The way we build and run software is changing. For decades, the model was simple: humans wrote instructions and machines executed them. Today a new layer is emerging, where humans express intent, AI interprets and plans, and machines execute. This shift is already reshaping how code is written, but its impact on deployment and production operations is less clear, and arguably where the stakes are highest.
Production environments demand reliability, predictability, and accountability. A mistake in a deployment or an operational change can mean service outages, SLA breaches, or security exposure. So the question is not just whether AI can help, but where it adds real value, how much autonomy it should have, and what guardrails are needed.
Practical examples, tools, lessons learned, and even failures are all welcome. The goal is to understand where AI genuinely strengthens production operations today, and where caution is still needed.
i’d give ai the narrowest reversible slice first: prepare the change, explain the expected state, run preflight checks, and collect the rollback command. execution should stop at a policy boundary unless the environment and blast radius are explicitly allowed.
the useful guardrail is independent readback. don’t accept “deployment succeeded” from the same agent that clicked deploy. verify the live version, health, one consequential user path, and the rollback state through separate checks. if any result is uncertain, the system should hand a human the exact action path and first mismatch, not improvise.
ai is strongest at assembling context and executing known runbooks. i’d be cautious where the oracle is subjective, the rollback is incomplete, or the change touches data and permissions.
I need to find a blog, which a colleague Eric wrote years ago, he continues to explain how AI is and will be used in Enterprise. How he explains it tallies better with how events in the 1760’s really did pan out. Sure it took about 80 years, but that lifetime has now been compresses into 2 decades, he refers to it in terms like a carpenter might have used, but a code-carpenter. As a hand-tools person myself, the AI is the table-saw and the chainsaw and the dovetail maker machine. All capable of much more quickly making beautiful furniture for indoors or outdoors, but none able to apply a final varnish. And if we continue to use small slow techniques, we will be left behind by competitors who are early adopters. and for me that’s the real driver.
We can separate the learning paths that the models have to go down, to cover security and to cover other things we value later. We will have to do this incrementally over time, because ATM the engines are yet to be trained on a variety of disciplines in which humans still excel. Circuit design is the current big frontier, others will follow, and at some point we will find we have to iterate and rewrite many of the older models, this will snowball… but.
If we try to predict the future we will not just fail to, but we will waste time. We will also not grasp the most vital lesson of being a human, and that is to learn from outcomes and adapt faster than any other animal on the planet ever has, and ever will do. So I disagree with giving it small slices of work, sure, test what it produces, but we are in a time of vast change. And making those same small incremental changes like the computer industry has seen other the last 60 years, is not going to see you staying in business for long.