AI Writes Code Faster Than Most Engineering Organizations Can Ship It 

Generative AI has made writing code dramatically faster, but shipping that code into production remains a different story. 

Engineering teams are discovering that while AI coding assistants can generate pull requests in minutes, the processes surrounding review, testing, provisioning, governance and deployment haven’t kept pace. Rather than eliminating software delivery bottlenecks, AI has shifted them further down the pipeline. 

That shift is placing platform engineering squarely at the center of modern software delivery. Organizations with mature internal developer platforms, standardized self-service infrastructure and embedded governance are finding it easier to absorb rapidly growing code volumes without sacrificing reliability. 

“The center of gravity has moved from generation to review,” says Adam Resnick, research manager for modern software development and developer trends at IDC. “Developers are not simply trusting AI more as it writes more code; they are reviewing more of it, making oversight the real constraint on how fast AI-assisted development can scale.”  

IDC’s forthcoming 2026 research illustrates how quickly AI has changed developer workflows.  

While the company’s 2025 survey found that most developers generated a quarter or less of their code with AI, the new data shows the opposite: 60% now generate more than a quarter of their code using AI. Yet the percentage of developers accepting AI-generated code without revision has fallen sharply, meaning review workloads continue to grow even as coding accelerates.  

Code Generation No Longer the Problem  

The industry has spent years trying to help developers write software faster. AI has largely solved that challenge–organizations now face a verification problem. 

“AI has dramatically shifted the primary software delivery bottleneck from writing code to trusting it,” says Itamar Friedman, CEO and co-founder of Qodo. “A developer can now generate a feature, migration, or test suite in minutes with tools like Cursor or Claude Code, but the hard part is knowing whether or not the code fits the architecture, respects the team’s rules, and won’t break something three repos away.”  

Manual code review has become increasingly difficult because AI-generated code often appears polished even when underlying architectural or security issues remain. 

According to Friedman, reviewers are now expected to evaluate larger volumes of code while simultaneously identifying subtle problems such as dependency conflicts, contract violations and architectural inconsistencies. The result is that review, testing and verification—not coding itself—have become the primary constraints on delivery throughput.  

Resnick notes this compounds an existing imbalance. Even before widespread AI adoption, developers spent considerably more time testing, debugging and performing quality assurance than writing new code. AI has accelerated only one half of that equation.  

Platform Engineering as Throughput Layer 

As code volume increases, platform engineering is increasingly determining whether organizations can translate developer productivity into production software. For many organizations, that starts with standardized self-service infrastructure. 

IDC’s research found that 81% of organizations are expanding, using or piloting internal developer platforms. 

Resnick says the greatest benefit comes from governed provisioning and “golden paths” that make secure, compliant deployments the default rather than requiring developers to navigate additional approval processes.  

Yasmin Rajabi, COO at CloudBolt, argues infrastructure automation has become essential rather than optional. 

“Self-service provisioning with golden paths, because if getting an environment is a multi-day ticket, none of the upstream speed matters,” she says. “You want a vending machine with opinionated defaults, not one-off infra per team.”  

Rajabi also points to merge queues, automated testing, embedded security scanning and centralized control planes with role-based access control as foundational capabilities for organizations increasingly relying on AI-generated software. 

These controls help prevent situations in which independently generated code changes conflict before reaching production.  

Friedman says he believes another critical capability is providing AI systems with organizational context rather than relying on generic models. 

“The biggest impact comes from giving AI review systems real organizational context,” he says. “The goal is not just to generate code faster; it is to build a verification layer that understands how this company actually ships software.”  

Governance Moves Into the Platform 

The rapid growth of AI-generated code is also changing how organizations think about governance. 

Rather than treating governance as a separate approval process, many are embedding policy directly into delivery platforms through automated verification, security controls and standardized workflows. 

Rajabi argues that organizations should stop viewing speed and governance as competing priorities. 

“Stop treating it as a balance, because framing speed and governance as a tradeoff is how you end up with neither,” she says. “The teams that get this right make the governed path the fastest path, so nobody has a reason to go around it.”  

Friedman echoes that view, recommending organizations separate code generation from code verification. 

“The same AI system that wrote the code should not be the only system deciding whether the code is safe,” he says. “Developers should spend less time policing every line and more time improving the rules, context, and guardrails that make the whole system trustworthy.”  

Ultimately, faster code generation alone does not produce faster software delivery. The organizations realizing the greatest gains are those investing just as heavily in the platforms, governance and verification processes that surround AI-generated code. 

“Review, not generation, is now the constraint on AI-assisted development,” Resnick says.