Claude, Gemini, and GPT-5 can handle every SDLC task. Almost none of them should.

Over the past couple of years, discussions surrounding AI in software development have primarily revolved around security concerns. These include ensuring AI-generated code remains secure, preventing intellectual property from leaving organizational boundaries, and managing prompt, model, and data access governance. While these are crucial considerations, they may not be the deciding factors for the success or failure of AI initiatives.

Instead, the real challenges are shifting from model outputs to system design when AI integrates into production. In other words, AI's impact on software delivery lies not in the generated code, but in the architectural layer it introduces. The risks posed by AI architecture, much like early cloud adoption, could lead to ballooning costs, increased governance complexity, and operational challenges.

Organizations may then consider repatriating workloads or adopting hybrid cloud strategies to regain control over these issues.

The future of AI in software development isn't centered around a single model, but rather an AI supply chain. Just like in software delivery, where various tasks span planning, coding, testing, validation, deployment, governance, compliance, and operations, AI-native software delivery will likely involve a network of AI capabilities.

These can include models that help define requirements and generate user stories, generate implementation logic, create unit and integration tests, review outputs against requirements, ensure policy and regulatory compliance, and monitor activity, costs, and model performance across the pipeline.

Large, general-purpose models like Claude, Gemini, and GPT-5, while highly capable, may not be the most efficient or cost-effective tool for every task in the software development lifecycle. In fact, most SDLC tasks might not require frontier intelligence. For instance, a specialized "R2-D2"-type model could be optimized for tasks like generating unit tests, code review, build validation, compliance checks, or even evaluating whether requirements have been met.

These specialized models could perform much of the work in-between, being cheaper to run, easier to fine-tune, and more predictable in their outputs. This does not mean frontier models will disappear. They might continue to be used at the beginning or end of the workflow, helping define requirements, map business objectives, or serve as an 'AI judge' to assess outputs, validate quality, or verify compliance.

However, most software delivery tasks could potentially be handled by a team of specialized AI systems, providing greater control, efficiency, and cost-effectiveness compared to relying on a single model.