[2/4] Panel Discussion on AI’s Impact on Software Development (Category AI)
Continuing my account of the panel questions, here are my thoughts on the first two:
- What has AI changed in development?
- How do we increase AI adoption: from the top down or the bottom up?
1. Real changes in development brought by AI I covered similar ground in my CTO Conf talk, Integrating AI into Development Processes at a Large Company. A couple of years ago, AI programming tools still seemed unusual; now the vast majority of developers use them. According to the DORA 2025 report, 90% of technology professionals, from programmers to product managers, use AI daily in development. What has changed in practice? Routine tasks have become faster. Boilerplate code, templates, tests and documentation are often delegated to AI. DORA lists these among common uses:
- Writing new code
- Modifying existing code, including migrations
- Generating tests
- Writing documentation with AI
Actual gains remain far from utopian claims of severalfold acceleration. There are improvements, but they are not yet revolutionary. Last year’s DORA report even found a slowdown associated with AI adoption as teams struggled to adapt. DORA 2025 found, for the first time, higher software delivery speed associated with greater AI use. Quality and release stability remain challenges: throughput increased, but instability and failures remained elevated. That is unsurprising: after chasing speed, we now need to catch up on quality.
2. AI adoption: spontaneous bottom-up use or managed top-down change? Adoption often starts with developers themselves. Engineers began experimenting with widely available tools such as ChatGPT and Copilot well before official guidance. A Microsoft study found that 3 out of 4 employees already used AI at work, with around 80% bringing their own tools. In large corporations, the share introducing AI on their own reached 78%. “Bring Your Own AI” means people connecting third-party services to workflows, often without IT or management knowing. We saw this with personal devices and BYOD in the previous decade; now it is happening with AI. MIT’s The GenAI Divide: State of AI in Business 2025, which I discussed earlier, also examined this.
The upside is enthusiasm and quick local improvements. Developers, analysts and testers find useful ways to summarise requirements, generate SQL queries or automated tests, and use them without waiting for permission. The downside is disorder and risk. Dozens of people bringing in unapproved services can expose the company to data leaks, compliance violations and unpredictable results.
I think leaders should guide engineers’ enthusiasm rather than suppress it. Open discussion of AI goals, training, internal experimental sandboxes and central security measures can turn scattered initiatives into managed change. The best approach is ultimately a hybrid: bottom-up momentum supported by top-down strategy and oversight. Adoption then serves shared goals, with fewer surprises. This resembles the proactive design of rules and infrastructure for agent markets proposed in Google’s Virtual Agent Economies whitepaper, which I reviewed. Applied to the SDLC, we need clear rules and a deliberate approach to internal AI tools versus external models and tools, both for piloting capabilities and for moving successful experiments into production.
#AI #Engineering #Architecture #ML #Software #Economics #Software