Dentsu Group is integrating generative AI into its accounting system's approval workflows to improve operational efficiency and ensure consistent quality. Within its accounting system, "CAS," the company has implemented a mechanism where AI verifies the matching of order information with quotations and assesses compliance with internal regulations.
Dentsu Group Integrates Gemini into Accounting Approvals, Adopting a Hybrid Approach that Retains "Human Judgment" Despite 99% Accuracy
This article is a translation. Read the Japanese original
The selection of a Large Language Model (LLM) proved to be a critical factor during the development process. In initial Proof of Concept (PoC) stages, the agreement rate between human and AI judgments was low at 15.4%, but switching to the Gemini 2.5 Pro model significantly improved the accuracy of the determinations.
Currently, the system operates by having the AI categorize items into three levels: "Problematic," "No Problem," and "Requires Verification." However, rather than leaving all approvals to the AI, the company has adopted a "hybrid approach" where humans make business context decisions based on the AI's findings. This structure ensures that humans remain responsible for long-term investment decisions and exceptional transactions.
The introduction of AI has virtually eliminated hallucinations (responses not based on facts) and has realized improved check quality, reduced costs, and faster processing speeds. Furthermore, from an internal control perspective, the company has implemented measures such as strict prompt management and "shadow deployment" to verify the impact of model updates.
Sources
- 電通、20万件超の会計承認をGeminiでAI化 精度99％でもAIだけに任せなかった理由 （ITmedia AI+、2026-09-29）