Kuaishou's AI Productivity System Takes Shape: From Tool Efficiency to Organizational Redesign

After three years of AI R&D practice, when Kuaishou replicated to a 10,000-person R&D organization in 2026, it discovered a counterintuitive phenomenon: the more developers involved in the same requirement, the smaller the efficiency gain from AI. Macro data showed that the proportion of L2+ requirements did not grow in sync with per-capita requirement delivery.

Drill-down analysis revealed three major bottlenecks: first, polarization in personnel AI capabilities; December 2025 data showed 30% of personnel had AI code generation rates above 40%, but 32% were below 10%; second, friction in process and division of labor; AI accelerated development and testing, but actual development time per day was only about 30%, with the rest spent on requirement alignment, collaboration communication, and task handoffs; third, under siloed organizational structures, benchmark team experiences were hard to replicate, and benchmark teams were mostly business-product-R&D closed-loop teams.

Kuaishou attributed this to the original framework's assumption that processes, roles, and division of labor remain unchanged, only providing AI tools, practices, and metrics. But the real bottleneck appeared in the framework's blind spot—organization and people.

Kuaishou used banking history to map AI R&D paradigms: L0→L1 (1960s) computers replaced manual work, but the organization remained unchanged, just a 'faster abacus'; L1→L2 (1970s) ATMs let machines handle deposits and withdrawals, forcing organizational change; Citibank adapted its organization, and during the 1977 New York blizzard, ATMs kept running, increasing retail deposit share from 4% to 13% between 1977 and 1981, while other banks were acquired or failed.

L2→L3 (1990s-2010s) online banking and mobile payments let systems work autonomously, transforming business from 'a place you go' to 'capabilities everywhere,' with organizations rebuilt around scenarios and customer journeys. The core difference: L2 changes organization but business essence remains; L3 changes both business and organization, mutually driving each other.

Mapping to R&D: L1 changes tools, L2 changes organization, L3 changes both business and organization. Kuaishou thus posed the proposition: How should the organization change in L2? What is the picture when both business and organizational forms must change in L3?

The old path 'AI×Tools→AI×R&D Personnel→Drive R&D Organizational Change' hit its limits because R&D couldn't push the entire chain. The new path starts from all employees: AI×Tools→AI×Employees→AI×Teams→AI×Organization, with the key turning point being 'AI×All Personnel,' not just R&D personnel.

Kuaishou shifted the proposition from 'R&D efficiency' to 'AI productivity': with AI capabilities as core supply, targeting all personnel and teams within the organization, systematically improving individual effectiveness, team collaboration efficiency, and business delivery quality. This is not just a name change; goals, supply side, paths, product forms, and organizational forms all need reconstruction.

No one in the industry has fully succeeded because there are three chasms: L0→L1 tool chasm (internal AI product wars, cost allocation difficulties), L1→L2 organizational chasm (Token cost allocation conflicts, siloed departmental walls), L2→L3 business chasm (business AI transformation is harder). Each chasm requires different solutions, beyond the scope of R&D efficiency.

Kuaishou instantiated the 'AI Productivity' system with a five-stage path for AI capabilities to permeate from tools to the organization, simultaneously advancing three main lines: organizational transformation, AI infrastructure, and institutional evolution.

In H1 2026, it adopted a dual-track strategy of 'main channel + fast track': the main channel scales AI R&D paradigm upgrades across business lines, targeting 80% L2+ requirements; the fast track selects over 30 AI pioneer teams, ranging from 10 to 100 people, exploring the upper limits of AI organizational evolution without boundaries, potentially exploring L2→L3 or directly from L1→L3.

Before practice, using 'team type' and 'delivery type' as two axes, R&D teams were divided into four quadrants, with different exploration goals and methods for each quadrant, avoiding being too aggressive or conservative.

After half a year of practice, Kuaishou distilled the 'AI Evolution Theory for R&D Organizations,' applicable to all types of R&D teams. The four team types differ, but there are clear commonalities and evolutionary trends.

Case 1: The commercialization risk control team, a high-adversity business, where traditional delivery models couldn't keep up with the pace of black-market AI attacks, chose to leap from L2 to L3, with a three-step transformation (specific steps not detailed in the summary).

Practice showed that changes in AI-Native organizations are not limited to R&D methods; what is delivered to users is also changing, for example, from SaaS platform services to AI services providing Agents, which becomes a key signal for understanding L3.