AI Coding Seems to Have Diminishing Returns Closer to Production

AI Coding Seems to Have Diminishing Returns Closer to Production

• AI & Machine Learning
• September 15, 2026

A recent longitudinal study followed 802 developers and nearly 200,000 pull requests during a company-wide AI rollout. Overall throughput eventually reached 2.09x the pre-mandate baseline.

But the split between codebases is more interesting. Repositories created from 2022 onward saw a 44% increase in throughput. In legacy repositories created before 2022, the increase was only 12% and wasn’t statistically significant. The researchers don’t establish exactly why. But the result matches something we often see in practice.

In our view, one possible explanation is the difference in risk. AI is extremely useful when the risk appetite is high: prototyping, early R&D, exploring implementations, getting the first version working. You can generate and iterate quickly because the cost of a wrong turn is relatively low. As software matures, that equation changes.

Production code carries accumulated complexity and technical debt. Changes require more context, review and verification. And mature codebases are more likely to support business- or mission-critical workflows, where an AI-generated mistake can be far more expensive than during prototyping.

So our hypothesis is that the 44% vs 12% gap may reflect not only codebase age, but a different risk profile. Teams can afford to use AI aggressively in newer development. In mature systems, using it selectively and keeping more human control can be the rational choice.

This also points to a broader limitation of AI coding ROI: faster code generation doesn't automatically mean proportionally faster software delivery. As generation accelerates, review and integration can become the constraint.

It would be interesting to know how much of the gap comes from codebase complexity, and how much from the way teams use AI in more critical production environments.

At DataObrii, we help engineering teams understand where AI can actually improve delivery by looking at the workflows and data behind the rollout instead of treating adoption itself as the KPI.

Source: Hao He et al., “AI Writes Faster Than Humans Can Review: A Longitudinal Study of an Enterprise ‘2×’ Mandate,” arXiv, July 2026.

Image credit: Fotis Fotopoulos / Unsplash.