Generative artificial intelligence (AI) is lifting software development productivity, but as the amount of AI-written code grows, costs for review, fixes and maintenance can also rise, an analysis said. It said the key task ahead may be not how much code AI generates, but how little it generates and how much it reuses verified assets.

Gigazine reported on May 29 that Deepak Anupal (디팍 아누팔리), co-founder and chief technology officer at WaveMaker, recently pointed to limits of AI code-generation tools and said, "The problem is not how to inspect the generated code, but how little to generate in the first place."

He said AI coding tools are clearly improving productivity in repetitive development work. He said they can quickly generate standardised code such as input forms, data tables and basic structures for API integration, helping prototype development and internal business application builds.

But he pointed to the problem that as the volume of generated code rises, so does the amount that people must review and fix.

AI is also rapidly taking a larger share of code written in real development settings. A 2026 developer survey released by software quality platform Sonar found 42 percent of code merged into repositories was written with AI support. About 29 percent of that was merged straight away without manual human review.

That means coding is faster, but quality assurance systems are increasingly unlikely to keep pace. The general response companies use now is to verify AI-generated code after the fact. They look for issues using static analysis tools, linters, security scans, accessibility checks and visual comparison tests.

Anupal argued this approach is unlikely to be a fundamental solution. He said review is feasible when there are few applications, but once dozens of services are running at the same time, the volume of code to review grows exponentially. He said AI may speed up development, but the burden of review and maintenance ultimately consumes developers' time again.

He proposed an alternative he called an "assembly model." Instead of having AI write new code from scratch each time, it prioritises reusing already verified components and libraries. For example, if a user asks to build a customer list screen with a search function, AI would choose an in-house table component that is already verified and apply only the necessary settings, rather than generate new table code.

The key is to reduce the amount of code that must be written anew. Reusing components that have already passed security and accessibility checks, such as buttons, input forms, authentication screens and data lists, removes the need to repeat the same inspections for each application. AI handles only minimal work such as data connections or screen transitions, and generates new code only for essential parts such as business logic and external system integration.

The same principle can apply to backend systems. Anupal stressed that in areas where design errors have major impact, such as data storage structures, APIs, authentication systems and access rights management, structural control is more important than code generation. He said companies need to build in advance measures such as secret management, role-based access control (RBAC) and standardised API contracts.

He said this approach is not cost-free. AI needs additional contextual information and token usage to understand internal company rules and component structures.

But Anupal said total long-term costs could be lower than a method of continuously generating the full code and then repeating fixes, regeneration and security checks. He said above all, because the amount of code subject to review falls, the chance of critical bugs passing through review also decreases.

This approach could become even more important in highly regulated industries such as finance, healthcare and the public sector. He said it is more favourable for accountability to show a structure in which verified parts are used and only newly generated parts receive additional review, rather than proving that every piece of generated code was verified one by one.

The next stage of AI coding is increasingly likely to be optimising the scope of generation, not competing on generation capability. Rather than how much code AI can write, the ability for companies to decide first what to build anew and what to reuse is emerging as a new competitiveness that will determine development productivity and cost efficiency.