How Tencent Games Leverages AI to Turn Data Governance into a Service

Research background and industry pain points – With the global expansion of Tencent Games and explosive growth of data volume, the existing data‑governance model suffered from resource waste, low collaboration efficiency, and passive decision‑making.

Four stages of data‑service evolution – The team identified four key phases: (1) custom game support, (2) control stage (similar to traditional data‑mid‑platform), (3) a stable‑settlement stage focused on method and tool consolidation, and (4) an AI‑optimization stage currently in progress.

Challenges in the stable stage – Even when processes were controllable, three major problems remained: passive resource‑cost control, severe storage‑resource waste (e.g., 30 % of historical partitions idle, 20 % of compute resources consumed by dead tasks), and fragmented collaboration where metadata, lineage, and metric definitions relied heavily on manual maintenance.

Limitations of traditional governance – Conventional workflows followed a linear “business request → development → asset” path, with governance relying on rule engines that only captured single‑point metadata (table schema, SQL task). This approach lacked semantic depth, suffered from delayed updates, and could not support high‑level service needs.

AI‑driven semantic governance paradigm – The new design replaces rule‑based control with a semantic‑centered framework. It includes three technical pillars: (1) metadata semantic learning that extracts naming conventions and development logic from historic SQL logs and table structures; (2) demand‑chain reverse inference using AI agents to map technical lineage (task‑table relationships) back to original business needs, achieving “SQL‑to‑Text” conversion; (3) a unified semantic model that fuses general‑model knowledge, industry‑specific gaming terminology, and internal “black‑talk” into a three‑layer knowledge base, ensuring consistent interpretation across teams.

Two concrete governance scenarios – (1) Resource intelligent control : By annotating data‑lineage with semantic tags, the system generates business‑readable cost‑analysis reports (e.g., a new holiday event adds 10 tasks, 5 tables, increasing storage cost by 15 %). It also automatically identifies stale dashboards (180 days of no access) and notifies owners, replacing manual tracing. (2) Collaboration efficiency : An AI‑powered Q&A search retrieves metric definitions, underlying SQL, cost, and reuse information, cutting metric‑definition alignment time by 50 %. An AI‑assisted demand assistant decomposes new requests (e.g., “National Day activity data”) and recommends high‑quality reusable assets, reducing development time from 1 hour to 30 minutes.

Practice outcomes – After semantic coverage, manual maintenance workload dropped by ~30 %, storage growth slowed dramatically, and resource‑decision latency vanished. Asset recommendation achieved a Top‑3 accuracy of 95 %, while development efficiency improved by 50 %. The initiative catalogued 5 500 asset tables and 100 000 fields, automatically evaluating and standardising them into a high‑quality asset pool.

Future direction – The team plans deeper data + AI integration, turning governance from a cost centre into a data‑service provider, and extending the semantic asset model to support broader industry scenarios beyond gaming.