT IP.On-Premises Connectivity: Integration with the bank’s legacy core systems (SAP, agree21, existing BRMS, Pricing Engine, Risk Engine, LOS, ECM) is established via Cloud Interconnect (dedicated or partner), terminating at a Cloud Router to enable secure, private BGP peering between GCP and the bank’s data centers.Secret Manager and Cloud KMS: All API keys (including SCHUFA credentials), certificates, and encryption keys are managed centrally via Secret Manager and Cloud KMS, with automatic rotation policies and audit logging.Chronical SIEM Integration: All security events, access logs, and decision audit trails are streamed to Google Chronicle for real-time security analytics and incident response, integrating with the bank’s existing Security Operations Center (SOC).Regulatory Compliance and GovernanceIn the German banking context, architecture is inseparable from compliance. This solution is designed specifically to address the regulatory frameworks governing automated lending and cloud deployments.BaFin: MaRisk and BAITMaRisk AT 4.3.1 (Credit Decisions): The architecture ensures clear competencies. The AI does not make the final credit decision; it prepares an evidence-backed recommendation. The human underwriter retains the final, legally binding authority. The deterministic BRMS remains the authoritative source of credit policy — the AI merely invokes it and interprets its output.BAIT (Bankaufsichtliche Anforderungen an die IT): The deployment adheres to BAIT requirements for cloud computing. Secret Manager and Cloud KMS handle all cryptographic keys centrally. All access is governed by Cloud IAM with Workload Identity Federation. The VPC Service Controls perimeter prevents data exfiltration even if an internal credential is compromised. Cloud Interconnect ensures that on-premises integration does not expose internal systems to the public internet.EBA Guidelines on Loan OriginationThe European Banking Authority required institutions to have robust, documented processes for creditworthiness assessments [1]. The Agentic AI layer supports this by enforcing consistency. Because the AI orchestrates data gathering and invokes the deterministic BRMS for every application, the bank can prove that every applicant’s data was subjected to the exact same rigorous evaluation process, reducing operational variance and satisfying the EBA’s expectations for standardized assessment.Digital Operational Resilience Act (DORA)DORA mandates strict management of ICT third-party risk and operational resilience [2]. This architecture addresses DORA through several mechanisms:Resilience and Availability: The architecture utilizes highly available, multi-zonal managed services (Cloud Run, AlloyDB). If a specific AI agent container fails, Cloud Workflows automatically retries or routes the application to a manual exception queue. There is no single point of failure.Third-Party Risk Management: Google Cloud provides contractual commitments for data residency, processing location, and incidents notification. The architecture minimizes external dependencies — SCHUFA and Valuation APIs are the only external calls, both secured via mTLS through Cloud NAT with static IPs.Monitoring and Incident Response: Cloud Logging and Cloud Monitoring capture all system metrics, feeding directly into Chronicle SIEM and Security Command Center for real-time incident detection and response.EU AI ActUnder the EU AI Act, credit scoring and lending decisions are classified as High-Risk AI systems [3]. This classification imposes specific obligations:GDPR CompliancePurpose Limitation: Data collected for underwriting is strictly segregated and is not used to train foundational LLMs. Vertex AI explicitly gurantees that customer data is not used to train Google’s based models.Data Minimization: Sensitive Data Protection (DLP) automatically redacts unnecessary PII (e.g., religion, health data potentially visible on bank statements) before the data is processed by the reasoning engines.Right to Explanation (Article 22): Because the Copilot Agent produces cited recommendations, the bank can proved any applicant with a clear explanation of why a decision was made, including which specific data points influenced the recommendation.Business Value and Measurable OutcomesImplementing this Agentic AI architecture transforms the underwriting department from a data-entry operation into a high-leverage risk analysis center. The bank can expect measurable improvements across the following Key Performance Indicators:Conclusion: The Path ForwardThe German banking sector is at an inflection point. The technology now exists to resolve the manual bottlenecks of loan origination without sacrificing the rigorous risk controls that define the industry. Banks do not need to rip out their existing BRMS, pricing engines, or core banking systems. They need an intelligent layer that can read unstructured documents, synthesize data from multiple sources, and prepare evidence-backed recommendations for human decision-makers.By deploying an Agentic AI Orchestration layer on Google Cloud Platform, banks can leverage their existing investments in digital channels, SCHUFA integration, and deterministic rule engines. The A2A protocol ensures agents collaborate in a governed, auditable manner. The MCP protocol ensures agents access legacy systems through standardized, secure adapters rather than brittle custom code. And the GCP infrastructure — with its VPC Service Controls, Private Service Connect, and data residency — provides the scalability, security, and explainability required to satisfy BaFin, the EBA, DORA, and the EU AI Act.This is not a theoritical exercise. It is an evolution of existing capabilities — a pragmatic, enterprise-ready path to faster, more transparent, and fully compliant lending for the German consumer.References[1] 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-monitoring[2] European Parliament. “Digital Operational Resilience Act (DORA).” Regulation (EU) 2022/2554. https://eur-lex.europa.eu/eli/reg/2022/2554/oj[3] European Parliament. “Artificial Intelligence Act.” https://artificialintelligenceact.eu/[4] BaFin. “MaRisk — Minimum Requirements for Risk Management.” Oct 2017. https://www.bafin.de/SharedDocs/Downloads/EN/Rundschreiben/dl_rs0917_marisk_Endfassung_2017_pdf_ba_en.pdf[5] BankingHub. “BaFin’s regulatory requirements for AI applications.” June 2022. https://www.bankinghub.eu/finance-risk/bafins-regulatory-requirements-ai[6] Google Developers. “Announcing the Agent2Agent Protocol (A2A).” Google Developers Blog, April 2025. https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/[7] Anthropic. “Introducing the Model Context Protocol.” Anthropic News, Nov 2024. https://www.anthropic.com/news/model-context-protocolThis 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