AI Coding Assistants Haven’t Sped up Delivery Because Coding Was Never the Bottleneck

2026-03-24 · Source: InfoQ · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, quick

Summary

Agoda's recent observations, supported by Faros AI research, indicate that while AI coding assistants significantly boost individual developer output, overall project delivery speed has not increased proportionally. This is because the bottleneck has shifted from coding to upstream activities like specification and verification, which demand human judgment. Faros AI data from over 10,000 developers across 1,255 teams shows that high AI adoption led to 21% more tasks completed and 98% more pull requests merged, but PR review time simultaneously increased by 91%. This shift necessitates a re-evaluation of traditional engineering team structures, moving from minimizing communication to prioritizing shared understanding and collaborative specification as the primary value-creating activities. Agoda engineer Leonardo Stern proposes a "grey box" approach for engineers interacting with AI-generated code, focusing on precise specification and evidence-based verification rather than line-by-line code review.

Key takeaway

For engineering leaders evaluating AI coding assistant adoption, recognize that individual developer velocity gains may not translate directly to project-level acceleration. Your teams should shift focus from code production to robust specification and rigorous, evidence-based verification. This implies restructuring teams to foster deeper collaborative understanding, as communication becomes the core work, not an overhead to minimize. Prioritize investing in tools and processes that enhance specification clarity and automated testing over those solely boosting code generation speed.

Key insights

AI coding tools shift bottlenecks from coding to specification and verification, demanding new team structures and workflows.

Principles

• Coding is rarely the primary bottleneck.
• Communication is essential for shared understanding.
• Engineers remain accountable for AI-generated code.

Method

Adopt a "grey box" approach: engineers write precise specifications and verify AI-generated results against evidence, rather than inspecting code line-by-line.

In practice

• Prioritize high-fidelity specifications.
• Focus on evidence-based verification.
• Structure teams for shared understanding.

Topics

• AI Coding Assistants
• Software Development Bottlenecks
• Software Engineering Workflow
• Spec-Driven Development
• Team Structure

Best for: CTO, VP of Engineering/Data, Director of AI/ML, Software Engineer, AI Engineer, Machine Learning Engineer

Related on AIssential

• Coding Agents Meet Data Science — AI coding agents boost developer productivity but introduce new bottlenecks and require specialized adaptation for data science workflows.
• Slow down to speed up — AI amplifies the cost of flawed initial decisions, making deliberate "slowness" in planning more critical for true speed.
• AI Made Your Engineers 10x Faster and Your Product 10x Worse — AI coding assistants boost speed but degrade product quality, necessitating a rebuilt quality layer.
• If AI was actually killing software engineering, why is there more code than ever? — AI shifts software engineering bottlenecks from code generation to validation, architecture, and complex system judgment.
• The Final Bottleneck — Accelerated code generation creates new bottlenecks in review and integration, leading to unsustainable backlogs.
• The 3 AI Skills That Will Matter More Than Coding in 2026 — Future AI engineering prioritizes judgment, architecture, and supervision over raw coding ability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by InfoQ.