China wants AI inside 20,000 software companies

01

WHAT ACTUALLY CHANGED

China's Ministry of Industry and Information Technology released a dedicated action plan for integrating AI into the software and information technology services industry. The State Council and State Council Information Office published matching English summaries on September 11. They describe a two-stage roadmap rather than a completed program, product launch, funding award, or regulatory mandate.

The first headline target lands in 2028. China aims to cultivate higher-standard AI coding tools and development platforms, then extend their use to more than 20,000 major software enterprises. The public English account does not define major enterprise, name participating companies, state a starting adoption level, or say whether a trial account, active developer seat, production deployment, or measured improvement will count toward the total.

The plan also calls for 100 benchmark applications for agent-based software across key industries. That suggests the government wants examples other organizations can copy, not just general-purpose chat tools. The summary does not identify the industries, applications, model providers, evaluation methods, safety controls, or performance thresholds that would earn the benchmark label.

Open source is a separate target. By 2028, the plan seeks to incubate at least five high-quality open-source projects and support internationally competitive open-source communities and industrial clusters. Five is a modest number beside 20,000 enterprises, which makes each project potentially important. Yet no repositories, licenses, governance bodies, maintenance commitments, security processes, or criteria for quality are named in the material available in English.

The second stage reaches to 2030. The plan envisions intelligent upgrades across critical software systems, with AI-powered programming, agent software, and intelligent services becoming new growth engines. It sits inside the wider AI Plus policy push. The official account also says authorities predict China's AI-related sectors will exceed 10 trillion yuan in combined scale by 2030. That is a broad government forecast, not revenue already earned and not a measured outcome of this software plan.

02

WHY THIS MATTERS

A national target can change software markets before the tools prove themselves. Large enterprises may revise procurement rules, train developers, connect internal codebases, and ask vendors to meet preferred standards because a policy direction is clear. That creates real demand. It can also reward visible installation over quiet usefulness if the easiest success metric is how many companies can be placed on a list.

AI-assisted programming is becoming infrastructure for producing other infrastructure. A coding tool can speed up documentation, tests, migration work, and routine implementation. It can also multiply insecure dependencies, licensing mistakes, fabricated interfaces, and code nobody understands. At 20,000 enterprises, the important unit is not prompts sent. It is reliable software shipped with fewer defects, faster review, and a clear record of who approved the change.

Agent software raises the stakes because it can choose steps and act through tools. In a factory, bank, logistics network, hospital supplier, or public service, an agent may touch data, accounts, build systems, or operating controls. A benchmark application should therefore demonstrate boundaries, logging, rollback, human authority, and incident handling as visibly as it demonstrates task completion. Otherwise the benchmark teaches imitation without teaching restraint.

The open-source promise could make the program useful beyond the companies receiving the first push. Public code lets developers inspect assumptions, repair defects, translate interfaces, and adapt systems to smaller organizations. But open source is not a ceremonial repository. A credible project needs an actual license, maintainers with decision rules, reproducible releases, a vulnerability process, documentation, and a community that can challenge the original sponsor.

Measurement will decide whether the roadmap produces evidence or theater. Reach, active use, accepted code, developer time, defect rates, security findings, maintenance cost, energy use, and business outcomes describe different things. China can hit a distribution target while companies quietly stop using the tools. Publishing definitions, baselines, subgroup results, failures, and repeatable evaluations would make the 2028 scorecard valuable inside China and legible beyond it.

03

WHERE IT COULD HELP

• Pilot coding assistants on bounded repositories with test coverage, dependency controls, and mandatory review
• Measure accepted changes, reverted changes, defects, security findings, review time, and maintenance cost instead of counting logins
• Require every agent application to expose permissions, tool calls, decision logs, human approval points, and a tested rollback path
• Release reference projects with clear licenses, reproducible builds, public issue trackers, named maintainers, and vulnerability reporting
• Publish benchmark tasks, baselines, failure cases, industry context, and evaluation methods so other teams can reproduce the result