For me, 2025 validated my relationship with AI: I found myself chatting less with bots and working more with “agentic” systems.
Wait, what is “Agentic” AI?
It sounds like technical buzzword soup, but the distinction is actually quite simple—and oh so useful.
Think of standard AI chatbot as a smart encyclopedia. You ask it a question, and it gives you an answer. It waits for you. It is reactive.
Agentic AI is more like a smart intern. Sometimes, a very smart and very quick intern. You don’t just ask it a question; you give it a job. Instead of saying, “Write an email about the project,” you might say, “Look through my last ten meeting notes, find the three biggest risks we discussed, and draft an email to the team summarizing them.”
The agent doesn’t just guess; it goes off, reads the files, “thinks” about the criteria, plans the email, and presents you with a result. It plans, executes, and solves problems to get you an outcome, and not just a paragraph of text.
Keeping it Personal
I have always believed that the true value of generative AI doesn’t lie in the “black box” of its training data, nor in its ability to search the web for generic answers. The real magic happens when we use it to reason over our own context—the information we provide it.
My main goal this year was to double down on that belief by keeping things personally grounded. I wanted to keep more of my data close to home rather than sending it out to the cloud. So, I focused on using these agentic models to organize my own library, manage my notes, and even build my own little software tools.
Here is a collection of the specific ways I used generative AI this year, from simple daily helpers to fun useful projects.
The Decipherer
For the jumble of the physical world—scribbled meeting notes, diagrams, and random brain dumps—I turn to Gemini.
Its vision capabilities feel almost surreal, grounded in deep context awareness and a profound understanding of the world. It doesn’t just “read” handwriting; it seems to understand the intent behind my messy scrawl, turning ink into text I can actually search and edit. It’s my favourite way to bridge the gap between my paper notebook and my laptop, when I’m not using my beloved e-ink tablet.
The Librarian vs. The Analyst
My digital archive is pretty big, so I use two different “librarians” to help me manage it, depending on what I need at that moment.
- Microsoft Copilot 365 is the careful archivist. I use it for internal stuff on the work environment—policy guidelines, minutes, shared notes, emails etc.. It drafts texts based strictly on the facts I already have access to. It sticks to the script and helps me keep the details accurate. Copilot Researcher scales up this way of working even further by ingesting huge amounts of files, chat messages, notes … into one, very well-written outcome.
- Gemini is the creative connector. When I want to find a theme connecting different ideas, I upload a specific set of documents in Notebook LM or use Deep Research. It reads through them, finds connections between files that I hadn’t noticed myself and surfaces new information from the web.
The Digital Vault (Local AI)
Privacy feels good, especially for things like financial records. I have plenty of paper—receipts, bank statements, recipes from cook books—that I want to digitise, but I prefer keeping them off cloud servers for peace of mind.
For this, I use Ollama on my desktop. It’s a friendly interface that runs Small Language Models (SLMs) completely offline. I use it without the interface, but plug it into self written tools. However, the outcomes are the same. I feed it a stack of scanned receipts and it tidies the numbers into a clean spreadsheet format with automatic, logical categories and metadata. The data never leaves my computer; I could unplug the internet, and it would still work perfectly. It feels private and secure. Just brilliant.
The Glue (MCP)
A really helpful addition to my workflow this year is the Model Context Protocol (MCP). It’s basically a standard way for AI (language models) to talk to my other apps, which saves me from a lot of searching, copying, pasting and other tedious manual jobs.
I use it in two main ways:
- Smarter Searching: I connected Claude to Zotero (where I keep my research). Now, instead of hunting for exact keywords, I can search for ideas semantically. I can ask for “papers that talk about privacy concerns,” and it finds them, even if the word “privacy” isn’t in the title.
- My Own Podcast: I connected Readeck (my bookmark manager) to Kokoro, a lovely text-to-speech model that runs on my computer (again, no cloud dependency for me here). It reads my saved articles aloud with surprising naturalness. It turns my reading list into a playlist to listen on the go.
“Vibe Coding”
And then there is the most fun part: building my own little and not so little anymore tools using Claude (code), Google’s AI Studio and Github Copilot.
People call it “Vibe Coding.” I don’t write the technical code; I just describe what I want—the “vibe” or the function—and the AI writes the code for me. It’s like being the architect while the AI handles the bricklaying.
My best and latest creation so far is the “Zotero Vector Explorer.” I wanted a specific way to analyse my research library. It didn’t exactly exist, so I described it, and the AI helped me build it. It’s a precise tool that fits my needs perfectly, and it exists simply because I asked (over many evenings of adding features and fixing things).
The Takeaway
So, that is where I stand in 2025. The shiny novelty of the “all-knowing chatbot” has worn off even more, replaced by something far more quiet and useful: utility.
By keeping my data close and my tools custom, I have stopped waiting for AI to be a creative genius and started letting it be the world’s most efficient assistant. It is less about the “artificial” intelligence out in the cloud, and more about the personal intelligence right here on my desk.
It turns out that the best way to use this technology wasn’t to let it take over, but to take ownership of it. And honestly? That is a vibe I can get used to.