Fujitsu, Human Resocia post AI-agent back-office wins

The article frames AI agents against two older tools. RPA only repeats pre-registered steps (a fixed scenario) and stalls when a screen layout or document format changes, while generative AI such as ChatGPT stops at producing text and leaves system entry to people. AI agents, by contrast, are given a goal and work out the steps themselves, which is why the article argues they can take on the judgment-heavy parts of paperwork.

The Japanese examples are presented as proof that this is already running in production. Human Resocia combined an AI-agent platform with RPA for standardized steps when writing recruiting ads, against a backdrop of about 16,000 hours a year previously spent at roughly 20 minutes per ad across 4,000 monthly ads. Nihon Densetsu Shizai applied AI to invoice and expense processes handling about 2,000 invoices a month. Osaka's demonstration with Hitachi covered commute-form applications, which run about 10,000 a year with about 5,000 concentrated in April.

The article's caution is that the outcome hinges less on the technology than on picking the right work and keeping a human in the loop. It advises starting with tasks that involve judgment, have varied input formats and carry small consequences if wrong, and warns that figures such as Osaka's up-to-40% result come from a demonstration, not full rollout. For accounting, HR and general-affairs teams, the practical question may be which single process to test first, not whether to adopt the technology at all.