Amazon is presenting one of the clearest internal examples yet of how an AI assistant can change routine software maintenance. CEO Andy Jassy says Amazon Q helped the company speed up Java upgrades that usually take development teams significant time but rarely get attention from users.
The headline claim is large: Amazon estimates the work saved about 4,500 developer years. The company also links the updates to better security, lower infrastructure costs, and estimated yearly efficiency gains of $260 million.
Why Java upgrades often get delayed
Jassy described core software updates as work that is both necessary and easy to postpone. They do not add visible new features. They usually do not create an obvious change in the user experience. For developers, that can make them feel less urgent than building something new.
That does not mean the work is optional. Updating core software keeps applications closer to current versions and can reduce technical drag over time. In Amazon's case, the task at the center of the example was moving Java applications to newer versions, including Java 17.
Before Amazon Q was applied to this work, Jassy said the average upgrade took about 50 developer days per application. With the new code transformation feature, he said the same kind of upgrade could be completed in just a few hours.
That difference matters because maintenance work compounds across a large software organization. A task that looks manageable for one application can become a major backlog when repeated across many production systems. Amazon's claim is that AI changed the economics of that backlog.
What Amazon Q Code Transformation does
Amazon Q Code Transformation is the feature Jassy highlighted for software development. According to the source article, it analyzes existing code, suggests changes, and implements them.
The work described is not limited to a simple version bump. Amazon Q Code Transformation can update package dependencies, revise outdated and inefficient code, and integrate security practices. That makes it relevant to the kind of maintenance that often requires careful review across multiple parts of an application.
In plain language, the tool is being used to move old application code toward a newer baseline. It looks at what exists, proposes the modifications needed, and carries out those changes. Developers still have to decide whether the generated work is acceptable, but the goal is to compress the time spent on repetitive upgrade steps.
The example also shows why code transformation is different from using AI only to generate new snippets. Here, the business value comes from changing existing production systems. That is usually more sensitive than starting from a blank file, because the code already supports running applications.
The numbers Amazon is reporting
Amazon says the impact became visible over six months. During that period, the company updated over half its Java production systems to newer versions much faster and with less effort.
Several figures stand out from Jassy's account:
- Average time to upgrade an app to Java 17 fell from about 50 developer days to a few hours.
- Amazon estimates the work saved about 4,500 developer years.
- Developers used 79 percent of Amazon Q's auto-generated code reviews without changes.
- Jassy said the updates helped improve security and lower infrastructure costs.
- Amazon estimates yearly efficiency gains of $260 million.
The 79 percent figure is especially important because it points to developer acceptance of AI-generated output. If most generated changes required heavy rewriting, the value would be lower. In Amazon's telling, developers used a large share of the generated code reviews as they were.
Still, the numbers should be read as Amazon's own estimates. Jassy is also promoting a product his company sells. That does not make the example meaningless, but it does mean the claims should be understood as coming from Amazon rather than from a neutral study.
What this says about AI in software work
Jassy sees the result as evidence that large companies can use AI to create major efficiency gains in core software maintenance. For Amazon, he called the result a "game changer." The reason is straightforward: maintenance work that once consumed large amounts of developer time was compressed into a much shorter workflow.
The case also highlights a practical area where AI may be easier to evaluate than more open-ended coding tasks. A Java upgrade has a clear target: the application must move to a newer version and continue to function. The work may be tedious, but the desired outcome is concrete.
That makes AI code transformation a useful test case. It is not only about whether an assistant can write code. It is about whether it can help teams clear work that they already know they need to do, but often delay because it competes with feature development.
At the same time, the source article notes that criticism of AI-generated code remains unresolved. Some critics argue that AI code can create more work than it solves. The long-term effect of AI on coding will be clearer only when neutral long-term studies are available.
For now, Amazon's example is a strong company-reported signal, not a final verdict. If the savings are close to what Jassy claims, the impact is substantial. Even if the real savings are much smaller, the scale of the reported improvement suggests why software teams are paying attention to AI-assisted maintenance.
The bigger takeaway
The Amazon Q example is not about replacing a glamorous part of software development. It is about compressing a necessary, repetitive, and often postponed task. That may be where AI assistants can show immediate value inside large codebases.
For organizations with many production systems, routine upgrades are not small chores. They affect security, infrastructure costs, and the amount of developer time available for other work. Amazon's claim is that Amazon Q changed that equation for Java upgrades at significant scale.
The unanswered question is how broadly that result applies beyond Amazon's own environment and Amazon's own reporting. The claim is still worth watching because it frames AI coding tools around measurable maintenance outcomes rather than vague productivity promises.