One Year In: What Applied AI Looks Like (A Fun List)
Today marks one year in my current role as Director of AI, Analytics, and Customer Care.
It's been one of those years that simultaneously feels like a flash and epoch. There have been many changes in my life, challenging work projects, and a consistent pressure to stay updated with the world of AI. It's also been an exciting year learning new things, seeing large technical projects deployed, and building a team that is objectively transforming Zinus.
So, with a year to reflect on, I wanted to make a list of things I have learned in the past 12-months. Some are deeper than others, but all of the below match conversations I have had, mainly with my wife, and others on the state of AI.
Fair warning: this article is AI-enhanced. I wrote this article in three phases:
I wrote a rough draft in Obsidian (my go to daily platform)
I then passed it to a Claude Project with Claude Skill that tracks all my writing and helps my grammar but has clear instructions to keep my dad jokes - those are non-negotiable
Finally, I did one more manual pass where I always make edits
Onto the List
"AI Brain" is real and I likely define it differently than others. It's when people use AI to be confidently wrong. This isn't just in my work life, but the number of times people have told me how to do something or a "fact" which can be verified as incorrect in less than a minute is shocking and scary.
AI is great at purpose-built tactical items but struggles with strategy. Yes, AI can analyze a report (with clear instructions), cross-reference documents, code a website, write schema markup, analyze a contract, write an email, and thousands of other clear sequential work items. But when it comes to strategy involving humans, corporate politics, internal policies; well, that's another story and a problem I have been trying to solve for months. If you have an answer OTHER THAN give it more context, I'm all ears.
It's frustrating to have AI work one day and then not the next because of quick model releases (looking at you Anthropic). I will go weeks making progress and then a new model will release and I have to change my workflows because each model has a personality of its own.
I worry about junior level employees (and my kids) as AI is great at semi-automating to fully automating redundant sequential tasks. These are the type of tasks that hone skills early in a career and provide experiential lessons. Not only that, but juniors come up with creative new ways to execute after being handed a redundant task that hasn't been optimized in years. We could lose a critical part of the learning cycle in the modern workforce.
Some days I feel like I'm working in a sci-fi movie with multiple computers running, agents moving, and projects being executed left and right.
Some days I deeply miss the manual work I used to do coding landing pages, building reports, or simply testing tech.
Fable is better than I thought it would be and built ~90% of the architecture of an enterprise AI agent in two prompts.
My default writing is now in markdown (MD) and I use Obsidian all day everyday even for personal items. A few months ago, with the help of Claude, I built a workflow where my Obsidian vault syncs to GitHub as a "mirrored" vault and then I connect AI to the mirrored repo. This keeps AI away from my primary vault as I fear having my data wiped.
Building simple video games in Claude with my kids has been a treat I did not have on my AI bingo card. Here are two examples:
My daughter built this one for a school project --> https://txcountrynerd.github.io/pixel-quest-assets/forest_run.html
My son and I built this one during a recent airport layover --> https://txcountrynerd.github.io/time-machine-five/
AI still has a long way to go.
I firmly believe in using AI to help people enhance their career and not as a tool to replace people. In my customer care responsibilities, our AI tools have helped streamline work to the point we promoted two people and took them off their previous work and placed them in more technical work, all while nothing was dropped. They are now doing projects they are passionate about and it shows.
I used to build Claude Projects routinely but now it's less often as I have gravitated toward Claude Skills. Agents are rarer and only built when I really need them since most of my job involves strategy (see above).
My ability to train people on how to use AI has accelerated, I look back on my first training sessions from a few years ago and cringe.
Since I have a background in websites, building HTML presentations has saved me so much time and helped improve my presentation game. Before you say, yea but corporate wants PPT, I know how to export the presentations as editable PPTs for those that need their PPT fix (seriously, corporate, y'all need to break the PPT addiction).
The most challenging part of AI is getting all the necessary knowledge into systems, many times the amount of time it takes to structure input data into AI for what you get out isn't worth it.
MCPs are amazing, and I think they point to where work is headed: fewer interfaces for non-power users. There will always be power users of a given application who need to log into their platforms. But managers and others who only need basic info will use MCPs and connectors to tap into those platforms instead.
Too many people say, "AI can do it," do what?
Having clean data has always been a struggle and it's even more evident and important with AI. One of the valuable aspects of AI is that it moves faster than humans, but bad data means that speed becomes a major liability.
Enterprise AI deployments are really no different than any large tech deployment, if anything, they are more difficult due to hallucinations compared to the past with deterministic code. I miss those old school tech deployments.
Speaking of hallucinations, they are mind-numbingly frustrating as you can build the orchestration layer, sub-layer, tools/functions with code, test for weeks, and when you go live the AI will still make things up at random.
I use 4 frontier platforms (Anthropic, OpenAI, Microsoft, and Google) and they are all good for various items, I use Google/Gemini the least because it doesn't really apply to my day-to-day work.
There are many niche AI companies doing really cool things, the hard part is finding use cases to justify the costs.
I worry that the future will be people begging for tokens. Have you ever seen the movie "In Time" where people fight over time, as in time they have to live? My fear is we won't be able to work without tokens and whoever holds them has all the power.
Working with AI makes me marvel at the human mind. This isn't new, I taught Anthropology for years, but I see the mind as even more amazing than before as the amount of input we take in at any given second would melt any current LLM (even the secret government models). We take in all senses, notice body language, can instinctively "feel" in ways that computers still can't.
I have dealt with more outages in the past year than the past ~5 combined. I have experienced website outages when I ran enterprise sites but they were few and far between compared to now where LLMs powering our tools seem to go down frequently (looking at you again Anthropic).
All to Say
It's been a fascinating year in AI. My team has been great at taking on challenges and it's inspiring to see them learn daily. The technology entices me regularly as I can't stop reading subreddits, watching videos, and simply experimenting nightly with AI. Yes, I have run large Fable experiments while watching movies.
I have no idea what the next year will hold, but if it's anything like the past, it will be filled with new experiences that I welcome.
