China's #1 Short Dramas Aren't Fully Automated—Why Expect AI Coding to Be?

I. Short Dramas: World's First, Still Not Fully Automated

China's micro-drama market reached 50.5 billion RMB in 2024, projected 63.4 billion in 2025, with nearly 700 million users. Overseas, the 2025 market is about $4 billion, over 70% produced by Chinese companies, covering 200+ countries. The production pipeline is extremely industrialized: vertical format, fast pace, strong conflict, script formula libraries, assembly-line shooting bases, data-driven ad buying—two dramas can be prepared in 10 days. Hengdian's industrialization level is a global ceiling for content industries.

Yet even this industry cannot achieve full automation . Scripts need human writers (AI assists but hit logic, paywall pacing, emotional rhythm require human judgment); actors are human (face-swap and AI redraw are nascent); ad buying needs human monitoring (platform rules, creative iteration, ROI games change daily); even translation and dubbing—seemingly the most mechanical step—require human localization review.

A striking figure: 80%–90% of overseas short drama projects fail to break even. The conclusion: the world's #1 ranked, most mature supply chain, most AI-embracing industry still cannot fully automate, and ROI remains perilous. This is the 2026 reality.

II. Why Do People Believe Programming and AI Infrastructure Can Be Fully Automated?

If content generation (widely considered AI's strongest domain) cannot be fully automated, why do so many believe AI coding, enterprise AI system building, and data infrastructure—work that demands deterministic outcomes and carries higher error costs—can be done "fully automatically" by AI? The answer lies not in technology but in cognition, which has been systematically lowered.

First layer: Big model companies' marketing.

U.S. foundation model firms, to tell a story, prop up valuations, and boost stock prices, make "AI full automation" their core narrative. The fuller the story, the larger the valuation imagination space. This is their business strategy—but the cost is a society-wide cognitive distortion.

Second layer: Selective viewing.

U.S. firms showcase "how powerful large models are" but never reveal the true cost of enterprise AI adoption. Internal AI build-out—data governance, process re-engineering, system integration, organizational change—has always been high-cost, heavy-investment. Watching only others' slide decks, not their cost sheets, creates the illusion that "AI build-out should be cheap and fast."

Third layer: Mistaking demos for delivery.

Demo videos show AI generating an app in one sentence, running a full workflow automatically. Those are controlled-environment demos. Real enterprise environments: 80% of Chinese manufacturing still mixes Excel with 20-year-old systems; data is scattered across seven or eight departments in a dozen formats. On such foundations, any "full automation" is a castle in the air.

III. Data Doesn't Lie: Failure Is the Norm, Not the Exception

MIT's 2025 report states: among enterprise generative AI pilot projects, 95% failed to achieve ROI; only about 5% truly deployed. The 95% figure is debated—it includes firms that never tried in the denominator and uses a strict success definition—but even after correction, multiple independent studies align: RAND estimates ~80% AI project failure rate; Gartner notes at least 30% of projects abandoned after POC; S&P Global shows 42% of enterprises scrapped most AI initiatives in 2025.

Thus the practitioner perception that "90% of China's AI build-out fails" matches international data. This isn't China being uniquely bad; it's the industry baseline for this phase.

MIT highlights a key detail: the root cause isn't model capability but "enterprise shortfalls at the AI application layer"—rigid processes, data silos, complex accountability, lack of AI-adapted workflows. In other words, AI build-out fails mostly on organization and management, not technology.

IV. Who's to Blame? The Software?

Clients ask "what can AI bring"; implementers should answer honestly—what's possible, what's not, preconditions, probabilities and boundaries. That's the implementer's duty.

But whether a project succeeds and its effect size depends on combined factors : client budget, data foundation, management maturity, organizational cooperation, top-leader resolve—none decidable by the software vendor alone. Vendors can only advise on internal management; whether advice is heard and acted on is the client's call.

Most common failure modes:

Low budget, high expectations. Spend a few hundred thousand RMB expecting results comparable to tens-of-millions investments. Not a technology limit; the budget structure doesn't support it.

Boss thinks AI is omnipotent, so refuses to pay. Logic: "You're AI, figure it out yourself, why need money?"

Mid-project abandonment. No stunning results in three months, project cut. But data governance and process re-engineering are inherently slow work.

Treating AI build-out as a sales job. Hype the boss, over-promise, sign contract, deliver, effect disappointing, reputation ruined, repeat.

None of these deaths are caused by "software itself being inadequate." 90% of failures are fundamentally cognitive failures.

V. What Now? Step One Isn't Buying Models—It's Raising Cognition

No shortcuts in the short term. Society's overall cognition can only rise incrementally—everyone hits a wall, feels pain, then wakes up. Some will harvest money in the process; that's the cost of this phase.

For enterprises serious about AI build-out, a few pragmatic judgments:

Admit AI is now in a "low budget, low return" phase. Anyone claiming high returns now should show data. MIT's numbers are on the table; mature enterprises still see 95% no ROI—why would your project be special?

Cognition precedes budget. Insufficient budget is manageable; unclear cognition is fatal. Before spending, clarify: what's your data state? Process state? Willingness to change organization? Without answers, money is wasted.

Distinguish "demo" from "delivery" boundaries. For what AI can do, watch controlled demos; for what AI can actually deliver, examine real peer-case deliveries—and ask how much they invested.

Accept "humans are a necessary condition." Like short dramas—the world's #1 industrialized content industry has humans embedded in every link. Coding and AI infrastructure demand higher correctness than content production, so humans step back less, not more.

Closing

China's short dramas are already global #1, yet cannot fully automate; why should AI coding and enterprise AI infrastructure be fully automated?

The real danger isn't insufficient AI capability, but blind trust in foundation model companies' narratives. They hype "full automation" as arriving tomorrow to boost stock prices, while their own enterprise clients face staggering build-out costs. Watching only the first half, not the second, skews cognition.

AI's highest-value use today is augmenting human efficiency, not replacing human responsibility. Raising cognition is the hardest step, but the only one that avoids backtracking—just as the short drama industry walked: first build the production system, then talk automation; first admit humans are indispensable, then figure out how to make humans work better.