In the September 2025 issue of Tax Research, Yao Jian discusses how to optimise AI application scenarios in taxation. Today's scenarios, he argues, neither fully exploit AI's potential nor match the core needs of tax work, and the root cause is that three transitions remain incomplete: scenario goals are still framed as operating tools rather than delegated agents; system capability still depends on preset rules rather than dynamic adaptation; and development is still human-designed rather than machine-learned. He proposes working from four directions: setting goals, running evaluations, laying a foundation and deepening coordination, by defining benchmark scenarios, designing test sets that measure a large model's tax competence, building a shared public foundation for application scenarios, and drawing on outside resources to enrich them, so that tax AI is embedded in real business and supports the strengthening of collection and administration.
News Detail
Optimising application scenarios for AI in taxation
Tax Research (2025, No.9) carries an article by Yao Jian on how to improve AI application scenarios in taxation. Existing scenarios, he argues, neither exploit AI's potential fully nor match the core needs of tax work, because three transitions remain incomplete: scenario goals stay at the level of tools rather than delegated agents, system capability relies on preset rules instead of dynamic adaptation, and development remains human-designed rather than machine-learned. He proposes setting benchmark scenarios, building test sets that measure a large model's tax capability, creating a shared foundation, and drawing on outside expertise.
2026-11-06