Hey friends,

There's a genre of video doing the rounds right now. Claude killed lawyers today. Claude killed designers today. Claude killed coders today. Claude, apparently, has been killing everyone.

I have spent more hours with Claude this year than I'd like to admit, and it has never once seemed that mean to me. It doesn't have a dark side. What's actually happening is people hyping up what the tools can do, and the tools genuinely can do a lot now, so the hype has somewhere real to stand.

I'll tell you where I was before I sat down to record this episode. Lying on my couch, trying to take a nap, because I'd been up since 3am. I've become a bit of a Claude addict. It's an actual term now, I think. And the reason is simple: I can get so much done that stopping feels harder than continuing.

Last year, the pattern was ask a question, get a response. That was the whole relationship. You typed, it replied, you copied the good bits out.

The intelligence behind the models hasn't fundamentally changed this year. What changed is what got built around them. If you download Claude desktop, there's an option called Cowork. You point it at a folder on your computer and say, go through these files, give me a summary, build me a report, make me a proper annual document. And it does the analysis and the actual work, not just the writing. Think of it this way: any work you can do on a computer, Claude can now help do on that computer. There's a connector that drives your Chrome browser. There's computer use that executes tasks directly. It's moving into Excel and the rest of the Microsoft suite too.

Then there's Dispatch. If you leave your computer on at home and you're driving back from a meeting, you can say, I just met this person, go check the meeting transcript, make sure the proposal is ready before I walk in the door. And it starts working while you're still on the road.

So I've stopped thinking of it as one assistant. I've got a crew now. Multiple versions of Claude, each doing a piece, and me checking their work. I come back, look at what they've produced, and say right, now you do this, and you do that. Most of my team runs on Claude APIs directly, but the desktop version has become powerful enough to carry a lot on its own.

My mother-in-law is an English literature professor in India. She took leave to write a book, and she asked me to help with the research, because I've got all these deep research tools. I knew what would happen next. She'd be coming back to me every day with questions.

So I built her an agent. I called it Cuckoo, which is her family name. I told it plainly: this is my mother-in-law, she's brilliant on literature and knows almost nothing about technology, she doesn't know how to talk to you, here are the six books she's working on. Then the important part. I said, don't just assist her. Teach her how to get the most out of you as you go.

Cuckoo messaged her on WhatsApp. Hey mama, I'm Cuckoo, I'm your assistant, made by Aamir. She thought I was pulling a prank. Aamir, is that you? It took a phone call to convince her it wasn't. But once she understood, she started asking for five-point summaries of her books, and Cuckoo would give her the summary, then gently offer more. What would really help, mama, is if I also did a fuller synopsis. Here's what I could do next.

That last part is the piece most people miss. Claude sitting underneath is genuinely good at understanding who it's talking to and then shaping itself around that person. It met my mother-in-law exactly where she was.

This Valentine's Day, I gave Sarah an agent. I figured, what's more romantic than giving your wife something that can do all the things she doesn't want to do herself.

The agent was called Miro. I connected it to our family calendar and the emails where all the school information lands. Miro introduced itself, told Sarah this was her gift, and asked to get to know her. What bothers you? What drives you crazy? Sarah took it as an open invitation to vent, which, fair enough.

At one point she mentioned the guest vanity had been broken for two years, and that she'd been asking me to fix it for roughly that long. Miro went to Airtasker, posted the job, found a plumber, negotiated him down on price, and came back to Sarah with, here's your plumber, want me to book him in? She said yes. The plumber came. The vanity is fixed.

Here's the honest part of both stories. There's now a real dependence on these agents at home. Two or three times, when the credits have run out, I've had the phone ring immediately. Cuckoo isn't responding. Miro's gone quiet. I need more tokens. Tanisha does the same with her content agent, Inky.

Which is worth sitting with for a second. These are business models. They're built to be so useful that you use them more, and then you pay more. The subscriptions are heavily subsidised right now, so it feels almost free. It isn't. Every message, coming and going, is compute that someone is paying for. If you're not paying for the product, you tend to be the product.

I bought Claude subscriptions for my project management team, and I knew straight away they wouldn't get value from it without setup. So before I handed it over, I built the scaffolding.

The first thing I built was Skills. A skill is basically a simple text file that gives Claude domain knowledge and tells it how to do a specific thing. My first one was an Enterprise Monkey brand skill: logos, tone, the whole voice. Now every document that comes out is properly on brand without anyone thinking about it. I added a humaniser skill, which you can find in a marketplace and drop in. Then an Enterprise Monkey proposal skill to help the team build proposals.

Alongside that, I built a brain. Think of the brain as a very curated folder, in Google Drive or SharePoint, full of clean information about the company. Inside it sits a readme that tells Claude how to move around. If the question is about a person, look here. If it's about this, go there. Clear instructions, nothing clever.

The thing worth understanding: the brain and the skills live outside Claude. They're mine. I've uploaded the skills into Claude, but they exist independently. All the magic that looks like it's happening inside Claude is actually happening outside it, and I'm just giving Claude access. That's the whole trick, and it's why it survives whichever model is winning this month.

To finish the setup for my team, I wrote a custom instruction: this is my team, they don't yet know how to use you, so help them; if they need information, use the brain; if they need a document, use the brand skill. Then I set up a scheduled task that sends each of them a daily brief pulled from their emails and correspondence: here's what needs attention today, here are your meeting minutes. A couple of hours walking them through it, and they were away. It felt a lot like the Jarvis idea from Iron Man, except it's real now.

Plenty of businesses can't just switch to Claude, because they've standardised on Microsoft Copilot for security and licensing reasons. One of our accounting clients put it perfectly: everyone is asking for Claude, everyone's asking for these tools, what do we tell them? The catch with Claude is that your data travels to their servers, outside the boundary.

So here's how I'd frame the decision, in three windows.

First window: your data legally has to stay in Australia. That's a hard no on the Claude features. Use Copilot on its own. If you're in healthcare or working closely with government holding personal data, this is you.

Second window: you care about security but you're not legally obligated to keep data onshore. Copilot has introduced Claude inside it, and there's a Cowork function arriving in their Frontier program, which is a beta you sign up for. You get a lot of the Claude capability while staying inside Microsoft. Worth saying plainly: with Claude the concern isn't a breach, it's that your data travels to US servers at all.

Third window: you're security-conscious but genuinely flexible. Maybe you hold important IP, say a manufacturing business with a novel process, but it isn't national-security sensitive. Claude has committed not to train on your data on the enterprise plans, and its security controls are strong enough that a lot of small and medium businesses can use it safely. For larger organisations there's the enterprise version.

One question cuts through most of the anxiety here: is data residency actually a requirement for you, or does it just feel like one? Plenty of businesses insist their data must stay onshore while already running HubSpot, Make, and a dozen other tools that send data offshore every day. The data's already out there. Be honest about which window you're really in. We did a whole episode on AI governance, and if this is live for your team, that's the one to go back to.

Here's something that surprised me. Claude's integration with the Microsoft stack is, in practice, better than Copilot's own, and it happened fast.

The reason is that Claude focused hard on one thing: being genuinely good at writing code. And nearly everything that happens on a computer runs on code underneath. So to make a Word document, Claude doesn't need to know how to click around Word. It writes the script that produces the document. Because it reads and writes code fluently, it can work inside almost any software by understanding the code beneath it.

Step back and there are really two shifts that got us here. First, machines learned to understand human language, which made them easy to teach. Second, they learned to read and write code, which lets them actually do things on a computer. Understand language well, operate a computer well, and the gap between wanting something done and it being done closes right up.

We should talk about cost, because it deserves its own conversation and most people are avoiding it.

The subscriptions are heavily subsidised. The usage you get in a thirty-dollar Claude or ChatGPT plan would often cost the provider a hundred dollars or more to deliver. They're funding that gap with raised money because they want as many people using AI as possible. The underlying cost of tokens, a token being roughly two words, hasn't really come down.

I'll give you my own number. One month, my AI bill came to around fifteen thousand dollars. APIs, consumption, all of it. I have a picture in my head of myself sitting there with a little fire, just burning money. But I was learning and building while it burned. On one weekend, across two late nights, I built an app. I asked my team, a genuinely skilled team, what that would have taken them. Three and a half months.

So here's the uncomfortable forecast. In a year, the twenty-dollar subscription might be a five-hundred-dollar subscription as the subsidies thin out. But that's exactly why now is the time to treat AI as a playground. Experiment, fail, burn tokens while they're cheap, and learn where the waste is, because right now, when you're building and designing, the waste is enormous. Better to learn that on subsidised pricing than discover it when the real bill lands.

It became fashionable to say Google fell behind on AI. I'd be careful with that.

Go back to the origin. A few people at Google were trying to fix Google Translate, which could swap words one at a time but couldn't connect a sentence. Their research became the Transformers paper, the idea that you read the whole document, take the words, and predict what belongs around them. OpenAI picked up that published paper, trained it on effectively the whole world's knowledge, and the large language model was born. The concept the entire field runs on came out of Google.

What Google also has is depth. Its AI is quietly moving into Gmail, into Maps, into all the things we've stopped noticing. Professor Raj Basser at Deakin has a line I love: if you notice the technology, it usually means it's broken. Google has become so woven into the devices we use that we don't notice it at all. And its ad revenue is rising, partly because as search shifts, businesses feel more pressure to advertise, and partly because its reach was never only about search in the first place.

There's a practical thread in here too. Claude doesn't have a direct connector for Google Ads, but there's a tool called Composio that connects through the MCP protocol, which is just a standard way to plug Claude into third-party systems. Through composio dev you can connect multiple Google and Meta accounts, then simply ask, look at my campaign, what am I doing wrong, want to set up a new one? And it can build the campaign for you. On the Meta side, the official advice now is to stop over-thinking your targeting. Give it your offer, work hard on the creative, and let Meta's AI find the right people. It gets it remarkably right.

The honest read is that Google and Apple are, to some extent, watching. Features Claude has shipped in weeks, their teams could build in weeks too. Claude is establishing its presence and will earn its own following. But nobody in this race is asleep.

Now the part I've been wanting to share.

I've been working with a small group of people on a project called Agents for Humanity. So many of the world's hard problems sit unsolved: how do you educate someone in rural India who needs quick outcomes, how do you get people into homes they can actually afford in Australia, how do you plan a city, how do you move the needle on hunger. There's no playbook for most of them.

Agents for Humanity is an open platform. Any not-for-profit, any individual, can put a problem into it. Then hundreds of agents take on defined roles and work the problem through a clear design-thinking process. They research it, debate it, steel-man each other, and drive toward a synthesis. A council of agents votes on the strongest proposals, a few are chosen, and from there an implementation plan gets built to actually solve the thing.

The project has been selected as one of the top five in Australia for the AI for Good pitching competition, which is supported by the United Nations. So I'm heading to Geneva. It started as a 3am, harebrained idea, which tracks.

There's one real constraint: compute. Solving a single problem properly costs somewhere between six and eight thousand dollars of compute. So the design is that you, as someone already paying for Claude or ChatGPT, can send your agent in to contribute a slice of work and come back, using capacity you're already paying for and mostly leaving idle. Pool enough of that, and we can work through hundreds and thousands of problems. Six of us have been building it, testing with five organisations across India and Australia, running on our own compute for now, with good commitments of compute coming from some universities. And the final synthesis will be open to the world, so a solution found in Kenya is available to someone in India.

I get asked constantly what's coming next, and the honest answer is that nobody can tell you which model wins or what ships next quarter. We're in the most uncertain stretch I've ever worked through, and everyone is guessing, including the people you'd expect to know.

But a few things I'll say with confidence. We are never going back to a world without AI. The toothpaste is out of the tube. This isn't a bubble waiting to pop, even if valuations move around. The way business works has already shifted, and it isn't shifting back. There will be displacement in the job market. The open questions are how much and how fast, not whether.

So when my kids ask what career to build in a world like this, I tell them the same thing I'd tell you. Focus on things that help people. Focus on where you actually add value. Focus on what you enjoy. Stay curious, and make peace with not knowing everything, because none of us will.

What keeps me hopeful is the same thing that's always kept me hopeful. We're building something genuinely powerful, and it won't be perfect, because we aren't. But it can solve problems in ways we simply couldn't before. If all of us, individually and then together, keep remembering that we're people first, I think we're going to be fine.

Stay curious. And go ring your mother-in-law, not just her agent.

See you next week,

— Aamir

🎧 Listen to the Podcast Episode on: Spotify | Apple Podcasts | YouTube

📘 Book: The CEO Who Mocked AI (Until It Made Him Millions) by Aamir Qutub

🚀 Dumb Monkey AI Academy

📱 Dumb Monkey AI Academy App: Apple | Android