How will AI reshape Software Engineering? Part 3: AI alleviates human constraints
- Yevgen Nebesov
- Mar 17
- 6 min read
Introduction
This is the third and final part of an article exploring how AI will reshape software engineering:
- Part 1 introduced the idea that complex sociotechnical systems, such as software engineering, evolve under constraints.
- Part 2 reinterpreted the social and technical innovations of recent decades as responses to two fundamental human-induced constraints: limited cognitive capacity and the time required to complete work.
Now, in Part 3, we project how software engineering might evolve once AI alleviates these constraints. In fact, AI possesses virtually unlimited computational and memory resources, mitigating the constraint of limited cognitive capacity. Additionally, its calculation speed surpasses that of humans by orders of magnitude, relaxing the second constraint. The following sections project how these AI capabilities will impact software engineering in the Co-Pilot and Pilot phases.
Co-pilot Phase
As discussed in Part 1, we are currently in the Co-Pilot Phase. Various artifacts—such as code, requirements, tests, and UI designs—are now co-created with the assistance of AI, as illustrated in the image below. In this phase, both humans and AI co-pilots engage with the same artifacts: they can read them and write them.
This marks a shift where the time constraint on human labor begins to be partially alleviated. However, while individual artifacts can be produced more quickly, the alignment between artifacts still relies on alignment between people, as in the picture below. As a result, human constraints remain the dominant limiting factor.
Many claims suggest that AI co-pilots will boost team productivity by orders of magnitude. However, such assertions often fail to consider the Theory of Constraints, demonstrating that local optimizations do not necessarily lead to global improvements.
Pilot Phase
The real transformations will occur in the second phase, where AI will evolve from a co-pilot to a pilot. In this phase, humans and AI will no longer work on the same artifacts. Humans will communicate with AI in natural language, while AI will independently edit artifacts.
Although this may seem like a linear extrapolation of the co-pilot phase, the pilot phase will introduce emergent properties that were not possible before: a reduced need for abstractions, AI-driven alignment, and, ultimately, the alleviation of human constraints.
Less need for abstraction
As stated in Part 2, we rely on abstractions to minimize rework because human effort and time are limited. However, if AI can create and recreate entire artifacts and systems within minutes, the necessity of avoiding rework disappears—along with the need for innovations designed to reduce it.
Example: Return of Machine Code. Consider the evolution of machine languages. The JVM's promise is "Write once, run everywhere." With AI capable of generating code, a new paradigm emerges: "Say once, deploy everywhere." AI can generate machine code for any target processor architecture.
If humans are no longer reading the code, why maintain the JVM abstraction? It is reasonable to suggest that in the Pilot Phase, a significant portion of code will be machine code rather than high-level languages designed for human maintainability.
Example: Thick Frontends. A typical pattern in software development is to place as much business logic as possible in the backend. After all, we don’t want to rewrite it for every frontend, and we aim to avoid tight coupling between the frontend and backend—since high coupling leads to increased rework, which humans tend to avoid.
However, these considerations do not apply to AI, as it can (re)generate both frontend and backend code quickly. This raises the question: why place calculations in the backend and pay for server resources when the load can be distributed to clients—such as web browsers and smartphones?
As a result, we can anticipate a partial return to thick frontends in the pilot phase.
AI-driven alignment
With AI creating artifacts, the nature of those artifacts will also change. The current types of artifacts stem from various categories in software engineering—such as programming, design, UI/UX, requirements engineering, and QA—driven by the human constraint of being unable to handle everything at once. However, since AI can generate all artifacts, it will likely eliminate many intermediary artifacts that humans traditionally create.
Moreover, if all artifacts are stored in a central knowledge-based system (KBS), AI can autonomously align knowledge across different domains.
This marks a significant shift. Imagine an engineer interacting with an AI-driven KBS:
"I need a new feature X. Here is what I want the system to do: ..."
The KBS could process the request and respond with suggestions such as:
- "Okay, let me show you a possible UI after the change. What do you think?"
- "Here is a list of test cases I generated for you. Am I missing anything?"
- "By the way, this feature conflicts with feature Y under context Z. How should I resolve this conflict? Here are a couple of ideas: ... After we agree on the requirements, I will redeploy the entire system in no time."
This is a game-changer—human-driven alignment becomes less relevant - AI takes this responsibility over.
Will we still need social and technical tools that facilitate human-driven alignment? Managers and Scrum Masters serve as such tools. So do refinement meetings and other sessions where people in different roles ensure that their artifacts remain consistent with one another. Established tools and methodologies like DDD, MBSE, ALM, Scrum, SAFe, and many other technical and social innovations will be challenged.
Will we still need them? Time will tell. What is certain is that the demand for human-driven alignment will decrease by an order of magnitude as AI-driven knowledge-based systems take center stage, managing all artifacts and becoming the core of information flow.
The role of humans
Now, the question arises: What will humans do after we reach the Pilot phase and AI takes over the role of the Great Aligner?
At least two aspects will remain inherently human.
1) People as a source of real-world data
Computer systems require data, and since we lack sensors for everything, humans continue to serve as a crucial interface for inputting data into the digital world. Consider doctors entering patient records, store workers assigning prices to products, researchers documenting experimental observations, border control officers recording travel documents, and warehouse employees scanning barcodes. While computer systems are expanding further into the physical world, reducing human involvement, we are still decades away from replacing German customs workers—operators of registries, staplers, and hole punchers—with €1 sensors.
2) People as a source of agency
The defining characteristic of human beings is agency—the ability to shape the environment according to our will. All computer systems to date have been built based on human-defined requirements. In an AI-driven Knowledge-Based System (KBS), humans will continue to be the primary source of system requirements.
However, even this uniquely human trait is under threat. Algorithms are already influencing our decisions—from buying habits to political choices (see the meme below).
For the past few hundred years, the Western world has lived in an era of individualism—a radical idea that every human life matters. Not just the king’s or the priest’s, but all individuals. As a result, modern software development revolves around User Stories, ensuring that every feature can be traced back to an individual user. Concepts like User-Centric Design put individuals at the center of decision-making.
This focus on individual rights and agency has shaped ethical discourse—we don’t prioritize the family, the tribe, the city, the nation, or even the planet above the individual. Instead, we believe every person deserves to live, be healthy, and pursue happiness. A novel and radical idea!
AI, however, may challenge this paradigm. If AI can deploy systems within minutes, why continue focusing solely on individuals? Perhaps shifting the focus to broader units, such as a nation, is also worth considering. Imagine a nation’s defense system that allocates specific individuals to war for the greater security of the nation. Or an environmental AI system that relocates entire populations to preserve ecological stability.
But until this distant future arrives, humans will remain the dominant source of agency for computer systems—the ones providing real-world data and defining system requirements.
Conclusion
While Generative AI began disrupting software engineering a few years ago, much of the focus has been on the technology itself rather than on how it will reshape the ways we develop systems.
This series of articles presents a guided projection of how AI adoption may unfold. It argues that software engineering today is shaped by two human constraints:
Limited cognitive capacity
The time and effort required for work
Most innovations in software engineering—both technical and social—can be traced back to these constraints. As AI alleviates these limitations, not only will technological toolsets evolve, but so will the communication topology within software organizations. With AI taking over the role of alignment, the demand for many existing tools and methodologies—ranging from technical frameworks to social approaches like Scrum—will decline.
The future of software engineering will likely revolve around a central AI-driven Knowledge Base System that generates and maintains all artifacts, while engineers will interact with this system to provide real-world data and define requirements.
Comentários