Risk‑Driven Delivery and AI‑Powered Intelligent Testing: Strategies and Technical Solutions

Risk‑driven delivery is a key research direction in Baidu's intelligent testing practice. It originates from three observations: (1) most projects have no associated bugs or online issues; (2) many testing tasks fail to uncover defects, resulting in a high proportion of ineffective quality actions; (3) testers can misjudge, leading to missed tests.

If we can accurately identify which projects need testing and assess risk precisely, both testing efficiency and defect recall can be greatly improved.

The article outlines three focus areas: (1) Baidu Search business delivery with unmanned operation and its exploration, highlighting the critical role of risk assessment; (2) Application of AI technologies throughout the risk‑based testing workflow, describing AI‑enhanced scenarios in each phase; (3) Quality assessment models that support risk‑driven decision making.

Background analysis shows a recurring “Q3” phenomenon where online issues spike due to high delivery volume and the influx of new graduates into development and testing. Similar spikes occur after organizational changes or hand‑overs, indicating that personnel turnover and loss of process rigor are root causes.

The testing process is examined from three perspectives: who should test (task allocation), how to test (scope definition), and what issues are observed in the release stage. Subjective human factors heavily influence outcomes, causing repeated inefficiencies.

To mitigate these problems, a risk‑based full‑process intelligent decision system is proposed, aiming to increase machine decision share and reduce reliance on human expertise. The solution includes:

Task allocation based on project risk characteristics to improve efficiency.

Structuring project experience and code analysis so machines can precisely define test scope and reduce missed tests.

White‑box code analysis and test‑case reduction to achieve high‑quality testing at low cost, with AI‑assisted problem localization.

Introducing a project quality‑score model at the release stage to make data‑driven release decisions.

Implementation details cover:

Automatic task distribution system that builds tester profiles from historical data and matches projects to suitable personnel or automation.

Precise test‑scope analysis using offline mining of code call graphs (both backend methods and frontend components) to recommend targeted regression tests.

Intelligent testing that dynamically creates CI pipelines based on white‑box analysis, pools execution containers, and filters test cases to the minimal necessary set.

Smart problem‑localization that aggregates environment, middleware, and log data, applies rule‑based knowledge bases, and ranks issues for automated remediation.

Risk‑based release assessment using a quality‑score model trained on historical delivery data, enabling early “准出” decisions and reducing manual gatekeeping.

Field trials within Baidu show significant gains: task execution time reduced from 90 minutes to under 30 minutes, test‑case volume cut to 10‑50 % of the original, and release‑gate efficiency improvements such as 2,172 person‑days saved and a reduction of release‑gate waiting time from 50 hours to 2 hours.

The overall conclusion is that introducing an AI‑driven, risk‑aware testing ecosystem can dramatically lower subjective human dependence, improve decision quality, and accelerate high‑quality software delivery.