From AI Tools to AI‑Native Teams: Paradigm Shifts and Organizational Evolution

Background and Motivation

Industry surveys (NBER, DORA 2025) show that while 89% of companies have put AI into production, the average productivity gain is merely 0.29%, indicating a gap between individual efficiency and organizational effectiveness. Kuaishou’s main‑site team of over a thousand engineers faces a similar dilemma: AI tools are deployed, yet the development pipeline remains unchanged.

Three‑Layer Paradigm (L1‑L3)

Kuaishou classifies AI‑enabled development into three stages:

L1 – AI Assistance : AI provides information in isolated steps.

L2 – AI Collaboration : AI generates first drafts; engineers refine them.

L3 – AI Autonomy : AI completes end‑to‑end delivery.

Since early 2024, code‑generation rates rose from 17% to 30%, but overall delivery cycles did not shrink. Detailed analysis showed that less than 10% of engineers truly changed their work style; the difference lay not in using AI, but in how they collaborated with it.

Key Insights from Practice

Insight 1: Faster AI amplifies human‑to‑human collaboration friction.

Insight 2: Human‑AI collaboration itself becomes a hidden cost, manifested in four friction types:

Manual bridging : Humans transfer context between AI and legacy systems.

Context alignment : Engineers must continuously supply business background for AI.

Verification & correction : AI‑generated code often requires hours of manual validation.

Capability boundary judgment : Mis‑ or over‑estimation of AI limits leads to rework.

These frictions stem from two constraints: (1) Capability constraints – AI’s uncertain reliability and the need for extensive human verification; (2) Structural constraints – existing development infrastructure is designed for human collaboration, not AI, causing information, process, and organizational mismatches.

Three‑Tier Reconstruction

Information layer : Re‑architect the knowledge base so AI can access accurate, up‑to‑date information; build AI‑friendly infrastructure and robust validation mechanisms.

Process layer : Shift from a human‑centric SDLC to an Agent‑driven workflow where intent is captured, structured, and consumed by AI agents, while human collaboration is confined to high‑level alignment and decision‑making.

Organization layer : Redefine roles and talent profiles; split delivery and guardianship, introduce “function owners” who co‑drive with AI, and let architects focus on standards and safety. Enable product, design, and operations teams to own parts of the delivery chain, dissolving traditional boundaries.

Organizational Evolution and Talent Model

The new talent model treats an engineer’s value as base traits × capability foundation × core lever . AI‑native engineers must combine deep domain knowledge with AI‑augmented skills and the ability to shape AI‑driven processes.

Case Study: Live‑Gift Production

In a live‑gift workflow, AI agents now handle direction, storyboard, asset generation, and review, reducing the release cycle from 20 days to under 4 days. This exemplifies the three‑layer reconstruction in a concrete product scenario.

Results and Ongoing Challenges

L2 has become the mainstream paradigm, shortening delivery cycles by 20‑30% compared with L1. L3 shows promising gains in specific operational contexts. However, scaling verification, especially for C‑end scenarios, and balancing pioneer‑team agility with enterprise‑wide standardization remain open problems.

Conclusion

AI’s true impact lies not in model capability alone but in redesigning the surrounding development structure. Only by reshaping information flow, process orchestration, and organizational topology can AI’s speed be transferred to real organizational efficiency.