AI Coding: Before We Convert Speed Into Savings, We need Quality
AI coding tools are clearly increasing developer output. The management question is what that increase actually means for cost, quality and risk.
A recent Faros study covering 22,000 developers across 4,000 teams found that task completion rose by roughly one-third and code-related throughput increased sharply. At the same time, bugs per developer increased by 54%. Production incidents rose. More code was merged without review, and review times became longer. The conclusion is not that AI coding tools are ineffective. It is that faster coding does not automatically create a more productive technology organisation.
Please note: The views expressed on this website are my own and reflect my personal professional experience, research and publicly available information. They do not represent the views, strategies or confidential information of my employer or its affiliated companies.
The risk is moving downstream
Writing code is only one part of software delivery. Code still needs to be reviewed, tested, secured, integrated and operated. If AI accelerates development but the rest of the system remains unchanged, the bottleneck moves. Developers complete more work, while senior engineers, testers and operations teams absorb more review, correction and incident management.
The apparent productivity gain can therefore be offset by higher rework and operational cost. For management, the relevant measure is not code volume. It is whether the full delivery chain improves:
- shorter time from idea to stable production,
- fewer defects and incidents,
- less rework,
- lower total cost of change, and
- better customer outcomes.
Be careful with headcount assumptions
The obvious response to higher output per developer is to assume that fewer developers will be needed. That may be premature. The people most critical in an AI-enabled engineering organisation are often not those writing the most code. Senior engineers, architects and security specialists review design, identify dependencies and prevent weak solutions from reaching production.
AI may reduce effort in code creation while increasing the need for experienced review. If these roles are reduced too early, the saving may return later as incidents, security issues, customer disruption, technical debt and rehiring.
The management decision should therefore not be:
How many developers can we remove?
It should be:
Where should we reinvest the capacity AI creates?
The strongest uses are likely to be better architecture, more automation in testing, stronger security, improved resilience and faster product development.
What management should ask for
Before AI productivity assumptions are reflected in budgets or workforce plans, management should request a clear before-and-after view. The most important indicators are:
- incidents per production change,
- rework and rollback rates,
- review times,
- code reaching production without meaningful human review,
- total lead time to stable production, and
- where the review and correction burden is increasing.
This should be assessed at team and product level, not only through overall averages.
The real issue is the operating model
The tool is not the strategic differentiator. Most companies will have access to similar AI coding assistants. The differentiator will be the operating model around them. That means clear standards for AI-generated code, stronger automated testing, architecture guardrails, defined accountability and enough senior capacity to challenge what AI produces.
Adding more approval gates at the end will not be sufficient. Quality has to be built into the development process from the start.
The key takeaway
AI coding tools can create meaningful capacity. But that capacity should not automatically be translated into a headcount reduction. First prove that the whole delivery system is improving-not only the speed of code creation.
The most important management principle is simple:
Do not convert AI coding gains into savings until quality, delivery speed and total cost have improved end to end.
Sources:
https://www.faros.ai/research/ai-acceleration-whiplash