Developers have folded AI into daily coding work. Still, the same tools remain largely absent from the systems that validate and ship software. New research from JetBrains points to a widening gap between how engineers write code on their own machines and what runs inside continuous integration and delivery pipelines.
Workplace use of AI among developers exceeds 90%. The figure comes from three JetBrains studies: the AI Pulse survey from January 2026, the State of Developer Ecosystem report from October 2025, and the State of CI/CD Tools report from the same month.
Most of that activity sits upstream from the pipeline. Developers use AI to generate code, refactor existing logic, explore unfamiliar APIs, and work through debugging sessions. The tools also help with documentation, code comprehension in large codebases, and early-stage security checks. These tasks share a common property: fast local feedback, low cost of error, and the ability to discard a suggestion with no lasting consequence.
Inside CI/CD pipelines, the picture changes. Seventy-three percent of organizations do not use AI in their pipelines at all, according to the State of CI/CD Tools survey. Among those that have held back, 60% cite unclear use cases or value, 36% cite a lack of trust in AI-generated results, and 33% cite data privacy concerns.
The AI Pulse survey puts the number higher still, with 78.2% of respondents reporting no use of AI inside CI/CD workflows.
The issue is less about technical integration and more about whether AI can deliver predictable, measurable value inside a system designed around reproducible signals. Pipelines exist to confirm that code compiles, tests pass, and deployments can be rolled back safely. Non-deterministic outputs sit uneasily alongside that purpose.
Where AI has entered the pipeline, it tends to cluster around three tasks.
Failure diagnosis is the most common entry point. AI processes pipeline logs, identifies recurring patterns, and correlates errors across runs to narrow down likely root causes. Engineers keep control over what gets fixed. In one current setup, the TeamCity CLI works with Claude Code to analyze failing builds and surface probable causes to engineers.
Security workflows represent a second area. AI layers on top of existing scanning tools to interpret findings, suggest fixes, and in some cases generate patches. Those patches still move through standard testing and review before merging.
Test optimization is the third. By analyzing historical test runs and recent code changes, AI prioritizes relevant tests, flags flaky behavior, and cuts redundant execution. Teams often retain periodic full test runs to catch anything that slips through a reduced set.
Teams tend to expand AI use in delivery systems through a predictable progression. The first stage is no AI in the pipeline at all, which covers the majority of organizations. The second involves AI-assisted understanding, where the system explains failures and logs to engineers who make every decision.
The third stage introduces AI-generated proposals such as pull requests, configuration changes, and test updates, with humans reviewing each one. The fourth stage permits agentic workflows in which AI can trigger actions directly inside the pipeline, subject to explicit permissions, full logging, and human approval for anything significant.
Most teams sit in the first two stages. Movement between stages depends on the reliability of validation, policy controls, and pipeline signals around the AI system.
CI/CD functions as an evidence system. Every build, test, and deployment step produces signals that answer whether a change is safe to release. AI raises both the volume and variability of changes entering that system, which puts pressure on three areas: the reliability of pipeline results, the controls governing how changes progress through approvals and audit trails, and the ability of CI/CD platforms to expose pipelines and logs to external tools, including AI agents.
Flaky tests, inconsistent builds, and unclear feedback loops become more visible and more costly once AI-generated changes arrive in higher numbers. Access controls, policy checks, and approval gates move to the center of pipeline design.
Several questions across the industry remain unresolved. Governance mechanisms that scale alongside more capable AI agents are still being worked out. The form of meaningful human oversight when a pipeline processes hundreds of AI-generated changes per day has no settled answer. The point at which CI/CD evidence itself may require auditing by AI systems has yet to be defined.
These questions will drive how CI/CD platforms evolve over the next several years, as AI participation in software delivery moves from assistance in the editor to direct involvement in the systems that decide what ships.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。