The seatbelt sign was still on when I opened the laptop. Early April, somewhere over Europe, on the way to Athens with the family. I’d promised myself I wouldn’t — the whole point of the week was the people two rows back, not the backlog. And there I was anyway, screen tilted slightly away, doing the one thing you are absolutely not supposed to do on a family holiday.

I opened it because I was frazzled and behind, carrying a stream of work that never drained no matter how many hours I poured into it. The worst part was never the volume. It was that the volume was crowding out the exact thing I’d built this life for — time with them, the reason you get on the plane at all.

I’d watched a video before we boarded, and I want to put the credit where it belongs before I go any further.

The idea wasn’t mine. It came from one of the personal-AI-team videos doing the rounds in early 2026, and I’ve tried and failed to re-find the specific one. I mention it anyway, imprecisely, because I’d rather credit a stranger badly than quietly keep their thinking and pass it off as my own.

What they were saying was simple, and at the time it didn’t even feel original: stop using AI one chat at a time, and build a team that works for you instead. You’d have a coordinator to dispatch the work, a researcher to go and dig, someone whose whole job was thinking about what the team itself should become. Plenty of people were saying versions of this. I want to be honest about that, because the part that turned out to matter wasn’t the part everyone was already saying.

So I gave myself permission to run a small experiment — from my phone, from my own laptop, in the gaps around the sightseeing. City heat, a hotel room, half an hour stolen here and there between one thing and the next. The tension is half the story. I was trying to build a machine to give me back my time by spending some of the most precious time I had, and I’m not sure I entirely got away with it.

The other half is that the experiment didn’t work the way the video promised. It worked because of two things the video never mentioned.

What flipped it

The first I noticed over dinner.

In the early days I’d been giving the team careful, step-by-step instructions — do this, then this, here’s exactly how. I was operating it, really, one keystroke removed. But somewhere in that week I’d started doing less of that and more of just setting the goal and letting the team work out the how. And once I’d crossed that line, this became possible. I kicked a task off from my phone before we sat down — nothing dramatic, something off the list I’d normally have had to grind through myself on a Sunday night. By the time dessert arrived, it was done — not half-finished, not a draft waiting on me, just done. And then another one. The to-do list was burning down in front of me, and I wasn’t the one holding the pen.

That was the acceleration nobody tells you about. The work got done without me doing it, so I could point the team at something over a starter and come back to a finished thing over coffee. And what made it possible was the shift I’d half-made without noticing: I’d stopped writing the script for every step and started just telling it where to end up. The AI was doing the work, and doing it while my attention was somewhere else entirely.

The second thing was the one that changed everything, and it arrived when I stopped feeding the team tasks and started feeding it my world.

The first real test was my home lab and the wiki. Instead of asking the AI to reason in the abstract, I wired it into the actual thing it was being asked to think about — the servers, the network, the notes I’d kept on how it all fit together. And it stopped feeling like a stranger. It started handing back things that were right because they were grounded in my setup.

That was the moment — hunched over a laptop I wasn’t supposed to have open — when I realised I’d had the emphasis wrong the entire time. I’d been chasing the novelty of the team, when the thing that mattered was the context. An AI with no context is just a clever stranger; wire it into yours and it starts to feel like it works for you.

And the context wasn’t a one-time upload. It grew. Every note, every decision, every correction I made became part of what the system knew the next morning — a bit less generic and a bit more mine, every single day.

It wasn’t a holiday fluke

The honest test of any idea you have on holiday is whether it survives a normal day.

This one did. The same move — wire the AI into the context it’s meant to operate in — worked just as well at work, in a completely different domain. There it wasn’t servers and a wiki. It was the meeting I’d just come out of, the email thread three replies deep, the transcript of a conversation I half-remembered. Feed it that, let it grow with it, and the same thing happened: it stopped being a tool I had to brief from scratch and started being something that already knew where I stood.

Home was where I proved the principle, and work was where I found out it wasn’t a one-off. What mattered in both cases was never the domain. It was the connection to context, and the fact that the context kept growing.

The part that was new

Here I want to be careful, because nearly everything I’ve described, people were already doing. Agents with roles. A coordinator. Connecting AI to your own data. None of that was new in April 2026, though it felt new enough to me at the time.

The genuinely new move came next, and it had nothing to do with the technology.

It was a change in my own role. Remember the dinner — the team burning down the list off a goal, not a script? That was the tell, and I didn’t read it for what it was at first. If the team only needs the destination and not the turn-by-turn, then I’m not the operator any more — I’m the person deciding where we’re headed.

So I stopped operating the AI and started managing it, as a team, with myself directing it rather than driving it. I wasn’t the one doing the work with a clever assistant beside me. I was the one deciding what work the team took on, who handled what, what good looked like, and what we were for — running it the way I’d run any team of people. You dispatch, you trust the specialists to do their bit, you hold the standard, and then you get out of their way.

That was the thing I hadn’t heard anyone talk about, and it changed everything. It wasn’t a better tool: it was a different chair to sit in.

And once I was in that chair, a question arrived that I couldn’t put back down. If the value was never in the tool — if it was in connecting AI to your context and then directing it as a team — then almost every organisation pouring money into AI right now is buying the wrong thing. They keep asking which tool to deploy. But the tool was never the point. The point is the bit nobody’s selling, which is the human direction.

So if you’re about to go and build your own, let me grab your sleeve for a second, because there are two things I’d want a friend to tell me first.

The first is a warning. The moment the friction to deliver faster is removed — the moment you’re directing a team of your own design instead of operating on your own — everything is in scope. There’s no task too big to attempt, nothing you can’t point the team at, no natural limit telling you to stop. That sounds like freedom, and it is, but it’s also exactly how you end up building for the sake of building — because it’s blooming fun, and somewhat addictive.

Which is why the second thing is the one that matters. Before you build anything, ask where you actually need help. What are you trying to achieve? What’s eating your time? Where could a team of your own help you achieve the most? Not which task to automate — that’s the small question. The real one is where you need help and what you’re for. Because if you’re building it without a purpose, then what’s the point? I started all of this on a plane, stealing time from my family to build a machine that would give that time back. The only honest measure of whether it worked is whether I’m more present in the hotel room than I was in the airport — whether the machine ever handed back the thing I built it for. Build toward that, or don’t build it.

That’s the argument I want to make next, and it’s the reason I built this whole series. So I’ll hand you straight to it.

Stop deploying AI tools. Start helping people build their own.

This is the first piece in a series on building a personal AI team. The next one — “Stop Deploying AI Tools. Start Helping People Build Their Own.” — is the why. And if you want the conversation at the dinner table that sits underneath all of this, that’s in “AI and Careers: What I Told My Daughters.”