AI can be a great help with tasks that are well defined, routine and require little independent judgement. Examples include producing certain pieces of code, writing repetitive solutions, or creating the first version of documentation or certain tests.
The developer time freed up this way can be spent on work where human experience and judgement add more value: designing the architecture, clarifying the business logic, understanding the system as a whole and ensuring long-term maintainability.
The study’s results also warn that the impact of AI tools should not be treated as uniform. Coding assistants and AI agents capable of carrying out longer tasks showed different results. At the companies studied, introducing agents came with a significant increase in coding activity metrics, while the review process also became more intensive. For assistants, the measured effects on coding were much more modest, and no significant increase was found in the review metrics examined either. This does not mean that one type of tool is better than the other in general. It simply shows that the outcome of introducing AI is determined jointly by the tool in question, the nature of the task and the development process.
AI can also be useful as a thinking partner. It can help look for alternatives, find the weak points of a plan, or even act as a “rubber duck” when we try to solve a problem by thinking it through out loud. In these situations, however, what matters is not only the answer AI gives, but also whether the developer is able to judge if that answer is correct. Choosing between alternatives, clarifying requirements and understanding the system as a whole remain professional tasks.
The experiences presented in the study also indicate that checking AI-generated code does not become unnecessary just because the code itself is produced faster. Part of the developer’s work shifts from writing code directly to evaluating suggestions, spotting errors and improving the solution. This, however, still requires the right professional knowledge. To recognise when a solution suggested by AI is good, and when it only looks good at first glance, we need to understand the underlying principles and the system as a whole.
In practice, therefore, three questions are worth asking before carrying out a task with AI support:
- How precisely can the expected result be defined?
- What reliable method can be used to verify that the result is correct?
- How serious would the consequences be if an error still got past the checks?
For a simple, easily testable task, AI can be given more room. For a change affecting business rules, permissions or critical data handling, however, stronger checks may be needed. At the same time, the aim is not to create a rigid set of rules, but to adapt the way work is organised to the expected quality and the risks. The goal is for AI to speed up development where the result can be checked properly, and for speed not to come at the expense of checking, quality and reliability.
So the use of AI has to be introduced and run deliberately. The next question is how we can make sure that the finished code not only works, but is also of high technical quality and provides a solid foundation for further development.
Source: Fiona Chen and James Stratton, Harvard University: Artificial Intelligence in the Firm: Bottlenecks in Software Production. This post is based on the version of 2026. 08. 04.
Was your software built with AI? Have it reviewed by an expert!