2025 Was the Year AI Became Infrastructure
Every December I look back at how my actual work got done, not how I imagined it would get done back in January. It is a habit I picked up years ago running production teams, where the gap between the plan on the whiteboard and what actually happened on Sunday morning was always instructive. Some years the review is humbling. This year it has one clear headline: AI stopped being a novelty I visited and became infrastructure I stand on.
That word, infrastructure, is doing specific work in that sentence. Infrastructure is the stuff you stop noticing. Nobody walks into a building and compliments the plumbing. You only think about the electrical panel when it fails. Twelve months ago, AI in my work was the opposite of that. It was a destination, a separate place I went to do a separate activity. Now it is under the floorboards of nearly everything I ship, and most days I do not think about it at all. That shift is the whole story of my year.
What changed this year
At the start of the year, I used AI the way most people did. A chat window. Paste something in, get something out, copy it somewhere useful. Helpful, but bounded. The model was smart and the workflow was dumb. Every session started from zero. Every result had to be hand-carried across the gap between the tool and the actual work, and I was the courier.
I remember the texture of it clearly because I did it dozens of times a week. Copy the meeting notes. Paste them in. Ask for a summary. Copy the summary. Open the email tool. Paste again. Fix the formatting the paste broke. Adjust the voice because the model did not know our voice. Send. Eight steps, and the model only did one of them.
By year end, the shape is completely different. Assistants connected to my real systems draft community posts in our brand voice, stage email for review, keep content calendars honest, and carry context from a meeting note all the way into a published piece without me playing courier in the middle. Agents handle multi-step work while I do something else. I can start a task, go run an errand, and come back to a draft waiting for review. Not a perfect draft. A reviewable one, which is the correct goal.
The standardized connectors that made this possible went from announcement to industry default in a single year. The Model Context Protocol is the one I have built the most on, and I want to be careful not to make it sound more glamorous than it is. A connector standard is about as exciting as a wall outlet. But that is exactly the point. Before standardized outlets, every appliance needed its own wiring job. Before standardized connectors, every AI integration was a custom project that broke when anything on either side changed. Now I wire a tool up once and every assistant I use can reach it. The boring standard is the thing that turned clever demos into dependable plumbing.
What I actually shipped with it
Talk is cheap in this space, so here is the inventory.
A brand identity stood up in 30 days. An organization I serve as communications director launched globally this month, and the brand behind that launch went from nothing to launch-ready in about a month: the visual system, the voice, the templates, the site content, all of it. I have built brands before. The creative decisions took the same amount of human judgment they always have. What collapsed was everything around the decisions: the variations, the resizing, the rewriting of one announcement into six formats for six channels. Work that used to consume the calendar now consumes an afternoon, which means the judgment gets more of my attention, not less.
A steady publishing cadence across email, social, and community platforms that one person could not have sustained alone two years ago. The pipeline is not exotic. Source material comes in, drafts get generated against a documented voice, everything stages for my review, and nothing goes out without a human eyeball on it. The magic is not any single step. It is that the pipeline runs every week whether I am inspired or not. Consistency used to be a staffing problem. Now it is a systems problem, and systems problems are solvable by one careful person.
Video curriculum work where transcription, cut planning, and drafts happen in minutes. I have produced teaching video for a ministry education organization for two decades, and the pattern was always the same: a day of recording followed by weeks of the slow parts. Transcribe, log, plan cuts, draft the supporting text. That tail is now measured in minutes and hours instead of weeks. The teaching itself, the part that actually matters, is untouched. The scaffolding around it evaporated.
App prototypes that went from idea to usable in days. Working forms, real data models, actual authentication, in front of real users fast enough that their feedback still mattered. Ideas that would have sat in a someday file became things I could hand to people and watch them use, and watching a real person use a rough version teaches you more in ten minutes than a month of planning documents.
None of it was AI doing my job. All of it was AI removing the eleven mechanical steps between my judgment and the finished thing.
What did not change
This part matters just as much, and it gets left out of most year-end retrospectives because it is less fun to write.
Judgment did not change. Every pipeline I run ends at a human review gate, and this year confirmed rather than weakened my conviction about that. The drafts are good. They are not accountable. I am. When a post goes out under an organization's name, a person decided it should, and that person can explain why. I spent twenty years in rooms where the livestream was going out whether we were ready or not, and that experience translates directly: the tools got faster, but somebody still has to be the one watching the output before it reaches the audience.
Relationships did not change. The trust that lets me publish on behalf of an organization was built over years of showing up, and no model shortens that. If anything, the ease of producing content has made the trust more valuable, because everyone can produce volume now, and very few people are trusted with voice.
Taste did not change, except that it became the bottleneck. When production is nearly free, the scarce skill is knowing what is worth producing and what good looks like. That is a strange inversion for someone who spent years being the person who could physically operate the gear, and I think it is the most underrated shift of the year.
The model race, and why I stopped watching it
The models themselves leapfrogged each other all year, and by fall the frontier releases were arriving weeks apart. I spent the early part of the year tracking every release the way I once tracked camera announcements, and I gradually realized I was consuming sports coverage, not doing my job.
My take on that is on record: the race matters less than your workflow. Here is the practical version of the argument. When my brand voice lives in documents I own, when my templates are files rather than memories, when my connectors follow a standard, a better model is a drop-in upgrade. I swap the engine and the car still knows where it is going. The teams I watched struggle this year were the ones who built everything around one model's particular personality, then had to start over every time the leaderboard shuffled. Portability turned out to be the real strategy, and portability comes from owning your context, not from betting on a winner.
There is a quieter benefit too. Stepping off the release treadmill returned a surprising amount of attention. The news cycle in this field is engineered to make you feel behind. The workflow, once built, is engineered to make the news cycle mostly irrelevant. I recommend the trade.
What this looked like from a small shop
I want to name something for the solo operators and small teams reading this, because most of the writing about AI infrastructure assumes an enterprise budget and a platform team.
I am a consulting practice of approximately one. The organizations I serve are ministries and small businesses, not companies with innovation departments. And this was the year the leverage genuinely reached us. The pipelines I described are not enterprise software. They are documents, templates, connectors, and habits, assembled by one person who reads documentation and is willing to be patient. The gap between what a large organization can automate and what a careful individual can automate closed dramatically this year, and I do not think most small organizations have noticed yet. The ones that notice first are going to punch far above their weight next year.
There is a caution inside that opportunity. The same leverage that lets a small team sound bigger also lets a careless team publish faster than it can think. The organizations that treat AI output as finished product are going to flood their own channels with words nobody chose. The review gate is not a speed bump. It is the difference between having a voice and having a volume knob.
What I expect next year
Boring is the goal. I mean that as a sincere prediction, not a shrug.
The demo era is ending. Everyone has seen the impressive thing. The organizations that win with this technology next year will not be the ones posting demos. They will be the ones with reliable, reviewed, slightly dull pipelines that ship every single week. The ones where the AI work is documented well enough to survive a staff transition. The ones that treated their prompts and context like assets, versioned them, and can hand them to a new hire on day one.
I expect the interesting arguments to move from what can the model do to how do we supervise what it does, and I welcome that, because supervision is a discipline production people have been practicing forever. Run sheets, redundancy, rehearsal, review. The vocabulary is new. The discipline is not.
Infrastructure does not trend. It does not demo well. It just holds everything up, quietly, every week, while the people standing on it get to spend their attention on the work that actually needed a human. That is what AI became for me in 2025. Next year I intend to notice it even less.