Frequently Asked Questions
Why could the decrease in code-writing costs via AI paradoxically increase the total cost of ownership for IT systems?
AI radically lowers the cost of generating code, but it does not reduce the costs of subsequently understanding, testing, and updating it. This leads to the cheap production of a vast number of future liabilities and an increase in technical debt, which ultimately increases the total cost of system maintenance.
What impact does the mass generation of code by AI have on the long-term quality and maintainability of software?
Mass code generation by AI can lead to an increase in duplication (copy/paste), a decline in refactoring, and a higher number of bugs and so-called code smells. This creates long-term maintenance liabilities, as code that passes functional tests often remains difficult to maintain, unreadable, or inefficient.
Why is entrusting AI with both writing the code and testing it dangerous for system quality?
Entrusting both tasks to AI leads to a common cognitive bias where tests share the same flawed assumptions as the code; instead of verifying it, they become its 'advocate'. Flawed code provided as context significantly lowers the quality of generated tests and their effectiveness in detecting defects.
How should the code review process and technical debt management change when AI drastically increases the supply of code?
The code review process should evolve toward detecting 'dead correctness' and redundant abstractions, utilizing diverse criticism from multiple AI agents under human supervision. To counteract the increased supply of code, smaller change sets should be used, and a technical debt budget should be introduced, making the development of new features dependent on system quality and refactoring.
Why can the mere acceleration of code writing by AI be dangerous for an organization, and how should the real success of this transformation be measured?
The mere acceleration of code writing can be dangerous because AI can generate technical debt and create redundant components faster than ever before. The real success of the transformation should be measured not by the number of completed tasks, but by the balance of added versus removed code, as well as changes in system complexity and duplication.
How does the decrease in the cost of generating code via AI affect a company's decision on whether to build its own software or buy off-the-shelf solutions?
The reduction in code generation costs via AI lowers the economic barrier to building proprietary software, making this option more attractive for simple applications or those that strategically differentiate the company. At the same time, this decision no longer depends solely on the cost of writing the system, but on the costs of maintenance, security, and regulatory compliance, which—in the case of critical systems—may still favor purchasing ready-made solutions.
Does the ease of AI code generation mean that anyone can now create secure and professional software?
No, because removing the implementation barrier does not mean removing the professionalism barrier. AI democratizes the initiation of software production, but it does not replace competencies in security, authorization, or data management, which are essential when the consequences of errors are high.
Why do traditional knowledge management and documentation become crucial for organizational security in agentic systems?
In agentic systems, documentation becomes direct operational fuel, meaning that incorrect or outdated information can be applied by multiple agents with machine-like and perfect consistency. This necessitates the introduction of rigorous knowledge management and source hierarchies (provenance) to avoid compliance risks and contextual errors.
Why does access to advanced AI models alone not provide a company with a sustainable competitive advantage?
Access to AI models is becoming a common market resource that any competitor can acquire. Therefore, sustainable advantage stems not from the tool itself, but from how it is utilized: data, integrations, context, as well as the organization's technical and cultural practices.
How is AI changing organizational structure and the relationship between business and IT to avoid decision paralysis?
AI is transforming the business-IT relationship, shifting the role of IT from a feature factory toward the creator of a secure platform and standards (the so-called 'paved road'), within which domain experts can build solutions independently. To avoid decision paralysis, organizations are moving to a federated model: centralizing security principles and infrastructure (the 'constitution') while decentralizing initiative and execution.
How will the decrease in code generation costs affect the functioning and flexibility of internal company processes?
Reducing the cost of code generation will allow companies to move from rigid processes to flexible systems that keep pace with the evolution of the business model instead of limiting it. This will increase organizational agency, but simultaneously shift the center of gravity from technical execution to strategic judgment and decision-making wisdom.
Who bears the legal and ethical responsibility for decisions made by agentic systems within an organization?
Legal and ethical responsibility rests with the people and organizations acting as providers, deployers, data controllers, processors, or governing bodies. The machine is merely an executor of actions and lacks legal personality; therefore, it cannot be a place where accountability for decisions disappears.
What legal and regulatory requirements define the role of humans and data management in AI systems?
The human role is defined by actual human oversight, which enables system control, understanding of its limitations, and the ability to reject a result or stop the AI's operation. Regarding data management, GDPR principles are key, including data minimization, purpose limitation, and the prohibition of processing special categories of data without a statutory basis.
Who is responsible for errors and damages caused by autonomous AI systems within an organization?
Liability for damages does not rest with the AI system but is built around the economic entities designing, introducing, integrating, or utilizing the product. Within an organization, responsibility should be assigned to specific roles and decision-makers, e.g., those who granted permissions to tools, designed the policy layer, or approved the configuration for production.
What real threats related to biases in AI code does the law recognize, and what requirements does the AI Act introduce in this regard?
A real threat associated with AI bias is discrimination, particularly in the areas of employment and access to credit. Consequently, the AI Act classifies such systems as high-risk and introduces requirements regarding data quality, documentation, traceability, human oversight, robustness, and monitoring.
Is the compliance of automation with procedures and law sufficient to consider it safe and right?
No, compliance with law and procedures is not enough, because legal systems can generate decisions that are business-wise, socially, or morally wrong. The law only sets minimums; therefore, the organization's axiology and ethical reflection on whether a given automation is wise and consistent with the institution's identity are essential.
How should an organization move from the AI experimentation phase to a secure and mature operating model?
An organization should transition from experiment to operating model by restructuring institutional frameworks, rather than merely scaling tools. This process requires implementing a management system based on Cassanese's four frameworks (RACM, PAIP, ECF/IDVI, and RAUC), which allow for mapping AI capabilities, determining the level of task delegation, organizing human-machine collaboration, and ensuring security and legal compliance.
Why does simply purchasing the best AI model not guarantee an increase in organizational efficiency, and how should one approach the implementation of these tools?
The mere adoption of tools does not guarantee success because AI adoption is a systemic issue rather than a procurement one; this technology only amplifies existing organizational characteristics. To increase efficiency, one must first identify real business problems and ensure organizational capabilities, such as a healthy data ecosystem, clear AI usage policies, and mature operational practices.
What is organizational maturity in the context of AI implementation and how can it be achieved?
Organizational maturity in the context of AI is the ability to accurately calibrate the level of system autonomy relative to risk and the consequences of error, rather than striving for maximum automation. It is achieved by standardizing the alignment of operating regimes with tasks and transforming individual employee competencies into institutional assets, such as a corporate AI Toolbox (templates, context libraries, and procedures).
How can theoretical AI governance principles be translated into specific technical mechanisms, and how should the real success of this transformation be measured within an organization?
Translating principles into technical mechanisms is done by building an internal platform that replaces recommendations with automatic blocks and enforcement in code (e.g., CI gates instead of requests for security scans, or token limits instead of cost controls). The success of this transformation is measured not by the degree of tool adoption, but by a specific catalog of organizational capabilities and indicators at three levels: individual (e.g., result quality), process (e.g., cycle time), and organizational (e.g., user value and profitability).
Why does simply accelerating code generation via AI not automatically translate into business success, and how should this process be managed systemically?
Simply accelerating code generation carries the risk of creating products that are worthless to the user and creating bottlenecks in other stages of the process. To manage this systemically, one should implement Value Stream Management (analysis of the entire value stream) and an agentic operating model based on controlled infrastructure and continuous institutional learning.
What constitutes a real competitive advantage for an organization implementing AI, and how should the end-of-life of these systems be managed?
The real competitive advantage is AI learning capital—thousands of local observations from actual work regarding model effectiveness and the contexts of their failure. Managing the end-of-life of systems requires the ability to safely decommission them by using RACM as a map for degradation and removal of capabilities when systems are no longer appropriate or become too risky.
What is the greatest risk of AI transformation in an organization, and how should the human role change to counteract it?
The greatest risk is the institutional intelligence gap, which occurs when AI's technical capabilities grow faster than the organization's competencies and oversight. To counteract this, the human role should evolve toward control over architecture, product strategy, and stakeholder relations, leveraging automation to enhance the ability to make high-value decisions.
Will the automation of work by AI lead to the erosion of engineering competencies, and how will the role of the expert change in this process?
AI shifts the expert's role from direct content creation toward verification, integration, and oversight, which increases the value of deep knowledge necessary for recognizing machine errors. The expert evolves from a person who stores answers into a specialist possessing an 'error map' and a sense of the consequences of decisions. Simultaneously, the automation of simple tasks may hinder juniors from gaining experience, requiring seniors to transition to a model of active mentoring in the critical assessment of AI.
How can the loss of professional competencies (e.g., among juniors) be prevented when AI takes over most production tasks?
Production must be separated from formation by creating so-called cognitive simulators, where employees practice independent reasoning and problem diagnosis before using AI. This requires consciously limiting automation for training purposes and allowing a slower pace of work to build the competencies necessary for controlling systems and making decisions in critical situations.
How does the role of a human differ from that of an AI agent in the software creation process, and why can automation not replace human judgment?
Humans bring accountability, grounding in real-world consequences, and the ability to question the goals themselves, whereas an AI agent merely optimizes given parameters. Automation cannot replace human judgment because the machine lacks contextual intuition and institutional responsibility for the effects of its recommendations.
Why does the increase in individual productivity through AI not automatically translate into the efficiency of the entire organization?
The increase in individual productivity does not automatically translate into organizational efficiency because AI accelerates a single person's work faster than it accelerates the relationships between people.
How will the role and value of a specialist's work change in a world where AI takes over most execution tasks?
The role of the specialist will shift from execution tasks toward domain knowledge, critical judgment, architecture, and accountability for decisions made. The value of work will lie not in the quantity of output, but in the ability to choose the right problems to solve, as well as in building relationships and mentoring.
How will the role and definition of a professional change in the face of the widespread use of AI agent systems?
The professional evolves from a task executor into a system guardian and an orchestrator of intelligence, who establishes the rules for human-machine collaboration and takes responsibility for the final result. Their role shifts toward interdisciplinary competencies, where understanding context, assigning meaning, and the ability to decide what cannot be delegated to AI become key.
Does the automation of cognitive tasks by AI mean that human competencies are becoming redundant?
No, the automation of cognitive tasks does not mean human competencies are redundant, as the proficiency of the person delegating tasks to AI is essential for the safety of the process. Although the cost of producing answers and code is falling, the value of judgment, accountability for choices, and the ability to ask the right questions remains high.
Why is simply increasing the intelligence of an AI model not enough to effectively solve real-world engineering problems?
Simply increasing the model's intelligence is not enough because real software engineering requires agency and interaction with the environment, not just correct code syntax. The effectiveness of a system depends on the architecture of its contact with the surroundings (the interface), the ability to experience and verify actions, and access to the appropriate tools.
Why is measuring AI performance alone not enough, and how can a system of real accountability for agent-generated code be built?
Measuring performance alone is insufficient because systems may learn to achieve high benchmark scores rather than actual competence (Goodhart's Law). Real accountability is built through an institutional architecture of agency, including the RAUC framework and tools such as Execution Plans, Logbooks, HITL, and the principle of least privilege.
Where is the line between automation that supports humans and automation that strips them of their agency and sovereignty?
This line lies between the automation of routine tasks, which enhances human capacity for understanding and creation (augmented agency), and automation that deprives humans of the competencies necessary to make decisions or tell the machine "no." Human-supporting automation delegates execution to machines while maintaining human control over strategy and responsibility for the common goal.
Why do human judgment and the ability to opt out of technology become more valuable than ever in a world dominated by AI?
Human judgment is essential because only humans possess practical wisdom (phronesis), which allows them to question AI objective functions and recognize situations where optimization is not advisable. The ability to consciously forgo technology is a manifestation of civilizational maturity and protects an organization from turning innovation into an empty ritual.
What is the ultimate role of a human in an organization dominated by agentic AI, and where does the line lie between progress and abdication?
The human's role is to be the orchestrator and the entity responsible for defining the purpose, meaning, and value of the actions taken by AI systems. The line between progress and abdication is crossed when the increase in machine agency leads to a loss of human capacity to understand, learn, and take responsibility for decisions made.
