Nowcast Begins Considering Commercialization of 'Local LLM Text Processing Infrastructure' for On-Premise AI Processing of Confidential Data

Nowcast Inc., a member of the Finatext Group providing financial infrastructure for the AI era, has announced that it is considering the commercialization of its 'Local LLM Text Processing Infrastructure,' which performs AI processing on personal and highly sensitive text data without sending it to external parties.

Nowcast developed a prototype of this technology during an internal AI contest, demonstrating its practicality in the task of mapping merchant names from credit card statements. The demonstration achieved a 95.7% cost reduction compared to the OpenAI API and a reduction of approximately 98% in manual verification work. Building on this success, Nowcast will generalize this technology to solve the common challenge in the financial industry: 'AI processing of text data that cannot be sent externally.'

Background The evolution of generative AI is rapidly automating tasks like text classification, entity resolution, and information extraction using LLMs. However, in the financial industry, many operations cannot utilize LLMs because data cannot be sent externally. For tasks handling data such as merchant names on credit card statements, customer addresses and names, B2B transaction data, and insurance claim documents, sending data to cloud LLM APIs like ChatGPT is unacceptable due to privacy and security concerns. This has created a structural issue where LLM capabilities are recognized but not applied, preserving manual labor. Another factor hindering adoption was the high cost of applying text processing to large-scale data using cloud LLM APIs.

With this awareness, Nowcast developed a prototype to automate the merchant name mapping task for credit card statements using a local LLM in an internal AI contest. The prototype achieved high accuracy and a 95.7% cost reduction compared to the OpenAI API without sending sensitive information externally, reaching a level ready for immediate application in internal operations. Following this success, Nowcast decided to consider commercializing this technology not just for merchant name mapping, but as a general-purpose platform to address common challenges in the financial industry when introducing LLM-based text processing.

Overview of the 'Local LLM Text Processing Infrastructure' This platform operates open-weight large language models within a company's own environment (on-premise or private cloud) to perform AI processing such as classification, entity resolution, and information extraction on text data with external transmission restrictions.

◯ Platform Features 1. Data Protection through Complete Local Processing All LLM inference is completed within the company's environment, and the text data being processed is never sent externally. The design is compliant with FISC security standards and Personal Information Protection Act requirements.

2. Overwhelmingly Low Cost A 95.7% cost reduction was verified compared to gpt-5-nano, the cheapest model currently available from the OpenAI API. Even when processing large amounts of data with LLMs, costs can be kept within a realistic range.

3. Human-in-the-Loop Design Based on Confidence Scores The system assigns a confidence score to LLM processing results, adopting a phased judgment flow where high-confidence cases are automatically confirmed, and only low-confidence cases are passed on for manual review. This balances processing accuracy and operational efficiency.

4. Model-Agnostic Architecture The processing pipeline is designed to be model-independent, allowing for flexible model switching as open-weight models evolve. This avoids lock-in to a specific AI vendor.

5. Horizontal Deployment with Business-Specific Templates The design allows for setting task definitions, confidence thresholds, and post-processing rules for each business task, enabling application to multiple operations on a single platform.

Prototype Verification Results for 'Credit Card Merchant Name Mapping Automation' The prototype developed in the internal AI contest targets the 'mapping' task of linking merchant names on credit card statements to their official brand names.

◯ Challenge Merchant names on credit cards are often recorded in formats different from the official brand name (e.g., abbreviations, katakana, strings with terminal-specific codes). This made manual mapping an essential pre-processing step for payment data analysis, with Nowcast handling about 300,000 cases per month.

◯ Verification Results - Throughput: 15.1 items/sec (processing 1 million items in approx. 18.4 hours) - Accuracy: Over 95% for cases with a confidence score of 90 or higher - Manual Check Reduction Rate: Approx. 98% - Processing Cost (1 million items): Approx. ¥4,500 (based on electricity and GPU depreciation) - Cost Reduction vs. OpenAI API: Approx. 95.7% (compared to performing the same process with gpt-5-nano)

◯ Supplement: Use of Web Search For cases where estimation by the LLM alone is difficult (e.g., new stores, special abbreviations), the system supplementarily retrieves brand name candidates via web search to improve accuracy. The search queries use only non-sensitive brand name candidates, and the merchant name itself is not sent externally.

Envisioned Application Areas This platform is envisioned for application not only in merchant name mapping but across all 'text processing tasks that cannot be sent externally' faced by financial institutions. - Merchant Name Mapping: Credit card companies, payment processors, household finance apps - Address Normalization: All financial institutions, real estate, logistics - Corporate Name Entity Resolution: Credit management, anti-social forces checks, CRM - Transaction Description Classification: Banks, accounting services - Fraud Detection Text Analysis: Insurance, securities, credit card companies - Structuring after Document OCR: Insurance assessment, loan screening

Expected Implementation Effects At Nowcast, the company anticipates annual savings of over ¥25 million from merchant name mapping alone, and over ¥40 million when including deployment to other tasks.

Regarding Commercialization Nowcast has begun considering the external provision of this platform to financial institutions and data businesses. Specifically, the following forms of provision are being considered: - Support for on-premise environment implementation (including platform construction, business-specific template design, and operational support) - Managed service provision in a private cloud environment - Licensing of business-specific mapping databases and dictionaries

Future Outlook For over 10 years, Nowcast has handled alternative data such as credit card payment and POS data, and has a track record of meeting the strict security requirements of data holders. By combining this experience with local LLM technology, Nowcast will contribute by providing an environment where even operations handling personal and confidential information can benefit from AI.