Cloud-based web development service Wix has written about a new approach to integrating artificial intelligence into continuous integration and continuous deployment (CI/CD) systems. In a blog post, Wix demonstrates how probabilistic AI can coexist with deterministic development processes, adding chaos without compromising reliability.
In the post, Wix’s Mobile Release Engineering team explains how they successfully implemented AI tools within their CI/CD infrastructure to use AI’s learning capabilities without undermining the predictability that CI/CD systems require. They make it clear that their use of AI is to augment human decision-making rather than replace automated processes.
Itai Schuftan, who leads the Mobile Release Engineering team, explains that the key insight was recognising that not every aspect of CI/CD needs to be deterministic. Whilst builds and deployments must remain bulletproof, tasks such as interpreting logs, triaging errors, and recommending fixes operate in grey areas where AI’s pattern recognition excels.
When it came to CI/CD, almost no one was going near it. Maybe for good reason — it’s a hard, unforgiving space where the margin for error is tiny. But my manager truly believed this was worth exploring. That there had to be a way to combine AI’s growing capabilities with the world of build pipelines and mobile tooling.
Wix was careful to ensure that boundaries were not overstepped, and AI is not used to trigger deployments or make critical infrastructure decisions. Instead, it guides the humans who do, making them more efficient and reducing the cognitive load of managing complex build systems.
One of Wix’s most successful AI implementations addresses analysing build logs, a persistent pain point for DevOps teams. Traditional build logs are lengthy, noisy, and inconsistent across platforms, often burying critical error information amongst dependency warnings and cryptic stack traces.
Wix deployed their internal AI assistant to parse these logs systematically, identifying actual errors rather than simply flagging the first problematic line. The system highlights key issues, attempts to explain them in plain language, and suggests potential solutions. This has reportedly saved hours of back-and-forth communication each week between developers and infrastructure teams.
This has also allowed developers to work more autonomously, as previously, many build failures required developers to contact DevOps engineers for help, creating a bottleneck and interrupting their work. The AI-powered log analysis frequently lets developers understand and resolve issues independently.
Similar approaches are being adopted across the industry. For example, Datadog’s AI-driven monitoring capabilities interpret logs in real time, highlighting errors and diagnosing root causes with plain-language explanations. Their system also has predictive failure analysis and can automatically restart failed services or roll back to stable versions, minimising downtime in production environments.
Wix has implemented Model Context Protocol (MCP), a specification that provides structured context to large language models before processing queries. The MCP server gives context to the AI by supplying build logs, module metadata, and relevant documentation. Connecting the AI to the MCP server lets the system narrow down faults, find where they occurred, and often correctly identify the affected module. The benefits are more accurate summaries, better root cause detection and actionable next steps.
Levelling up the AI by adding aid from MCP to understand domain context has allowed Wix to reduce the verbosity needed to prompt the AI and the hand-holding required in a more generic AI approach.
MCP shifted us from ‘AI assistant’ to ‘AI teammate.’ It no longer needs hand-holding or verbose prompts — it just gets it.
Wix has also developed an automated migration system that addresses code degradation across its mobile applications, which can be very complex and consist of dozens of subprojects. The migration system uses static analysis, heuristics, and LLM capabilities to identify and update relevant module code while suggesting tailored fixes. The migration framework understands specific module contexts, including dependencies, code style, and project settings, and avoids applying generic updates. It generates pull requests with proposed migrations and explanations of changes, allowing migration owners to approve, modify, or reject suggestions.
This system has dramatically improved migration timelines. Wix reports that migrations previously requiring up to three months can now be completed across 100 modules within 24-48 hours, whilst maintaining code quality through human oversight.
The success reflects broader industry adoption of AI-powered migration tools. Sourcegraph’s AI code intelligence drives automated migrations for organisations including Reddit and FactSet, reportedly improving migration speed and reducing bugs by up to 60%. Google uses similar strategies in its generative AI-powered workflows for automated code migrations. It allows AI to autonomously produce verified code changes that pass unit tests, with human engineers retaining approval controls. Google’s implementation has reportedly accelerated migration workflows and cut overall migration time by up to 50% while maintaining code stability and auditability through human oversight.
Carefully and thoughtfully adding AI to their CI/CD pipelines has produced measurable improvements in system stability and developer experience at Wix. By introducing controlled non-determinism in appropriate areas, the CI/CD process has become more stable, predictable, and user-friendly. Wix attributes this success to clearly defining boundaries and keeping human oversight where it’s needed. This methodology is gaining traction across the technology industry. Healthcare platforms like Epic and Meditech use deterministic AI agents embedded in CI/CD pipelines for secure, explainable validation and deployment workflows, prioritising reproducibility and compliance requirements.
Schuftan closes out the blog piece by pointing out that they are still iteratively improving when they use the new AI-enabled tools, using them to find and fix bottlenecks and repetitive tasks.
The irony is: by introducing a bit of non-determinism in the right places, we’ve made our CI/CD process feel more stable, more predictable, and way more human-friendly.