It's been a while. How are you? I got a little hyper-focused on my MSc studies in organisational psychology. And while we weren't watching, AI has had a bit of an adulting phase. Now, by mid-2025, the models got a bit better, cheaper and faster, but most of all - easier to use. Over the next three months, I'll send you an update every two weeks, so we can catch up on all that has happened and what it might mean for you.
Think back to summer 2024. AI was exciting, sure. There were lots of new things popping up everywhere, and yet nobody really quite figured out what we wanted to do with it yet. AI was this thing that we'd open a chat box, type things into, and get answers from. Maybe it would make a slightly awkward image or help us write an email, but it was very much a pull application. You ask, and it delivers an answer.
We'd do things one at a time, probably write prompts one at a time, sort of on the go. Over time, we learned how to write better prompts, how to use techniques like giving the AI a role: "act as a marketing specialist," for example. And we played. We experimented. But that was about it.
Now, in the last eight months, something fundamental has changed. AI is moving from being a curious experiment to becoming essential business infrastructure. The numbers tell part of the story: 78% of organisations now use AI in at least one business function, up from just 55% last year. In 2025, IT leaders are allocating approximately 20% of their technology budgets to AI initiatives.
The First Steps: Projects
The most basic version of this evolution that's been popping up everywhere is what I call "projects". They have different names on different platforms. In ChatGPT, they're called CustomGPTs (and now also projects). In Claude, they're Projects. In Gemini, they're Gems. And Copilot calls them Agents. (slightly confusing as we'll learn soon). The idea is simple but powerful: instead of writing prompts from scratch every time, we save more specific, detailed prompts that we can reuse.
Say you might have a preset chat for email editing. Another one for drafting blog posts. Yet another for brainstorming ideas or breaking down tasks into small, actionable chunks. You can spend extra time writing a prompt that helps instruct the AI to do things the way you like them done. You can attach files, PDFs, databases, maybe even live data from a customer database or email to inform how the AI works on that prompt.
For a copywriter, add example copy or a style guide, so you don't have to add this information manually every time. Now you have specialised collaborators that you can work with for different tasks. You go to a different project, a different gem, a different custom GPT, and you shortcut a lot of that prompting and context setting because you've saved it ahead of time.
Learning walk: Canvas
Another useful extension to the basic chatbot, pioneered by Anthropic and quickly adopted by every major AI platform, was the canvas. Rather than going back and forth, we now have the ability to work on a document collaboratively. Even if you didn't do it on purpose, you probably saw the chat split in two, and a document appear on the right. In Claude, that's often referred to as an artifact, ChatGPT calls it a canvas. Now you can refer to the document, your AI can make edits, pass it back to you, etc. If you're using Microsoft Office, it's almost the other way around: the document is your regular Word document, and the chat pops up in the CoPilot sidebar. Either way, having a virtual colleague work on a document with you can be extremely powerful.
And if you haven't tried it yet, the canvas features are amazing once you ask the AI to code something. Have a boring spreadsheet? Export it as CSV, or another format your AI understands (it will tell you), and ask it to turn the spreadsheet into a data visualisation, if you haven't yet. Or need to visualise a quick idea for an app or feature? Ask to get a quick prototype.
Now that we've taken the first steps, let's recap.
What "Adulting" Actually Means
This idea of grouping together information and instruction to be reused later takes a general AI and makes it more specialised for specific tasks. It's the difference between having a helpful but generic assistant and having a team of specialists who understand your business, your style, and your needs.
Where it gets really interesting is having those agents interact with each other. In software development, there has been a lot of progress around how agents can work together. You might have whole armies of agents working on drafting software solutions to solve specific bugs or feature requests.
Picture this: once a bug report has been filed with an example and references, the first agent might pick up that ticket and figure out what exactly the problem is and what kind of knowledge and tools it needs to solve it. The second agent might try to replicate the bug, see if it can make the bug happen on its own, so it can later test the fix. Once the bug has been replicated, a third agent might pinpoint where exactly in the codebase that bug is created. Once it finds that, it might pass that file to yet another agent that's trying to fix it.
Finally, a last agent might test the fix and see if the bug that was reported in the first place is solved when the code change is implemented. Once that's all done, it might write a bit of documentation and submit the change to the codebase for human review.
Suddenly, we have a whole team of different specialised agents working together, collaborating, going back and forth between themselves depending on what they deem the next step should be.
The January Watershed Moment
The turning showed its first signs in October 2024. That's when Anthropic introduced something called 'computer use', teaching Claude to view screens, type, move cursors, and execute commands. Around the same time, Google was developing Project Jarvis, an AI agent designed to control web browsers and complete everyday tasks.
This wasn't just an incremental improvement. This is AI learning to stop waiting for us.
By January 2025, this shift became undeniable. OpenAI launched Operator, an AI agent that could independently navigate web browsers to complete everyday tasks. This moment marked what many are calling "the official beginning of agentic AI."
Around the same time, Google DeepMind unveiled Gemini 2.0 Flash Thinking, setting new highs in mathematical and scientific reasoning with a massive 1M token context window. Last year's models could grasp the nuances of a few dozen pages, remember key plot points or arguments, and answer questions based on that limited scope. If you gave it a whole novel or even just a long PDF report, it would only be able to "remember" the most recent few chapters and would struggle to connect themes from the beginning to the end. This year's models process and understand an entire multi-volume encyclopedia. They can cross-reference information between different volumes, synthesise knowledge from diverse topics, and answer complex questions that require a holistic understanding of a vast amount of information.
At this point in time, there's very little difference between the big frontier models. Google, OpenAI, and Anthropic are falling over themselves to launch the next better version, mere days faster than their competitors. Now, all have huge context windows, allowing for a lot of contextual information to be digested. They can all browse the internet for us, even do "deep research", where they take a simple question we type in the chatbox and then launch several "research agents" to Google for us, make connections, and argue which sources are useful or accurate. Once they have enough information, they compile a well-grounded report with citations. Suddenly, research that would have taken days takes 5-10 minutes.
From Pull to Push: A New Relationship
In our relationship with AI, we're moving from us typing things into a box and getting an answer, to us saving prompts and steps that we do often so we can reuse them more quickly, to eventually having different pre-made agents working with each other, even using tools like web browsing or taking screenshots, making airline bookings on our behalf.
As we're moving through 2025, I expect a lot more of this to become embedded in how we work. As the technology gets better, it's back to us to update our relationship with AI. We're moving from a world where AI waits for our instructions to one where AI can take initiative, plan ahead, and work alongside us as genuine collaborators.
Your Homework: Start Small, Think Big
So here's what I invite you to do. Think of it as a little homework to explore how you might collaborate with AI agents, but also how you want to see yourself evolve in this new landscape.
Look at your day-to-day work and start identifying tasks you do repeatedly. These might be simple admin tasks like screening emails and using a tagging or filing system. Maybe you get inquiries about specific features or products you sell, and you find yourself answering similar questions again and again.
Examine the steps in these processes. How do you answer those emails? Are they highly personal, or are they based on information you can save into a PDF?
Try building one of those custom GPTs, Claude projects, Copilot Agents, or Gemini Gems. The platform doesn't matter much. But do give it a try. Write up the process. Give your AI an instruction like: "I will share an email with a customer inquiry. Act as a personal assistant. Use the attached information about our products and services to draft a response answering the customer inquiry and offering a one-on-one meeting to discuss further."
Then attach a few documents that are key to answering routine inquiries—product fact sheets, your website saved as PDFs, whatever information you typically reference.
See how it feels to use this to draft your first email rather than writing it by hand. Is it adding something? Is it saving time? Is it giving better responses? Is it worse? Maybe there's information missing—is that because you didn't attach it, or because it's something only you know that isn't documented anywhere?
Notice which kinds of processes or tasks in your life are easy to turn into these little projects, and which parts need you specifically. How might that inform how you think about what you do in your role?
If the custom agent doesn't quite work the way you want, copy the prompt and get your AI to help you make the prompt better. Open a new (normal) chat and say:
Act as a prompt engineer. Help me refine a prompt for a [custom GPT/Claude Project/Gem/Agent]. I pasted the current prompt below. Please read it and critique it, then ask me at least three questions, one at a time, to understand my process and context. Once you have enough information, draft the best prompt you possibly can for me to use in my [custom GPT/Claude Project/Gem/Agent].
< -Current Prompt- > [paste your hand-made prompt here]</ -Current Prompt- >
The Bigger Picture
Over the next week we'll build on this experiment with projects/custom GPTs and start exploring:
- What agentic AI actually means and how we might think about it
- How to stay smart in a world that does all the thinking for you
- What prompt engineering and systems thinking have in common
- Measuring what matters, where we revisit Metrics for a GenAI World
- How to say yes to both, technology and humanity
This is the first post in a six-part series exploring AI's transformation from hype to practical implementation. In a couple of weeks, we'll dive deeper into the rise of autonomous agents and what happens when AI truly stops waiting for our instructions.
References
Where I got the stats cited above:
State of Generative AI in the Enterprise 2024, Deloitte