You can write code faster. Can you deliver it faster?

Summary

At DPE Summit 2025 Hans Dockter argues that GenAI’s value isn’t more code but more working software — yet DORA data shows AI can increase developer productivity while reducing delivery throughput and stability. He compares AI mispredictions to unavoidable biological mutations, arguing errors are inherent and must be engineered around rather than wished away. Strengthening the software development lifecycle — tests, faster MTTR, human-maintainable code, supply-chain defenses — is essential. AI should bolster continuous delivery capabilities, not replace human understanding; otherwise organizations risk shipping more defective software. Trisha Gee closes by urging practical GenAI practices to accelerate delivery without sacrificing reliability.

Takeaways

• 🚀 The central question for AI in software development is not how much more code can be produced, but how much more working software can be delivered to customers.
• 📉 Early DORA research shows that while developers feel more productive using GenAI, actual delivery throughput drops by about 1.5% for every 25% increase in AI adoption, and stability drops even more sharply by 7%.
• 💡 Hans Dockter emphasizes that organizations may currently be investing more in AI than they are getting out, making it essential to understand the full impact of AI on the software development lifecycle.
• 🧬 Dockter draws an analogy between AI and biology, comparing software failures to DNA mutations—both are unavoidable and must be managed, not eliminated.
• 🧠 Large language model 'hallucinations' are better understood as 'mispredictions'—a fundamental property of probabilistic systems that cannot be fully removed, only mitigated.
• 🔄 Increasing AI-generated code accelerates production but also raises the absolute number of defects and regressions, mirroring how higher replication rates in biology increase mutation frequency.
• 🛡️ Continuous Delivery (CD) capabilities act as the immune system of software, filtering out defects and ensuring stability; strengthening them is vital as AI amplifies the forces of software engineering.
• ⚙️ The fundamental forces of software engineering—mistakes, side effects, vulnerability, and constant change—remain unchanged even in the era of AI, but their magnitude increases.
• 🧰 Modern Continuous Delivery practices, like faster mean time to repair (MTTR), robust testing, and maintainable code, become more critical as AI scales development complexity.
• 🔍 Dockter and Martin Odersky warn against treating AI-generated code as a black box; human understanding of code remains essential for debugging, maintenance, and sustainable software delivery.
• 🐘 The cancer analogy extends to evolution—elephants evolved ultra-sensitive defenses against cancer; similarly, software systems must evolve defenses (strong pipelines, testing, governance) to survive AI-driven complexity.
• 🎯 The takeaway: AI should not replace fundamental engineering practices but strengthen them—organizations must adapt their toolkits and processes to handle the intensified ‘forces’ of modern software delivery.

Q & A

What is the key concern when it comes to AI's role in software delivery, as discussed by Hans Dockter?

The key concern is not how much more code AI can produce, but how much more working software can be delivered to customers. Despite increased productivity, the actual delivery throughput and stability can decrease as AI adoption rises.

How does the first DORA report on GenAI impact our understanding of AI in software delivery?

The DORA report reveals that while AI can make developers feel more productive, it can lead to a drop in delivery throughput (about 1.5% for every 25% increase in AI adoption) and a larger decrease in stability (around 7%). Developers may feel faster, but organizations often deliver less value.

How does Hans Dockter compare AI's impact on software delivery to cancer research?

Dockter compares AI’s impact to the historical understanding of cancer. Just like cancer research transitioned from speculation to a system-level understanding, we need to understand the boundaries and risks of AI in software development. This insight can guide us toward solutions, much like cancer research has led to more effective treatments.

What does Hans mean by AI coding assistants creating 'mispredictions' rather than 'hallucinations'?

Dockter suggests that the term 'hallucinations' is misleading because it implies a mysterious or pathological issue. Instead, AI's errors should be understood as 'mispredictions' that naturally occur due to how large language models are trained and evaluated. These errors are inevitable and a part of the system's behavior.

What is the relationship between AI-generated code and defects, according to Dockter?

As AI helps generate more code faster, the absolute number of defects and regressions also increases. This is similar to how, in biology, faster replication results in more mutations. The challenge is managing these defects within a robust software delivery pipeline (SDLC).

Why does Dockter emphasize the importance of strengthening the SDLC pipeline when using AI in software delivery?

Dockter stresses that the SDLC pipeline is crucial for controlling the defects and regressions introduced by AI. Without strengthening this pipeline, AI’s increased throughput could destabilize the entire delivery process, reducing both speed and stability.

What analogy does Hans use to explain the inevitability of mistakes in software development and AI?

Hans compares AI-generated software to human biology, explaining that just as mutations are inevitable in cell replication, mistakes in AI-generated code are inevitable. The solution is not to eliminate mistakes, but to design systems that can handle and correct them efficiently.

How does Hans Dockter propose we address AI’s limitations in software development?

Dockter suggests that we must design around the permanent limitations of AI, particularly its tendency to produce mispredictions. He advocates for maintaining strong, human-readable code and ensuring that troubleshooting remains efficient despite AI's involvement in software creation.

What is Hans’ stance on the idea that AI will soon replace human understanding of production code?

Hans rejects the idea that production code will soon become a black box created solely by AI. He argues that understanding code is essential for troubleshooting, and without this understanding, any failure of the AI system could lead to total breakdowns, particularly in complex enterprise software.

How does Dockter view the future of cancer treatment and its relation to AI in software?

Dockter compares the predictable path to successful cancer treatment with the future of AI in software development. Just as understanding cancer's fundamental forces has led to effective treatment strategies, a deeper understanding of AI's inherent limitations will allow us to improve software delivery and develop reliable solutions.