We noticed that MIT's data has resurfaced this week: 95% of enterprise AI projects ultimately fail to become sustainable productivity, with each failed project burning through an average of ¥1.89 million. The problem isn't the model—AI can't access precise business context, and it runs on stateless infrastructure (memory is lost the instant a container restarts). That's the real culprit. Zadig's solution this week: turn AI into a regular node in the workflow.
What this is
Zadig (a domestic DevOps platform—simply put, a tool that helps companies automate the entire "code-to-deployment" pipeline) added an "AI Task" node to its workflow task list in v5.0, open to the community for free. The underlying approach isn't complicated: stuff AI into the code commit → build → test → deploy → release pipeline, making it a called-upon step rather than a separate system.
Three typical scenarios stand out. First, pre-release SRE (Site Reliability Engineer—responsible for keeping production services stable) check: AI automatically scans error logs, inspects dependency status, and confirms database migrations—turning 15 minutes into 3, with unified standards. Second, code review: AI flags risky areas first; developers only need to verify "did AI mark it correctly?" Third, integrating existing Agents (AI programs that autonomously execute multi-step tasks): self-built intelligent agents can be invoked by workflows with just an HTTP address—no redundant development needed.
The core principle: AI handles comprehensive checking and organization; humans make the final call. Every execution leaves a trace, fully traceable.
Industry view
The bullish side believes this approach hits the real pain point of enterprise AI. GitLab surveyed 1,528 developers—85% said the bottleneck has shifted from writing code to reviewing code. AI doing the grunt work while humans call the shots is a reasonable division of labor. IDC data from the same period: 90% of enterprise AI POCs (Proof of Concept—think "pilot projects") fail to pass production acceptance. Zadig's solution at least answers the "how do we actually land this?" question.
But reservations exist. One skepticism: this is essentially an extension of workflow orchestration (stringing multiple steps together to run automatically), not a real Agent breakthrough. Turning AI into a called-upon tool versus letting AI make autonomous decisions are two different things—model capability hasn't changed, and the ceiling remains. Another concern is more pragmatic: the "context governance" problem isn't truly solved. AI still can't access precise business context and will still confidently fabricate nonsense. The ¥1.89 million pit won't disappear just because you switched tools.
Impact on regular people
For enterprise IT: Traditional industries dipping toes into AI don't need to build teams from scratch. Open a slot in existing DevOps processes first, validate ROI (Return on Investment) with small scenarios, and avoid heavy upfront investment.
For individual careers: "Writing code" is depreciating; "reviewing code, setting standards, making judgments" is appreciating. The technical worker's core competitiveness is shifting from hands-on ability to defining problems and gating quality.
For the consumer market: You won't see Zadig directly, but you'll feel it—faster software iteration, fewer bugs, more stable SaaS services. AI quietly working in the background is the most underestimated trend of this wave.