Industrial Fieldwork Through an Inclusive Lens: Inclusive AI Agent Benchmark (1)

Starting from the question of who generative AI should reach and how, this redefines inclusive AI for people who are easily left behind in the tech environment, such as the elderly, people with disabilities, and patients.

The core focus is not simple conversational assistance but an inclusive AI agent that helps people who actually work in industrial fields. It argues that beyond merely providing information, the agent needs Agentic Tool Use capability to solve real problems by connecting with external systems.

From this perspective, unlike existing approaches that viewed industry mainly in terms of economic contribution, the focus here is on people who work in industry, particularly fields where many workers aged 50 and above are employed. After examining the proportion of workers aged 50 and above by industry in Korea as of Q2 2024, four industries were selected that have both a high proportion of elderly workers and heavy reliance on digital systems: agriculture/forestry/fisheries, caregiving, convenience store operation, and freight truck dispatching.

The representative tasks derived from each industry are as follows.

• Handling farm machinery breakdowns: A field problem so frequent that local government phone consultation services exist for it
• Managing caregiver duties: Work requiring records and reports to be entered into digital systems
• Managing convenience store products: Complex operational tasks including ordering, inventory checks, and return processing
• Managing freight truck dispatching: Dispatch and operation management tasks directly tied to income

In designing the evaluation, not only the task outcome but also the process of reaching that outcome is examined together. In conversations with elderly users, situations repeatedly occur where necessary information isn't given in time, terminology is forgotten, or unnecessary context is mixed in, so the agent must be able to naturally compensate for this and infer intent to solve the problem.

Based on these criteria, the benchmark was constructed to reflect the real constraints and conversational environment of actual industrial fields. Part 2, coming next, will reveal how effectively various AI models perform in this evaluation environment.