Nov 10, 2025
AI Writes Code. People Ship Software.
AI has transformed how fast software gets written, not how fast it gets shipped. For executives weighing a legacy modernisation, understanding this dynamic is key. The technology that accelerates code generation does not, on its own, accelerate the delivery of working systems. The decisive work stays human, and it sits on either side of the code.
The productivity gains are real, but they leak
The scale of AI’s effect on coding is not in dispute. A 2026 study of more than 100,000 GitHub developers found that successive tool generations raised coding activity sharply: roughly 40 percent for autocomplete, about 140 percent once synchronous agents arrived, and about 180 percent with autonomous agents. The difficulty is what happens next. The gains attenuate as work moves toward release. Synchronous agents lifted lines of code by 741 percent and pull requests by 65 percent, yet releases rose only about 20 percent. The study puts the elasticity of substitution between AI output and human effort at about 0.25, the signature of a weak-link production chain: automate one stage and the human stages around it cap the gain.
Discovery is where modernisation value is won or lost
In legacy modernisation the binding constraint appears first upstream, before any code is generated. Source code records what a system does, not why. This tacit knowledge sits with the engineers and product staff who build and run the application: the business rules customers depend on, the exceptions earned over years, and the logic behind features that look strange until their history is known. Capturing this knowledge, and setting a target architecture matched to the skills of the people who will own it, is the highest-value work in the programme. AI can propose designs; people must judge them.
Review and integration decide what reaches users
The constraint returns downstream. Review and integration are where AI’s gains fall away most: the difference between writing code and shipping it. Generated code cannot replace the human work of validating behaviour, catching regressions, and merging changes into a system that still runs. If you produce code faster without expanding this capacity, the bottleneck simply moves to the people who must review it.
Three moves for modernisation leaders
Leaders who treat AI as the whole answer will accumulate code without delivering systems. Three moves keep the gains flowing to the finish line.
Invest in discovery first. Fund the capture of tacit knowledge and the design of an architecture your team can own, before code generation begins.
Build review and integration capacity. Scale the human stages that turn generated code into shipped software, rather than assuming AI removes the need for them.
Treat AI as one instrument, not the orchestra. Use it to accelerate the labour-intensive middle, and surround it with the human work that protects quality.
Frequently asked questions
Can AI coding tools modernise a legacy system on their own?
No. They accelerate one stage, writing code. Delivery still depends on human discovery, the capture of tacit knowledge and a workable architecture, and on human review and integration. Large coding gains shrink to far smaller gains in shipped software.
Why don’t AI’s coding gains translate into delivered software?
Because the production chain has a weak link. A 180 percent rise in coding activity corresponds to only about a 30 percent rise in releases; the human stages downstream compress the upstream gains.
Evidence cited: Mert Demirer (MIT), Leon Musolff (Wharton), and Liyuan Yang (MIT), “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools” (SSRN working paper, 2026). Statistics are drawn from the paper and referenced in the footnotes.
