Mistaking Coding Speed for Delivery Speed - AI Productivity Myths

Headline: Mistaking Coding Speed for Delivery Speed - AI Productivity Myths

AI Productivity Myths: Lessons from the Real World #2 Why coding is not the bottleneck — and why speeding it up barely moves the needle on actual delivery.

A few weeks ago, I attended a company social event. The atmosphere was informal, with people mingling and sharing updates, personal stories, and weekend plans. However, the conversation during this event was dominated by AI talk, focusing on personal experiences and applications of AI in their work. None of the discussions touched upon the complex domain I work in daily or the upstream dependencies that were crucial to the team's success.

This experience made me reflect on the recent conversations I observed, which seemed to revolve around AI tools without questioning whether they were addressing the right problem. In software delivery, the real issue is not typing code faster, but rather aligning with the right problem. Most software engineers would likely agree with this perspective if asked directly.

To understand the true bottlenecks in software delivery, I asked engineering managers what percentage of their delivery time is spent writing new code. The responses ranged from 10% to 20%, sometimes even lower. The remaining time is consumed by various non-coding activities, such as aligning with shifting requirements, debugging production systems, coordinating with other teams, navigating unfamiliar codebases, and addressing security issues.

Research from GitHub supports this notion, showing that developers spend as much time waiting for builds and tests as they do writing new code. Additionally, DORA's report on the ROI of AI-assisted software development found that AI coding tools deliver significant gains only in simple tasks and for legacy code, with gains shrinking in areas where the real work lies.

The gains reported from AI coding benchmarks are not representative of real-world deliveries, which involve complex stakeholder interactions and organizational challenges.

The most impactful engineering activity I have witnessed occurred during a forty-minute conversation with a product owner, preventing the team from building the wrong thing. Sprint velocity remained flat during this activity, yet it saved days or weeks of wasted effort. The majority of projects' challenges lie in aligning priorities among team members and making informed trade-off decisions. These activities require judgment, trust, and interpersonal skills that AI coding tools cannot provide.

AI coding assistants are incredibly proficient at answering "how do I implement X?" However, they fail to help teams determine whether X is the right thing to implement. This critical question necessitates human context, including business strategy, regulatory constraints, and accumulated institutional knowledge. AI cannot grasp these aspects, which live in the minds of people and are often not documented.

When AI coding assistants became part of my daily routine, the time saved was mostly spent on reviewing generated code and correcting suggestions that missed the mark. The time investment did not lead to improved requirements gathering or deeper thinking. Instead, it consumed time that could have been allocated to other critical tasks, such as aligning with stakeholders, debugging, and coordinating with other teams.

The AI coding productivity myth suggests that speeding up coding will significantly impact delivery. However, the reality is far more nuanced. AI tools can assist in finding configuration issues faster, but the real value lies in the human expertise required to understand the broader context and make informed decisions. In software delivery, the bottleneck is not coding speed, but rather aligning priorities, coordinating with stakeholders, and navigating complex organizational dynamics.