Earlier this year I wrote about the AI Engineering Maturity Model — a five-level framework for understanding how engineers actually work with AI, from ad-hoc chatbot use all the way through to bounded autonomous workflows. This post is the first set of results.

We all feel that the use of AI in software engineering will be transformative, and there are a lot of enthusiastic evangelists around. Going behind the hype reveals a much more mixed picture of adoption. My goal was to get behind the great anecdotes and get an honest, org-wide picture of how engineers are really using these new tools.

For those who haven’t seen the original post: the AIMM is a five-level maturity scale scored from 1.0 to 4.0. L1 (Explorer) is where engineers are experimenting individually with AI tools. L2 (Tool Adopter) is active, habitual use — IDE plugins installed, regular AI interaction — but practice is still individual.

Last year, I dismissed AI coding assistants as fancy autocomplete tools. After five decades of programming, I thought I'd seen every productivity promise come and go. I was wrong.

My journey started simply enough with vanilla Visual Studio Code. Microsoft Copilot came next, making big promises about revolutionizing coding. Skeptical but curious, I tried that, then Cursor, a VS Code fork that actually delivered when pointed at the right code along with the manufacturers documentation. Adding Claude through Cline opened new possibilities, followed by Roo Code with its specialized prompts for architecture, coding, and code review.

Then DeepSeek-R1 arrived, unlike the other leading AI LLms, it is open-source, meaning anyone can use, modify, or share it for free. matching Claude's capabilities at a tenth of the cost. This constant evolution taught me something: yesterday's cutting-edge tool could be tomorrow's expensive luxury.

5 new tools to learn in one year, it’s what I live for.