Transforming Loan Underwriting with Agentic AI on Google Cloud
An Enterprise Architecture Proposal for German Banks (Part 2)In Part 1 of this series, we established that German banks already possess highly automated loan origination pipelines. The true bottlenecks lies in the “exception” cases — complex applications requiring human underwriters to manually…
An Enterprise Architecture Proposal for German Banks (Part 2)In Part 1 of this series, we established that German banks already possess highly automated loan origination pipelines. The true bottlenecks lies in the “exception” cases — complex applications requiring human underwriters to manually interpret unstructured documents and synthesize data across multiple legacy systems. We introduced Agentic AI is not as a replacement for core banking platforms or SCHUFA, but as an intelligent orchestration layer designed to perform this cognitive heavy lifting while the existing deterministic systems remain automated.In Part 2, we explore the architecture of this orchestration layer in detail. We will describe the multi-agent ecosystem required to handle the complexities of German mortgage lending, and explain how these AI agents securely communicate with existing enterprise systems using the A2A and MCP protocols.The Multi-Agent Ecosystem: Bridging the Cognitive GapTo satisfy BaFin’s requirements for traceability and to ensure the system remains modular, we do not rely on a single, monolithic Large Language Model (LLM) to underwrite a loan. Instead, we deploy a tightly governed ecosystem of specialized agents. Each agent has a distinct domain, specific permissions, and access to specific tools.Crucially, these agents do not replace the bank’s existing logic. They act as intelligent bridge between unstructured data and the bank’s deterministic systems. The following diagram illustrates how the future-state process works, clearly showing which steps remain unchanged (existing systems in green) and where the new Agentic AI layer augments the workflow (in purple).Figure 2: Future-State Loan Processing Flow — Green boxes represent existing banking systems that remain completely unchanged (KYC, SCHUFA, BRMS, Pricing Engine). Purple boxes represent the new Agentic AI layer that augments the exception handling path. The human underwriter (yellow) retains final decision authority. Compare this with Figure 1 in Part 1 to see exactly what changes and what stays the same.The Specialist Agents1. Document Understanding AgentThis agent tackles the unstructured data problem head-on. Utilizing Google Cloud Document AI, it ingests complex, varied documentation: inconsistent Gehaltsabrechnung (payslips) from different employers, handwritten employer letters, foreign tax documents, and pension income statements. It classifies the documents, extracts the relevant financial figures, and normalizes the data into a structured JSON format that the rest of the bank’s systems can consume. It immediately flags blurry scans, missing parts, or incomplete uploads — reducing the back-and-forth rework cycles that currently plague the process.2. Borrower Profile AgentThis agent synthesizes the applicant’s complete financial identity. It does not calculate a new credit score; rather, it aggregates existing data from multiple sources. It queries the SCHUFA API, pulls internal repayment history and relationship data from the core banking system (e.g., SAP or agree21), retrieves existing exposure limits, and combines this with the income data extracted by the Document Agent. It builds a holistic, 360-degree view of the borrower’s financial health, reconciling any discrepancies between stated income and documented income.3. Collateral Agent (Mortgage Specific)For real estate lending, collateral is paramount. This agent is responsible for property valuation and land registry (Grundbuch) validation. It integrates with external valuation providers like SkenData or Spengnetter, extracts key metrics from validation reports, validates the property against the land register, and computes the preliminary Loan-to-Value (LTV) ratio based on the extracted data.4. Rule Engine Integration AgentThis is a critical architectural component that reflects the core philosophy of this proposal. The AI does not hold credit policy logic. Instead, this agent takes unstructured data prepared by the Borrower Profile and Collateral agents and invokes the bank’s existing Business Rules Management System (BRMS). The AI reads the output of the BRMS. If the BRMS flags a Debt Service Ration failure, the agent notes the specific failure reason, identifies which input data caused the failure, and prepares this context for the underwriter. The deterministic credit policy remains untouched and fully governed by the risk department.5. Pricing AgentSimilar to the Rule Engine Agent, the Pricing Agent does not invent interest rates. It takes the finalized LTV, the BRMS risk grade, and the customer relationship history, and passes them to the bank’s existing Pricing Engine. It retrieves the calculated interest rate, margin, and product eligibility, appending this information to the case file. The bank’s pricing logic remains entirely within its existing, audited pricing system.6. Fraud & Integrity AgentOperating continuously in the background, this agent cross-references data points to detect anomalies. It compares the employer listed on the application with the employer on the extracted salary slip. It checks for signs of document tampering. It utilizes BigQuery ML and Vertex AI Vector Search to compare the current application against historical fraud patterns stored in the bank’s data warehouse.7. Underwrite Copilot AgentThis is the synthesis layer and the primary interface for the human underwriter. The Copilot Agent compiles the outputs from all previous agents and the BRMS into a comprehensive, nautral-language underwriting memo. It highlights the SCHUFA score, summarizes verified income from multiple sources, lists the specific rules the BRMS flagged, notes the collateral valuation, and provides a clear recommendation. Crucially, every claim in the memo includes a citation linking back to the source document or the specific system API response. If the Copilot states “Applicant’s net income verified at €4,500/month,” it provides a direct reference to the specific line item in the extracted Gehaltsabrechnung.Agent Communication: A2A and MCP ProtocolsA multi-agent ecosystem requires standardized, secure communication. How does the Borrower Profile Agent query SCHUFA? How does the Copilot know the Document Agent has finished extracting the payslip? How do agents invole the bank’s existing BRMS without writing brittle, custom integration code?We utilize two complementary, open protocols to manage these interactions: the Agent-to-Agent (A2A) protocol and the Model Context Protocol (MCP).Figure 3: A2A and MCP Protocol Communication — The left side (blue) shows the A2A Protocol Layer where Cloud Workflows (Orchestrator) delegates tasks to specialist agents. The center (green) shows the MCP Server Layer acting as secure adapters. The right side (red) shows the existing systems and external APIs that agents access through MCP. Note how agents never directly touch legacy systems — they always go through a standardized MCP server.Agent-to-Agent (A2A) Protocol for Task DelegationThe A2A protocol, announced by Google in April 2025, governs how agents delegate tasks to one another [1]. It utilized standardized Task and Artifact payloads, along with Agent Cards that describe each agent’s capabilities.In our architecture, Cloud Workflows acts as the central deterministic orchestrator. It assigns a Task via A2A to the Document Agent: “Extract income from this uploaded PDF.” The Document Agent completes the work and returns A2A Artifact (the structured data). Workflows then assigns a new Task to the Rule Engine Integration Agent, passing the Artifact along. This ensures the workflow remains deterministic and auditable, even while the individual cognitive steps are handled by AI.Model Context Protocol (MCP) for Tool and System AccessThe Model Context Protocol (MCP), created by Anthropic, standardizes how AI agents interact with external data sources and legacy systems [2]. Instead of writing custom, brittle integration code within the LLM prompt for every system, the bank deploys MCP Servers. An MCP Server acts as a secure adapter for a specific system:When the Borrower Profile Agent needs a credit score, it connects to the SCHUFA MCP Server. The server executes the secure API call, handles the authentication and mTLS, retrieves the JSON response, and translates it into a standardized context window for the agent. The AI agent synthesizes the data without ever directly touching the underlying API credentials or legacy system code.The End-to-End Workflow: A2A and MCP in ActionThe following sequence diagram illustrates how these agents, protocols, and existing systems work together to process a complex loan application from submission to final human decision.Figure 4: End-to-End Sequence Diagram — Shows the complete flow from application submission through Apigee Gateway to Cloud Workflows (Orchestrator). Each colored band respresents an A2A task delegation. Within each band, you can see the MCP calls to external systems (SCHUFA, Core Banking, BRMS, Pricing Engine, Document AI, BigQuery, Valuation). The final step is always the human underwriter making the decision based on the Copilot’s synthesized memo.How A2A and MCP Work TogetherTo summarize the relationship between these two protocols clearly:Human-in-the-Loop: The Non-Negotiable SafeguardUnder the GDPR (Article 22) and the EU AI Act, automated decisions that significantly affect consumers require meaningful human intervention [3]. Furthermore, BaFin’s MaRisk AT 4.3.1 requires clear competencies in the credit decision process [4]. The EBA Guidelines on Loan Origination further mandate robust, documented creditworthiness assessment processes [5].Therefore, the Agentic AI system is designed to route all high-risk, borderline, or regulatory-sensitive cases to a human underwriter. The underwriter is presented with the Copilot’s synthesized memo, the extracted data, the BRMS output, and the source documents. The human makes the final, legally binding decision, utilizing the AI purely for accelerated insight and data preparation.The AI never approves a loan. The AI never rejects a loan. The AI prepares the evidence and recommends. The human decides.Looking Ahead to Part 3We have established that Agentic AI can serve as a powerful cognitive orchestration layer, seamlessly integrating with existing BRMS, pricing systems, core banking systems, and external bureaus like SCHUFA via standardized protocols.However, designing the agent ecosystem is only half the challenge. Deploying this system in a German enteprise environment requires an infrastructure that can withstand the scrunity of BaFin and European regulators. In Part 3, we will detail the Google Cloud Platform (GCP) enterprise architecture. We will map the specific GCP services to each architectural layer, outline the secure network topology (VPC, Private Service Connect, Cloud NAT for SCHUFA), and explicitly address compliance with MaRisk, BAIT, DORA, and the EU AI Act.References[1] Google Developers. “Announcing the Agent2Agent Protocol (A2A).” Google Developers Blog, April 2025. https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/[2] Anthropic. “Introducing the Model Context Protocol.” Anthropic News, Nov 2024. https://www.anthropic.com/news/model-context-protocol[3] Hogan Lovells. “Automated decisions by financial institutions under the GDPR and the AI Act.” Feb 2024. https://www.hoganlovells.com/en/publications/automated-decisions-by-financial-institutions-under-the-gdpr-and-the-ai-act[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] 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