Skip to content
back to the archive page
#Engineering

T-Sync Conf and T-Bank’s AI4SDLC 2025 Research Results (Category Engineering)

#Engineering #AI #RnD #Software #Management #Leadership #Metrics #Processes #Productivity

Today I will spend the day at T-Sync, a conference built around discussions and spaces for conversation rather than talks. I will have a space on the second floor where we can discuss the AI4SDLC research we launched last year with a community survey. You can find it at ai4sdlc.tbank.ru under the Research tab. It has two parts:

1️⃣ The 2023–2025 meta-study reviews around 50 individual studies from recent years. The recurring annual studies were particularly interesting: they show changing attitudes towards AI in development, adoption across use cases and assessments of its effects. These include DORA, GitHub, Stack Overflow, JetBrains, Atlassian, McKinsey, Yakov and Partners / Yandex, Express42 and MIPT. The site has a list of research series with their main trends and a timeline of all included studies, each with a short summary.

2️⃣ The survey results cover the survey we ran late last year. It was extensive, and I want to thank everyone who participated. Around 1k people started it. Luckily we put AI use and impact questions first, because not everyone made it to the end :). The respondents’ profile was:

  • 50% developers, 17% technical leaders, 7% senior executives, 6% QA engineers, 4% AI/ML engineers, and other roles.
  • 24% worked at large companies with 1k–10k staff; 23% at very large companies with 10k+; 18% at medium-sized companies with 100–500; and 16% at small companies with 10–100.
  • Main industries: finance 25%, technology 20%, retail and ecommerce 15%, telecoms 7%, education 5%.
  • Application and service types were fairly evenly distributed: internal versus external users, 45% versus 37%; B2C versus B2B, 34% versus 41%.

The executive summary of the results:

  • 58% use AI for code generation or completion “often/always”; 24% for code review; and 18% for legacy modernisation, where 42% answered “never.”
  • 64% reported higher productivity, including 18% reporting a significant increase; 37% reported improved ability to write code.
  • Code quality improved for 32%, worsened for 14% and was unchanged for 42%.
  • Trust in AI code was low: 49% did not trust it, 11% did, and the rest were neutral.
  • AI’s role within the company grew for 54%. Adoption plans felt transparent to 50%, but not to 21%.
  • On system readiness, commit-to-production time was measured in weeks or months for a substantial group. Documentation worked poorly as a source of truth: around 39–42% did not rely on it in critical situations.

These findings resonate with the meta-study:

  • AI adoption is approaching “almost everyone,” for example around 90% in DORA’s 2025 findings. Individual effects are usually stronger than team-level effects.
  • Bottlenecks move to integration, testing, review, releases and communication. Studies repeatedly note that coding occupies only part of an engineer’s time and that organisational barriers consume a substantial part of the week.
  • Quality and trust are central maturity issues: “almost right” answers create additional verification and debugging work.
  • Experiments show that effects on complex tasks may be zero or negative. AI amplifies the system; it does not replace engineering.
  • Without tests, quality gates, small batches, review and observability, “acceleration” can easily become “accelerating chaos.”

If you are interested, come to the stand. I will explain how the research fits together and answer your questions.

#AI #RnD #Software #Engineering #Management #Leadership #Metrics #Processes #Productivity