How Zhihu Boosted R&D Efficiency by 38% with AI‑Powered End‑to‑End Collaboration

Background and Challenges

Traditional product‑research‑development (R&D) at Zhihu suffered from low collaboration efficiency, high labor cost, and difficulty preserving knowledge. Single‑point AI tools could not break these bottlenecks, prompting the team to embed AI throughout the entire R&D lifecycle.

AI‑Empowered End‑to‑End Workflow

The organization built a full‑chain AI empowerment system covering six stages: requirement analysis, architecture design, development, testing, release, and operation. AI agents performed the following core tasks:

Indexed raw requirement documents stored as Markdown and generated detailed PRDs via conversational queries.

Converted PRDs into product prototype diagrams by generating HTML code and rendering it.

Created operation flowcharts with Mermaid based on business processes.

Produced initial architecture proposals; after iterative dialogue the proposals reached ~70% usability and were refined into complete designs.

Generated test cases, of which 52.4% required no modification; the remainder were refined through AI‑human dialogue.

Generated front‑end page code from HTML prototypes and validated it against test cases in batches.

Generated backend HTTP interface definitions, mock servers, database schemas, and data‑loading scripts, cutting backend effort from 10 person‑days to 6 person‑days (≈40% time saved).

Quantitative Results

Overall team efficiency increased up to 38.6% .

Weekly labor savings of 22.5–25 person‑days for an 18‑person team.

Front‑end development speed up by 40% .

Backend development time reduced by 40% .

Test‑case generation time dropped from 30 min to 6 min per feature.

Problem‑resolution time improved by 49% .

New‑hire onboarding time shortened by 90% (from two weeks of mentorship to two days of self‑learning via AI agents).

Knowledge‑Base Construction Methodology

A structured knowledge base was created to store not only code but also business logic, requirement documents, design specs, and test cases. When the repository grew beyond 1,000 entries, a hierarchical “catalog” (directory) was introduced to improve retrieval precision and reduce context overload. Continuous accumulation of knowledge enabled the AI to handle legacy project refactoring as effectively as new projects.

Success Factors

Systematic knowledge organization – a “knowledge‑first” approach reduced learning cost and enabled document‑first collaboration.

Standardized collaborative processes – minimized synchronous meetings and clarified hand‑offs.

Shift to human‑AI collaboration – team members focused on task decomposition, instruction formulation, and result review rather than low‑level coding.

Future Outlook

Zhihu plans to evolve from relying on external experts to internal AI‑augmented experts and eventually “digital employees.” The roadmap includes deeper automation of the toolchain, tighter AI integration in decision‑making, and scaling the knowledge base into a core strategic asset for the enterprise.