Mitsubishi Corporation: Proof-of-concept project to reform accounting operations using generative AI

Japanese

PwC Tax Japan has supported Mitsubishi Corporation in a proof-of-concept project to reform accounting operations using generative AI. The project was conducted over two months from April to May 2024. It achieved high accuracy, including an average accuracy rate of 97% in extracting information from contracts and balance certificates related to guarantees, as well as a 98% recall rate in determining the necessity of submitting payment records.

While use cases such as inquiry responses (e.g. chatbots) are common for generative AI in the financial and accounting fields, this proof of concept demonstrated the potential for integrating generative AI into actual business processes, showcasing its practical application. Future plans include exploring further support to expand use cases.

Mitsubishi Corporation, as a leading trading company, operates numerous domestic and international offices, engaging in diverse business activities. The company aims to foster an environment where employees can focus on value creation towards realising the company’s management principle of ‘Creating MC Shared Value’. This pursuit involves reforming business processes, including accounting operations, where substantial manual work is involved. Accounting staff often manually extract necessary information from large volumes of documents, such as contracts and invoices, conducting tasks like categorising and inputting data. In collaboration with Mitsubishi Corporation’s Accounting Department (Business Process Coordination Office), PwC Tax Japan has verified the potential for automating and streamlining these processes via AI-OCR (artificial intelligence-powered optical character recognition) combined with generative AI. This marks a first step towards automated and efficient document reading and subsequent accounting processing.

• To prepare disclosure documents such as securities reports, extract information related to guarantee obligations (e.g. ‘guaranteed company’, ‘maximum guarantee amount’ and ‘guarantee balance’) from various documents, including contracts and balance certificates, achieving an average accuracy rate of 97%, and compile a list to serve as a database. 
• For creating payment records, which are part of tax filing, extract information such as the company name and description from invoices, and determine the necessity of submitting these payment records, achieving a recall rate of 98%.

Figure 1: Overview of automated processing flow utilising AI

PwC Tax Japan provided comprehensive support for the proof-of-concept project, offering services ranging from requirements definition to development and verification. This included the construction of an automated processing workflow using generative AI and the integration of the insights and experience of tax professionals into prompts.

Specifically, in April 2024, PwC Tax Japan established an automated processing flow encompassing PDF data preprocessing, text conversion via AI-OCR, automatic data extraction with generative AI and the determination of the necessity for payment record submission. In May 2024, an agile development approach was adopted, organising weekly sprints to incorporate feedback from Mitsubishi Corporation representatives, implement cutting-edge methodologies and refine prompts. This iterative process focused on aligning key areas as needed to enhance accuracy.

Three effective strategies employed by PwC Tax Japan in the accuracy improvement process, utilising the expertise and experience of generative AI task forces and tax professionals, are highlighted below:

1. For determining the necessity of submitting payment records, merely reflecting legal definitions in prompts did not lead to improved accuracy. Therefore, PwC Tax Japan embedded tax advisor-specific knowledge into prompts, incorporating more specific examples and explanations.
2. Utilising a technique known as ‘Dynamic Few-Shot Learning’, which combines search AI with databases, they integrated information from Mitsubishi Corporation’s past determination results and internal manuals into prompts.
3. In preparing foundational disclosure documents for guarantee obligations, extracting necessary items from diverse data such as contracts, balance certificates and guarantee documents proved challenging with prompt improvement alone. Thus, they implemented a ‘Multi-Agent’ system, deploying specialised AIs such as ‘contract-specialised AI’ and ‘balance certificate-specialised AI’, allowing multiple AIs to collaborate effectively.

Figure 2: Future vision for improving efficiency and automation of accounting operations using generative AI

This proof-of-concept experiment demonstrated the potential for long-term efficiency and automation of accounting processes by designing business processes with an eye towards the exponential evolution of generative AI and establishing a framework that enables a cycle of accumulating in-house expertise and updating AI models. PwC Tax Japan will leverage the insights gained from this experiment to expand the scope of accounting processes targeted for AI adoption, while continuing to provide support for practical implementation, including organising use cases for AI processing and conducting additional verification. We will continue to provide support for practical implementation, including the organisation of use cases suitable for AI processing and additional verification. 

PwC Tax Japan is strengthening its support for business reform utilising generative AI in collaboration with the PwC network in Japan and overseas.^*3 In particular, as an example of the utilisation of generative AI in the tax field, it is now possible to improve the efficiency of internal document analysis and tax-related considerations, as well as the quality of tax-related operations. We are incorporating these insights into our support services for clients. Through these initiatives, we will further strengthen our support for clients in reforming tax-related operations and creating value through the practical application of generative AI.