[2/4] Panel discussion on the impact of AI on software development (AI column)
Continue. story I will share my thoughts on the first two questions.
- What has changed because of AI?
- How to increase AI adoption? Top or bottom?
1. Real changes in development influenced by AI I talked about that in my report.Integrate AI into development processes in a large company" at CTO Conf. But in a nutshell, a couple of years ago, AI tools in programming seemed like a curiosity, and today they are used by the vast majority of developers. I agree. DORA report 2025, 90The percentage of technology professionals now use AI every day in development, from programmers to product managers. What has changed in practice? The developers felt the acceleration of routine tasks. The generation of generic code, templates, tests, and documentation is now often delegated to AI. According to Dora, one of the most common cases
- Writing new code. Modification of the existing (ala code migration)
- Test generation
- Writing documentation using AI
At the same time, the real effect is far from a utopian multiple of acceleration – there are improvements, but not revolutionary ones yet. If last year in the DORA report there was even a slowdown due to the introduction of AI (The teams did not have time to adapt.)Now the situation is clear: Dora-2025 For the first time recorded an increase in the speed of software release with an increase in the use of AI. At the same time, the quality and stability of production remain a challenge. DORA notes that while throughput has grown, instability and disruptions are still elevated. This is expected: first, everyone was chasing speed, now you need to tighten the quality.
2. AI implementation: spontaneously “from below” or controlled “from above”? Practice shows that the initiative to introduce AI often comes from the bottom – from the developers themselves. After the advent of mass tools such as ChatGPT, Copilot, etc., engineers began to experiment with them long before official instructions. For example, Microsoft showed: 3 from 4 Employees are already using AI in their work.80Percent do so through personal, home-branded tools (BYOAI). And in large corporations, the share of those who willfully implement AI reaches 78%. That is, the phenomenon of “Bring Your Own AI” has arisen everywhere: people connect third-party AI services to work processes, often without the knowledge of the IT department and management. So it was with personal gadgets. (BYOD) In the last decade, it's now the same with AI tools. That was the study.The GenAI Divide: State of AI in Business 2025from MIT that I'm already handler.
The upside is enthusiasm and rapid local improvements. Developers, analysts, testers find useful AI chips for their tasks (Summarizing requirements, generating SQL queries, autotests, etc.) They don’t wait for permission to use them. The downside is chaos and risk. When dozens of people are dragging uncertified AI services, the company faces threats: data leakage, compliance violation, unpredictable quality of results.
I think leaders need to oversee, not suppress, the enthusiasm of engineers. We need an open dialogue about AI goals, training, internal sandboxes for testing, and centralized security measures to turn disparate initiatives into manageable transformation. Ultimately, the optimal approach is a hybrid one: take advantage of the bottom-up momentum, backed up by a top-down strategy and control. Then AI is introduced not randomly, but in line with general goals, and the company extracts the maximum benefit without unnecessary surprises. This is similar to the approach proposed for the development of agent economics in the whitepaper.Virtual Agent EconomiesFrom Google, which I already have. handler. It proposes proactive design of rules and infrastructure for agent markets. If this is portable to SDLC, then we need clear rules and design for using internal AI tools vs external models/tools. This is necessary for piloting new capabilities and clear rules of transition to production of these developments.
#AI #Engineering #Architecture #ML #Software #Economics #Software