When I first started experimenting with AI-assisted development — and was listening to the echo chamber of how “AI will replace human developers soon” — I couldn’t quite agree with the hype, but I also couldn’t fully explain why.
In some projects, AI worked brilliantly. It felt like a very knowledgeable (if slightly unimaginative) assistant — quick to refactor, eager to document, and sometimes even helpful in debugging.
In others, it was a disaster: incoherent code, endless loops of half-baked suggestions, and contradictions that took longer to fix than to write from scratch.
Over time, I noticed something. What applies to AI, matched my personal experience as well. Across different codebases — sometimes even within the same company — I felt like being a completely different developer. In some environments, I felt like a genius that could build anything in hours; in others, I felt like a monkey banging on a keyboard, barely producing anything useful - and definitely not quickly.
It can’t be just me that feels that way, right?
Now, after reading the 2025 DORA Report, I have the data to match that gut feeling. It’s not me, it’s you! Or rather - the environment. The report confirms what I felt when trying AI on these different projects: AI struggles in the same environments that humans do — only faster and louder.
“AI magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones.” — DORA 2025, p. 3
In other words, what’s good for developer velocity is even better for AI. What’s bad for humans becomes catastrophic for machines.
⸻
The Environments Where AI Fails Link to heading
Looking back at projects and codebases where AI fell flat, the same issues keep appearing — all the things that slow down or frustrate human engineers become deal-breakers for AI-driven workflows:
- Code spread across multiple repositories with no clear deployment path
- Long, fragile build pipelines and heavy dependencies
- No caching or incremental builds
- No reliable way to debug applications locally
- Undocumented credentials and missing configuration context
- Sparse or outdated documentation
- Missing logs, metrics, and traces or no direct access to them
- No consistent or automated testing
- No end-to-end or integration tests
- No stable test data for performance or regression testing
- Slow compile times and feedback loops that make iteration painful
- No way to deploy, test, and run a local copy of the full application stack
Humans can cope with some of these missing feedback loops through intuition, tribal knowledge, or tapping a teammate on the shoulder. AI can’t. AI doesn’t have context beyond what the system exposes — so if your build, test, and deployment processes are opaque, it will be flying blind and take wild guesses.
That’s why DORA’s 2025 findings resonate so strongly: AI isn’t a productivity tool you can drop into chaos. It’s a mirror that reflects exactly how chaotic your system already is.
⸻
Building Systems That Work for AI — and Humans Too Link to heading
The DORA team identified seven capabilities that make AI effective:
- a clear and communicated AI stance (is your company supporting AI on an organizational level?)
- a healthy data ecosystem (in the context of feature development - debug information, test results, benchmarks, logs, and metrics)
- AI-accessible data (can your AI agent see that data?)
- strong version control practices (are you using semantic commits and clean repository structures?)
- working in small batches (de-risk development with incremental changes, fast feedback loops)
- user-centric focus (provide the AI with product and persona descriptions, mockups, jobs-to-be-done, end-to-end tests, UI tests)
- quality internal platforms (have fast feedback loops, good developer tools, and focus on developer velocity).
You don’t need to memorize the details though to grasp the takeaway:
The same systems that help humans perform better, help AI even more.
Clean and fast build pipelines, observability, debug information, test results, logs, metrics and benchmarks, automated testing, short feedback loops — these aren’t “AI readiness” initiatives. They’re what good engineering already looks like.
And that’s the beauty of it: if you start building your environment for AI, humans will profit too.
Because AI doesn’t fix bad systems — it exposes them. And the best way to prepare for an AI-driven future is to build an engineering culture where both humans and machines can thrive.
⸻