Why Strong AI Models Still Fail: Managing AI Employees in Enterprises

Since the release of GPT‑3.5, the pace of AI advancement has accelerated, and at NetEase’s 2026 "Intelligent Journey" conference Ruan Liang, Vice President of NetEase and General Manager of NetEase Zhìqǐ, argued that the problem is no longer model capability but how enterprises manage AI employees.

In 2023‑2024 NetEase’s internal teams were anxious that employees wouldn’t use AI; by 2026 the anxiety reversed to fearing uncontrolled AI use. The concern is not token waste but security‑related risks such as data leakage, unauthorized calls, and accidental deletions caused by AI agents.

Ruan distilled the transition from pilot to production into five “gaps”:

Knowledge gap – enterprises must supply their own domain knowledge, methodologies, client records, and SOPs as context for AI.

Data gap – CRM, ERP, OA, and email data reside in isolated silos, preventing AI from accessing a unified context.

Process gap – sales, pre‑sales, product, R&D, delivery, legal, finance, etc., require coordinated workflows; AI that handles only one segment cannot deliver real efficiency.

Governance gap – AI agents introduce prompt‑injection, sensitive‑data exposure, and over‑privileged calls, creating new security challenges.

Value gap – competition is no longer about the best model (which converges by 2026) but about turning AI into a reliable, deliverable employee.

Based on these gaps, NetEase Zhìqǐ has operationalized four AI‑employee categories:

AI Sales – automates CRM entry, visit summaries, daily reports, and provides instant SKU knowledge and 360° customer portraits, while also acting as an always‑on mentor for junior sales.

AI Private‑Domain Assistant – reconstructs the entire enterprise‑WeChat operation chain, using user tags, behavior traces, and data models to generate personalized audience segments and scripts, and embeds a 20‑day WeChat nurturing SOP.

AI Coding (SDD) – replaces the “hero‑developer” model of Vibe Coding with Spec‑Driven Development, where AI first reads specifications, then generates code under unified constraints, enabling reusable project assets and eliminating drift.

AI Security Governance – provides a full‑stack AI‑agent protection suite (prompt shielding, sensitive‑data masking, real‑time high‑risk behavior interception, end‑to‑end audit) to make AI usage safe enough for core business processes.

Treating AI as an employee requires an HR‑style management system. NetEase’s “帝王蟹” platform offers recruitment (model selection), role definition (agent job description), compliance checks at every step, and performance assessment with continuous benchmarking, allowing underperforming agents to be retrained or reassigned.

The platform also acknowledges that model capabilities evolve; instead of a one‑time purchase, models become on‑the‑job staff that can be re‑allocated, trained, or promoted as their abilities change, thereby continuously bridging the governance and value gaps.

In summary, AI must move from a showcase to a reliable employee. Overcoming the five gaps—knowledge, data, process, governance, and value—through a systematic AI‑employee lifecycle is NetEase Zhìqǐ’s new mission to unleash enterprise productivity.