Microsoft Says AI Agent Coding Cuts Faults 20%, Support Cases 25%, Escalations 30%

Microsoft said artificial intelligence's role in software development is no longer limited to writing code, as the company seeks to fundamentally reshape how engineering teams build software. Charles Lamanna, executive vice president of Microsoft's Copilot, Agent and Platform business, outlined how AI is being deeply integrated into development workflows while maintaining a focus on customer success metrics.

Lamanna stressed that the team's biggest milestone in 2026 is not the launch of a flagship product or feature, but the complete transformation of software construction through emerging AI techniques such as agentic coding. In this model, software engineers no longer write code line by line; instead, they guide AI agent systems to carry out development work, allowing them to focus on strategic planning, architecture design, and quality control. Product development is no longer constrained by coding ability; the bottleneck shifts to making the right decisions to build compliant products.

Lamanna said that although AI gives engineers extra time, he has guided the team to keep the number of features developed unchanged, instead using the saved time to focus on software reliability, performance, and security. This approach has produced notable results, such as a double-digit year-over-year decline in service tickets related to Power Automate. Internal engineering team data shows that after adopting agentic coding, fault incidents fell by about 20%, support cases by about 25%, and critical escalations by about 30%, while accessibility work progressed significantly faster.

Lamanna concluded that engineering teams should reshape their work habits around AI, and successful transformation teams serve as role models. Based on this approach, Microsoft itself has shortened the cycle from product concept to mass market to a few months. However, Lamanna also acknowledged that Microsoft has not solved all problems, and how to more efficiently address security vulnerabilities within a large development system remains an ongoing challenge.