Building the “Pipeline” for AI: Why Most Corporate Pilots Fail — and Why China Is Betting Everything on Diffusion

In the summer of 2025, MIT’s NANDA initiative dropped a sobering report titled The GenAI Divide: State of AI in Business 2025. Despite $30–40 billion poured into generative AI pilots and experiments across enterprises, 95% of organizations saw zero measurable return on their investment — at least in terms of profit-and-loss impact.

The models themselves were not the problem. By mid-2025, frontier AI systems were already demonstrably capable across a wide range of tasks. The bottleneck was something far more mundane and human: getting from pilot to production.

Companies were happy to run experiments in controlled sandboxes. Far fewer were willing to redesign workflows, retrain staff, rewire decision-making processes, or accept short-term friction and risk to integrate AI into core operations.

As a result, most initiatives remained stuck in the “interesting demo” phase, delivering neither revenue acceleration nor cost savings at scale. Only the top 5% of pilots succeeded by embracing organizational resistance as the necessary crucible for real adaptation.

---

China’s Different Playbook: AI as “New Quality Productive Force”

Now shift focus to China. In its 15th Five-Year Plan (2026–2030) and the accompanying “AI+” action plan, artificial intelligence is not framed primarily as a corporate productivity tool or a flashy new technology to experiment with. It is declared one of the core “new quality productive forces” — a strategic driver meant to permeate the entire real economy.

The language is deliberate and systemic. The goal is not merely to develop better models in research labs, but to build the conditions, infrastructure, incentives, and integration pathways so that AI flows into every factory, hospital, logistics network, farm, and government service.

Concrete targets underscore the ambition:

• By 2027: >70% penetration of new-generation intelligent terminals and AI agents in key sectors.
• By 2030: >90% penetration.
• Longer term: Ubiquitous deployment across the economy and society.

This is classic technology diffusion thinking at national scale. China is investing heavily in the “pipes” — compute infrastructure, high-quality sector-specific data, agent ecosystems, embodied AI (robots), integration standards, and policy incentives that encourage (or require) adoption in traditional industries. The bet is that once the infrastructure and incentives are in place, AI will stop being a series of isolated experiments and become embedded infrastructure itself.

---

The “Pipeline” Metaphor

Think of AI not as a single brilliant invention sitting on a shelf, but as a fluid that needs an extensive distribution network to reach end users at scale.

Western corporate efforts have largely focused on improving the fluid (better models, more capable agents). China, by contrast, is prioritizing the entire delivery system:

• Compute clusters and data pipelines;
• Standardized interfaces for integration into legacy systems;
• Training and reskilling programs tied to real workflows;
• Regulatory and financial incentives aligned with national adoption goals;
• Feedback loops from real-world deployment back to model improvement.

In scientific and innovation studies, this distinction between invention and diffusion is well-established. Many transformative technologies (electricity, the internet, smartphones) only delivered their full economic impact once the supporting infrastructure and organizational changes caught up. China appears determined to compress that timeline through coordinated national effort.

Conceptual infrastructure for large-scale AI deployment — the “pipes” that turn models into economy-wide productivity gains.

Also read:

Sony’s Table-Tennis Robot “Ace” Just Beat Human Pros — And It’s Far More Impressive Than a Marathon-Winning Bot

Hey CEO Who Can’t Stop Asking ChatGPT for Strategy Advice — HBR Just Dropped the Receipts

Chatbots and the Crazies: Danish Study Confirms What We All Suspected — ChatGPT Is Rocket Fuel for Psychosis

DeviantArt Bet Big on Generative AI — And It Paid Off Handsomely

---

Lessons Without Copying the Model

No one is seriously suggesting that market economies simply replicate China’s centralized, party-guided approach. The political and institutional differences are profound.

That said, several elements of the Chinese strategy are worth serious study and selective adaptation:

1. Infrastructure as a public/private good — Coordinated investment in compute, data, and integration standards that individual companies struggle to build alone.
2. Focus on integration over isolated pilots — Policy and funding mechanisms that reward (or require) moving from experiments to production deployment in real operations.
3. Sector-specific roadmaps — Clear targets and support for embedding AI in traditional industries (manufacturing, agriculture, healthcare, logistics), not just digital-native sectors.
4. Feedback loops between deployment and improvement — Treating real-world usage as the primary driver of capability advancement.

The MIT report makes one thing crystal clear: in the current environment, superior models alone are insufficient. The organizations that succeed are those willing to change how they work, not just what software they trial.

China’s approach may be uniquely enabled by its governance model, but the underlying diagnosis — that the real challenge is scaling and embedding rather than inventing — is universal.

Other countries and companies that want to capture AI’s productivity promise will need their own versions of robust “pipelines.”

Whether through smarter public-private coordination, new financing mechanisms for integration projects, or cultural shifts that treat process redesign as core to AI strategy, the lesson from both the MIT data and China’s five-year ambitions is the same:

The models are ready.
The real work is building the system that lets them actually flow.