Mitsubishi Corporation deploys MC-GPT on Azure OpenAI for accounting automation

Title

Mitsubishi Corporation deploys MC-GPT on Azure OpenAI for accounting automation

Content

Mitsubishi Corporation, Japan’s largest general trading company (sōgō shōsha), deployed its MC-GPT generative-AI platform built on Microsoft Azure OpenAI Service to more than 1,000 monthly active users within two years of launch, and reached 97% accuracy on contract and balance-certificate information extraction and 98% recall on withholding-slip filing decisions in a controlled accounting-automation experiment with PwC Tax Corporation. The deployment was framed by Mitsubishi’s Mid-term Management Strategy 2024 around the principle of creating an environment where employees can focus on value creation — addressing the gap between that aspiration and the day-to-day reality of manual document handling in finance and accounting. As the case study frames it, in trading companies the saying “people are the greatest asset” has long coexisted with a mountain of work dependent on individuals and manual labor, leaving little time for the strategic work that should have been the priority. Two forces pushed Mitsubishi toward generative AI: a backlog of manual document processing — contracts and invoices in finance and accounting that absorbed large amounts of time and headcount — and the cost structure of previous AI initiatives, which required dedicated systems per use case at budgets of tens of millions of yen each. The November 2022 release of ChatGPT demonstrated that a single technology could cover multiple use cases while collapsing development cost from tens of millions of yen to a fraction of that, but security could not be compromised for a general trading company handling customer information, contracts, and financial data — so a GDPR-compatible, enterprise-grade environment was non-negotiable. Mitsubishi selected Microsoft Azure OpenAI Service for that reason and chose its 100% subsidiary MC Digital as the development partner, in a co-build model where MC Digital ran workshops with each business unit to find use cases rather than operating as a pure vendor. Phase 1 ran March to June 2023: the MC-GPT vision was formalized in March, and a document-summarization system was rolled out company-wide in June — the first step on generative AI. Phase 2 ran April to May 2024: a joint experiment with PwC Tax Corporation targeted accounting automation — specifically information extraction from guarantee-debt contracts and balance certificates, plus judgment on whether withholding slips needed to be filed. Initial accuracy was around 50%, far from usable; switching to a Multi-Agent system that assigned specialist agents per document type, and embedding tax-accountant practical knowledge into prompt design, took the run to 97% accuracy on contract and balance-certificate extraction and 98% recall on withholding-slip filing decisions. Phase 3, from August 2024, launched the MC-GPT Portal — an integrated platform combining document summarization, internal search, translation, and email-drafting support — which has since passed 1,000 monthly active users company-wide. Mitsubishi’s generative-AI rollout has moved beyond efficiency. From fiscal year 2027 the company will require the JDLA G-certificate (an AI literacy benchmark from the Japan Deep Learning Association) for promotion into management, with the goal of building AI literacy across all 5,000+ employees over time. In 2024 Mitsubishi co-founded the AI cloud company PFCI with Preferred Networks and IIJ, and in 2025 it established a roughly 50-billion-yen CVC fund for generative-AI and bio startups. Looking toward 2030, the company plans to extend AI-led workflows beyond accounting into investment decisions, risk analysis, and trading support, and to rebuild core ERP systems on AI-native assumptions — moving from people-driven processes to AI-driven ones with human oversight on the final calls.