Processing a real estate loan has traditionally been one of the slowest, most costly, and error-prone tasks for lending institutions. In 2026, the integration of AI into mortgage processing is radically transforming this landscape, enabling the analysis of dozens of complex documents in mere seconds. Payslips, tax returns, bank statements, and property valuations are no longer bottlenecks, but structured data ready for automated processing.

In this technical tutorial, led by engineer Francesco Zinghinì —an expert in Fintech systems and CRM development for credit management—we will explore how advanced prompt engineering and Large Language Models (LLMs) are revolutionizing financial back-office operations. We will build an enterprise-grade document processing pipeline using Retrieval-Augmented Generation (RAG) techniques on leading cloud platforms such as Google Cloud Vertex AI and AWS Bedrock . The goal? To reduce approval times from weeks to just a few hours, while ensuring maximum security and privacy for sensitive data (PII).

Prerequisites and System Architecture

Before writing the first line of code or the first prompt, it is essential to define a robust architecture. Analyzing financial documents requires a deterministic approach: we cannot afford AI model hallucinations when assessing an applicant's income.

The tools and prerequisites for implementing this solution include:

• Cloud Platform: Google Cloud Platform (GCP) with Vertex AI RAG Engine, or AWS with Amazon Bedrock and Bedrock Data Automation.
• OCR (Optical Character Recognition) engine: Google Document AI or Amazon Textract for extracting raw text and layout from scanned PDFs.
• Vector Database: AlloyDB for PostgreSQL (on GCP) or Amazon OpenSearch Serverless to store document embeddings.
• Orchestrator: LangChain or LlamaIndex (in Python) to manage the logic flow, or native serverless frameworks such as AWS Step Functions.
• Target CRM: Salesforce, Microsoft Dynamics, or a proprietary CRM exposed via REST API.

According to official AWS Bedrock documentation, using Agents for Amazon Bedrock enables the orchestration of complex workflows, securely invoking enterprise APIs (such as a CRM) only after validating the extracted data. On the Google side, Vertex AI Search serves as an optimized retrieval backend, ensuring that the LLM (such as Gemini 1.5 Pro) bases its responses exclusively on the documents uploaded for the specific mortgage application.

The Role of Retrieval-Augmented Generation (RAG) in Financial Back-Office Operations

RAG is the beating heart of our pipeline. Generic language models do not know the details of "Mr. Rossi's" mortgage application. RAG solves this problem by injecting specific context directly into the model's prompt.

In the context of mortgage underwriting, the RAG process comprises three critical phases:

1. Ingestion and Chunking: Documents (e.g., Form 730, Certificazione Unica, valuation reports) are processed using OCR. The extracted text is divided into semantic "chunks" (fragments). For financial documents, it is vital to use a chunking method that respects tables and logical sections, avoiding the splitting of a financial statement line item in the middle.
2. Embedding: Chunks are converted into high-dimensional numerical vectors and saved in the vector database.
3. Retrieval and Generation: When the system needs to calculate net income, it queries the Vector DB to find the most relevant chunks (e.g., Section RN of Form 730) and passes them to the LLM with a prompt structured for extraction.

"The most common mistake when implementing AI for mortgages is treating financial documents as simple continuous text. Tables, merged cells, and data hierarchies require advanced OCR and RAG that is aware of the document's spatial structure." – Francesco Zinghinì

Document Processing Pipeline: Step-by-Step

Let's see how to build the pipeline step by step, simulating an architecture based on AWS Bedrock and Lambda functions (or their Cloud Run equivalents on GCP).

Step 1: Acquisition and Classification

The client uploads a batch of mixed PDFs via the web portal. The AI's first task is document classification . We use a fast LLM (such as Claude 3 Haiku on Bedrock or Gemini 1.5 Flash) to analyze the first page of each document and categorize it.

The system will label the files as: BUSTA_PAGA , ESTRATTO_CONTO , CARTA_IDENTITA , COMPROMESSO . If a mandatory document is missing, the system immediately sends a notification to the client, eliminating back-office downtime.

Step 2: Data Extraction

Once classified, the documents proceed to the extraction module. Here, we use more capable models (Claude 3.5 Sonnet or Gemini 1.5 Pro) configured with a temperature of 0 to ensure maximum determinism and eliminate creativity (and, consequently, hallucinations).

Step 3: Cross-referencing and Validation

AI does not simply read one document at a time; its true added value lies in cross-referencing . The system verifies that the net salary credited to the bank statement (e.g., €2,150 on April 27) exactly matches the net pay shown on the payslip for the same month. Any discrepancy triggers a flag for the human analyst.

Advanced Prompt Engineering: Practical Examples for Financial Data

The secret to perfect extraction lies in prompt engineering . It is not enough to simply ask the LLM, "What is the income?" We must provide rigorous system instructions, define the output format (JSON Schema), and supply examples (few-shot prompting).

Here is an example of a system prompt optimized for extraction from an Italian payslip:

 Sei un analista del credito senior specializzato in mutui ipotecari italiani. Il tuo compito è estrarre dati finanziari chiave dal testo OCR di una busta paga fornita nel tag <document>. REGOLE TASSATIVE: 1. Estrai SOLO i dati esplicitamente presenti nel documento. 2. Se un dato non è presente o è illeggibile, restituisci null. NON indovinare o calcolare valori mancanti. 3. Formatta tutti gli importi monetari come numeri decimali (es. 2150.50), rimuovendo il simbolo dell'Euro ei separatori delle migliaia. 4. L'output DEVE essere un JSON valido conforme al seguente schema: { "mese_competenza": "MM/YYYY", "datore_di_lavoro": "Nome Azienda", "tipo_contratto": "Indeterminato | Determinato | Apprendistato | Altro", "netto_in_busta": 0.00, "trattenute_cessione_quinto": 0.00 } # Sei un analista del credito senior specializzato in mutui ipotecari italiani. Il tuo compito è estrarre dati finanziari chiave dal testo OCR di una busta paga fornita nel tag <document>. REGOLE TASSATIVE: 1. Estrai SOLO i dati esplicitamente presenti nel documento. 2. Se un dato non è presente o è illeggibile, restituisci null. NON indovinare o calcolare valori mancanti. 3. Formatta tutti gli importi monetari come numeri decimali (es. 2150.50), rimuovendo il simbolo dell'Euro ei separatori delle migliaia. 4. L'output DEVE essere un JSON valido conforme al seguente schema: { "mese_competenza": "MM/YYYY", "datore_di_lavoro": "Nome Azienda", "tipo_contratto": "Indeterminato | Determinato | Apprendistato | Altro", "netto_in_busta": 0.00, "trattenute_cessione_quinto": 0.00 }

By providing this prompt to a model that supports JSON Mode (such as the Vertex AI or Bedrock APIs), we obtain a structured payload ready to be injected into the CRM's relational database.

Calculation of the Debt-to-Income (DTI) Ratio and Anomaly Identification

One of the key parameters for mortgage approval is the Debt-to-Income (DTI) ratio —specifically, the ratio between total monthly debt payments (including the new mortgage) and net monthly income. Italian banking policies typically set the maximum sustainability threshold at around 30–35%.

AI can automatically calculate this value by aggregating data extracted from payslips and CRIF (Credit Bureau) reports. Below is an interactive widget simulating the calculation logic implemented in the CRM frontend for analysts:

Beyond mathematical calculations, AI excels at anomaly detection (fraud detection). A specific prompt can be configured to compare the employment start date declared by the client with the one shown on their payslip, or to flag recurring outgoing transfers on the bank statement that might indicate a loan not reported to the Central Credit Register.

CRM Integration and Workflow Automation

Data extraction is useless unless it is seamlessly integrated into business processes. Modern architecture entails sending the JSON output generated by the LLM directly to the banking CRM via webhooks or REST APIs.

However, full automation (Straight-Through Processing) for mortgage approval is still not recommended due to regulatory and risk management considerations. The correct approach is the Human-in-the-Loop (HITL) model :

• If the LLM extracts all data with a high confidence score and the DTI is below 30%, the application is pre-approved and sent to the analyst solely for a final signature.
• If the LLM detects anomalies, illegible documents, or a borderline DTI, the application is routed to a senior agent, accompanied by an AI-generated summary highlighting the exact location of the issue (e.g., "Warning: discrepancy between declared income and CUD").

Troubleshooting and Hallucination Management

Working with Large Language Models in the financial sector requires rigorous error management. "Hallucinations" (when the model invents data) are the number one enemy.

How can these risks be mitigated according to Google Cloud and AWS best practices?

1. Strict grounding: Use grounding APIs (such as Vertex AI Grounding) to force the model to cite the exact source (PDF page and paragraph) for every extracted number.
2. Downstream validation: Do not blindly trust the JSON. Implement Python scripts to verify data types (e.g., ensuring the "income" field is a float rather than a string) before sending the data to the CRM.
3. Context Window Management: Mortgage files can exceed 500 pages. Although models like Gemini 1.5 Pro support millions of tokens, including too much noise degrades performance. It is crucial to filter out irrelevant documents (e.g., advertising pages in bank statements) before passing them to the LLM.

Conclusions

Applying prompt engineering and generative AI to mortgage application analysis represents a quantum leap for the banking sector. As demonstrated in this technical guide, the combined use of advanced OCR, RAG architectures on AWS Bedrock or Google Cloud Vertex AI, and rigorously structured prompts makes it possible to transform a manual process taking weeks into a digital workflow completed in just a few hours.

The goal is not to replace the credit analyst, but to empower them. By eliminating the tedious tasks of data entry and document verification, credit professionals can focus on complex risk analysis and client advisory services. Banks and credit brokers adopting these technologies in 2026 will not only slash operating costs but also deliver an unprecedented customer experience, ensuring fast, transparent, and secure approvals.

Frequently Asked Questions