In mid 2025, the industry was shaken when industry leaders like CEO Satya Nadella disclosed that AI generates 20 to 30% of Microsoft’s code. Similar messages came from other industry leaders. Fast forward to Feburary 2026, we have the Spotify leadership disclosing that none of their senior developers have manually written a single line of code since December 2025. Quite a jump from 20-25% to 100%. The velocity of development has changed dramatically and so has been the case of quality engineering. Microsoft’s recent appointment of Charlie Bell as its first Engineering Quality Head makes this tension explicit. The sequence reflects a pattern emerging across the industry: teams adopted AI coding assistants for speed and are now redesigning how quality needs to be managed. This is especially true in large enterprises.   

At QualityKiosk, our view is that as more and more code is written by AI, investment in Quality Engineering must extend beyond the application that being developed. In large non-software product companies, application landscape is quite heterogenous. It’s a mix of custom developed, closely integrated with Commercial Off the Shelf (CoTS) product which in turn integrates with SaaS solutions and more. AI does a fantastic job of writing code, however, its context window is limited to the repo or application being developed. It needs experts to validate seamless integration within its ecosystem and factor domain constraints. AI rarely understands enterprise architecture, cross-system dependencies, or the business context that governs how software should behave end-to-end. Without filling this gap, reliability, security, and maintainability risks accumulate over time.  

 What the Data Shows 

Uplevel studied nearly 800 developers in 2024. Those using GitHub Copilot introduced 41% more bugs compared to their previous performance. 

GitClear analyzed 211 million changed lines of code across 2020 to 2024. Duplicated code blocks increased eight-fold during 2024, with duplication signals measured at roughly 10 times higher than two years earlier. 

Google’s DORA research also highlights the speed-versus-stability trade-off. It reports a ~3.1% increase in code review speed, yet an estimated 7.2% reduction in delivery stability for every 25% increase in AI adoption. This means faster reviews, but less predictable systems. 

The Quality Challenge 

AI coding assistants can produce syntactically correct code quickly. But they often miss system-wide architecture, cross-service dependencies, domain constraints, and compliance requirements, unless teams provide guardrails and context. This is where quality engineering must scale alongside coding velocity. 

The trend we have observed in recent past are:  

• Dependency complexity: AI optimizes for the local function while missing downstream dependencies. Authentication flows work in isolation but break session management elsewhere. Data validation passes unit tests yet violates domain constraints in production. 
• Self-reinforcing blind spots: When AI generates both code and tests, it can end up validating its own assumptions. Tests cover happy paths while missing edge cases, concurrency issues, and adversarial inputs where AI-generated logic proves most fragile. 
• Missing know-how – Since the code has been written by AI, humans have difficulty explaining the logic in defect triage leading to lower quality root cause analysis and pattern learnings.
• Missing integration – Ability to integrate with existing process, methodologies or frameworks.
• Code consistency – every matured enterprise has a set of coding standards. It gets compromised when code is written through AI.

What Changes in Practice 

Quality engineering as a practice is rapidly evolving to address these new risks. It means setting architectural constraints and test strategies up front. It continues during generation, with instrumentation tracking duplication, churn, and dependency risk. It extends after generation, with independent test suites designed to challenge AI outputs. 

Traditional QA treated testing as a validation step after development. Teams wrote code, then tested it. With AI generating thousands of lines of code per week, this sequential model breaks down. The review surface expands faster than review capacity.  

Organizations adapting to this trend successfully maintain independent quality streams. They track metrics specific to AI-generated code: duplication rates, code churn within two-week windows, and defect escape rates by generation source. These signals expose accumulating technical debt before it reaches production.  

Generating code costs almost nothing. But maintaining non-contextual code costs the same as it always did – it is high. Over time, these costs compound. 

What This Means for You 

Microsoft’s move signals recognition of this reality. Code generation has accelerated. Quality practices must match that pace, or the gap between generation speed and validation capacity becomes an operational risk. 

If you already run AI-assisted development in production, the practical question is: have your quality practices evolved to address the new risk profile? The answer determines whether AI coding assistants deliver durable short-term value or create a growing long-term maintenance burden. 

The companies that succeed with AI coding agents tend to share a few traits. They invest in quality engineering with domain expertise. They build automated quality gates that operate at the system level, beyond individual functions. They act on quality signals even when velocity metrics look strong.  

AI changes how code gets written. It also changes how code must be tested. Microsoft’s shift acknowledges that reality, and the rest of the industry is moving the same way.