Table of contents
Generative AI was expected to eliminate talent shortages and dramatically accelerate technology projects. Instead, companies have encountered a phenomenon that can be described as the delivery paradox — the volume of code in repositories is growing rapidly, but the time required to deploy production-ready systems does not automatically decrease. Organizations are falling into the trap of verification debt, where reviewing and correcting code becomes the new bottleneck. In the era of generative artificial intelligence, the real challenge is therefore not the technology itself, but the economics of the entire software development process.
Code Was Supposed to Solve the Biggest Challenge in IT Projects
Until recently, one of the key constraints in technology projects was the availability and productivity of software developers. The speed at which an organization could deliver a new feature, digital service, or application depended on their expertise and delivery capacity. The first wave of generative AI adoption in IT was therefore built on an intuitive promise: automating code generation would deliver an immediate breakthrough.
Early data appeared to confirm these expectations. In an experiment conducted by GitHub, developers using Copilot completed a specific programming task on average 55% faster than those working without its support.[1] Businesses therefore anticipated a dramatic acceleration of the entire software delivery process. If a machine can generate code within seconds, the journey from idea to working application should become significantly shorter.
However, it quickly became apparent that accelerating one stage of the process does not necessarily improve the efficiency of the entire delivery pipeline.
More Code, But No Automatic Acceleration of Delivery
According to a Sonar study conducted among more than 1,100 professional developers, code generated or significantly assisted by AI already accounts for an average of 42% of the code committed to repositories. Respondents predict that this figure will reach approximately 65% by 2027.[2]
However, the growing volume of code has introduced a new bottleneck. As many as 38% of respondents believe that reviewing AI-assisted code requires more effort than evaluating code written entirely by humans. At the same time, 96% of respondents do not fully trust AI-generated code.[2]
The industry is increasingly referring to this phenomenon as verification debt — a situation in which the time and cost of reviewing, validating, and correcting code begin to consume a significant share of the benefits gained from generating it faster.
The widespread adoption of so-called vibe coding means that IT departments must deal with growing streams of code requiring testing, validation, security reviews, and integration with existing systems. As a result, the bottleneck does not disappear. It simply shifts from the coding stage to later phases of the software lifecycle.
After all, code itself was never the sole constraint in the process. It was merely its most visible component.
The New Economics of Software Development
Escaping this deadlock requires abandoning the view of artificial intelligence as merely a digital assistant for developers. AI is fundamentally changing the economics of the entire software lifecycle — the Software Development Life Cycle.
In the traditional model, the cost of creating code was directly linked to an engineer’s time and specialized expertise. Today, large language models make it possible to scale code-generation capabilities to an extent previously unavailable. The cost of execution itself — the technical act of translating requirements into code — is decreasing.
As code generation becomes increasingly accessible, the center of gravity shifts toward the beginning of the process — business analysis, architecture design, and the precise definition of business requirements.
“If AI generates vast amounts of code, someone still has to validate that code. And the less clearly the requirements are defined, the more difficult and expensive that validation becomes,” emphasizes Karol Drążek, Head of the IT Modern Architecture business line at MINDBOX.
This shift requires a new discipline of work in which specification-based approaches, such as Specification-Driven Development, become increasingly important. Documentation ceases to be merely a formal project artifact and becomes the primary source of context for AI systems.
The principle of garbage in, garbage out takes on particular importance in this environment. If business requirements are incomplete, inconsistent, or ambiguous, AI can rapidly generate software that appears correct but fails to solve the organization’s actual problem.
The Developer as an Orchestrator, Architect, and Auditor
In the new model, traditional software craftsmanship is gradually giving way to technology orchestration. The role of the experienced engineer is evolving from that of a direct code creator toward that of a quality auditor, security guardian, and system architect.
“The most important capabilities become systems thinking and the ability to break down complex business problems into smaller components in order to manage the context provided to AI models with precision. Knowledge of programming language syntax alone is no longer sufficient,” says Łukasz Ewertowski, AI Transformation Leader at MINDBOX.
Increasing value is being created by individuals who can properly define a problem, break it down into specific tasks, delegate those tasks to AI tools, and then assess whether the generated components form a coherent, secure, and scalable solution.
This change is also affecting the development paths of junior professionals. Companies are becoming increasingly cautious about hiring entry-level programmers because simple and repetitive tasks are currently the easiest to automate, at least partially.
However, completely limiting investment in junior talent would be short-sighted. Unless the industry redesigns career paths and engages younger engineers in emerging roles — such as AI tool operators, context engineers, or first-line validators — a significant skills gap may emerge over the longer term.
Experienced architects and technology leaders do not appear out of thin air. Organizations must create environments in which the next generation of engineers can develop expertise not only in coding, but also in architecture, quality, security, and accountability for the solutions they create.
Three Implications for Business Strategy
The transformation of the IT operating model has concrete implications for executive teams and technology leaders:
- Build or Buy? Until now, companies have often preferred subscription-based software solutions to avoid the cost of developing and maintaining their own systems. AI, however, is lowering the barrier to rapidly creating prototypes, dedicated microservices, and applications that support specific business processes. As a result, building custom solutions may become a viable alternative for some organizations. This does not mean that every system should be developed from scratch. Off-the-shelf solutions may still be cheaper, safer, and easier to maintain. The key question is no longer: “Can we build it?” but rather: “Should we build it, given that AI allows us to do so much faster?”
- AI Cost Control: A lower barrier to entering the development process does not automatically translate into lower overall costs. Poor prompt quality, insufficient context, and repeated inefficient attempts at code generation can quickly erode the anticipated savings. In some cases, the cost of repeated iterations and resource consumption may make AI-assisted development more expensive than traditional approaches. Effective management of AI budgets is therefore becoming a new management discipline. Just as the rise of cloud computing led to the emergence of FinOps practices, the widespread adoption of artificial intelligence is creating demand for AI FinOps capabilities focused on monitoring model costs, token consumption, iteration volumes, and selecting the right technology for specific tasks.
- Changing the Operating Model, Not Just Buying Tools: Simply providing coding assistants will not transform the software delivery process. Such tools may reduce the time required to complete certain tasks, but they do not solve issues related to unclear requirements, architectural deficiencies, insufficient validation, or a lack of accountability. Meaningful gains are likely to emerge only after reshaping the operating model itself — how requirements are defined, teams are structured, validation is performed, accountability is assigned, and value delivery is measured. Rather than focusing exclusively on development speed, organizations should analyze the entire journey from a business need to a secure, functioning solution available to end users.
Six Questions Every CIO Should Ask
Code is becoming increasingly accessible, but decisions regarding what code should be created, how its quality should be maintained, and what business value it should deliver require greater precision than ever before.
Technology leaders should therefore not focus solely on which AI tools to deploy within development teams. Organizations that gain a competitive advantage will be those capable of answering six critical questions:
- Can we define business problems precisely enough to provide valuable context for artificial intelligence?
- Do we measure the effectiveness of the entire value-delivery process—from identifying a business need to deployment—or only developer productivity?
- Do we have clear and transparent accountability for code created with AI assistance?
- Can we effectively control the costs of tokens, additional iterations, and model queries?
- Do we know when it is strategically beneficial to build our own solutions and when it is safer and more cost-effective to rely on existing market offerings?
- Are we developing people for new roles—architects, auditors, and context management specialists—instead of simply automating existing tasks?
According to Deloitte, only 15% of organizations using generative AI currently report significant and measurable returns on investment, although another 38% expect to achieve them within the next year.[3] Organizations leading in terms of ROI distinguish themselves by treating AI as a business transformation enabler rather than merely a technological enhancement.
Success in a world where code is no longer the primary engineering constraint depends on how effectively an organization manages the broader system of decisions, costs, and accountability surrounding it. In the era of generative AI, the winners will not be the companies that produce the most code. They will be the organizations that best understand why they are creating it and what value it is intended to deliver.
References:
[1] GitHub Research: The Economic Impact of the AI-Powered Developer Lifecycle (GitHub Copilot productivity study).
[2] Sonar Global Report 2025/2026: Survey of more than 1,100 professional developers on AI adoption and the challenges associated with machine-generated code.
[3] Deloitte GenAI EMEA Report: Analysis of return on investment (ROI) levels and GenAI and Agentic AI implementation strategies among 1,854 technology leaders.
Authors
Łukasz Ewertowski
AI Business Transformation & Technology Director, Head of Microsoft Centre of Excellence
Karol Drążek
Director of Modern IT Architecture COE, Head of AWS Competence Centre
