I started writing this with a bit of a bittersweet taste in my mind. I'm happy with where I am right now — a better version of myself, in a way, though one still built on the foundations that made me a decent engineer before any of this started. If everything went wrong tomorrow and I had to go back to working the way I did two years ago, I'd be fine. Slower, but fine.
What surprises me most is how much my scepticism has shifted — not just about whether to use AI, but about how many things I've started using it for that I never would have considered before.
I usually joke about feeling like Tony Stark, and in some ways I do. Right now I'm talking to a computer using voice-to-text, and if I mis-speak, it just understands what I meant and corrects it. We've quietly crossed into a world where that doesn't surprise us anymore, and a few years ago we couldn't do it to this level.
But the real story is the journey from being skeptical of anything AI told me, to trusting it as a platform for becoming a super user of my own life and work.
How it started: talking to ChatGPT and not trusting a word of it
I remember starting to use ChatGPT at work, around two years ago, give or take. It felt like cheating, if I'm honest — the policies around AI usage weren't clear yet, so I wasn't even sure if my company would be happy about it. But the first real surprise was just how close it felt to talking to a human. It was breaking the Turing test for me, repeatedly.
I started small: a function here, a piece of an algorithm there, mostly in Python. It worked best when I gave it tight, well-defined inputs and outputs — I wasn't doing BDD yet, wasn't thinking in Gherkin, just thinking like I always had about edge cases and optimisations.
Trust was the real constraint, though, and not just in code. RAG didn't really exist yet, so you were always conscious of where the model's training data ended. I remember pasting a large legal document into ChatGPT to reformat for the web and being genuinely nervous — would it come back complete? Would it be accurate? That was a critical feature in a system I was responsible for, and trusting something that consequential to a chatbot felt reckless at the time. It's strange to think how far we've come from that — models now feel like they're defined by their capabilities, not by what they happen to know.
How the industry shaped a new mindset
Early on, everyone was figuring this out alone. There were a few prompt engineering courses, scattered conversations, the odd company experimenting with how to fold this into people's workflows — but nothing settled. I remember watching demos of Microsoft's early multi-agent work and thinking that looked like the future. I couldn't have told you we'd be living inside that future, daily, within two years.
The bigger shift came when the industry stopped thinking about AI as a feature and started thinking about it as a workflow. I've watched people go through something like an identity crisis when their workflow gets disrupted — and I never really felt that. I've always seen my job as solving problems, not as being the person who's good at a particular stack. So when the ground shifted, it didn't feel personal.
The real turning point for me was the arrival of spec-driven development — I'll call it a "style" rather than naming a specific framework, since what mattered was the underlying idea: pushing human interaction to a higher level of abstraction. For a long time I resisted this. I treated agents purely as co-pilots, sitting beside me while I did the real breakdown and validation myself, scepticism intact.
That's changed. I'm far more comfortable now moving up a level and letting the agent surprise me with the outcome. Even on complex, brownfield projects that still need real guardrails, the results have become noticeably more reliable. Some people complain about agents going off the rails — that hasn't really been my experience. Give it a clear scope and tight verification boundaries, and it tends to do a solid job. That shift moved my whole focus up a level.
My role isn't really "architect" anymore, technically speaking — it's closer to product thinker. I still invest heavily in technical depth, because understanding how things work underneath is what lets me make good calls. But I don't tell a plumber which exact fitting to use. I tell him what I need, draw a few hard boundaries — don't break this floor — and let him work. That's roughly where my trust in AI sits now: I can let it be, and it's going to be right, mostly.
Software has one big advantage here: we can always go back. We can revert. As long as we hold a minimum bar of professionalism, we can throw things together quickly, evaluate, and discard what doesn't work. Some people see that as failure — aiming low on the first shot. I don't. A large part of our value as engineers is still in the iteration itself: discovering the problem, testing assumptions, being willing to come back and change them as the picture sharpens.
AI in everyday life
Lately AI has crept into my everyday life too, and I'm starting to enjoy it. A few days ago I delegated two fairly significant tasks to it — one personal, one professional — both involving some risk, in the sense that I gave it access to chats and tools that could be misused, even if no actual responsibility was ever delegated (nobody should do that). What surprised me was how much trust I'd built up, and how consistently it's earned that trust. Of course, there are limits — it's not getting anywhere near my bank account.
I've started delegating things like drafting expenses, or digging through a long chat history to surface something specific. It does a genuinely good job, and it's taken real pressure off me. Learning how to delegate well is its own skill, one I'm still building — and I suspect that's true for a lot of people. Maybe that's the next thing this technology forces us to get better at: not becoming a boss, just becoming more effective with our own time.
Final thoughts
I'm not the super user I'm aiming to become yet, but the trust, the tooling, and a better sense of what to hand off and how to explain it are getting me there. A lot of the background noise — I need to do this, I'm running late, I haven't done that — has gone quiet. When something comes up now, I open a session, lay out the goal and how we'll verify it, and let it run. I won't get 100%, but 80% with minimal input from me is a very good trade. I think we should be a bit more optimistic about where this goes. Yes, it will change things. But I'm hopeful it pushes more people away from repetitive manual work and toward something closer to coordination.