Last week, what did an AI do for you while you weren’t watching?

The question sounds silly. It isn’t.

Because pretty much everyone uses AI to summarize a doc or rewrite an email, that’s become standard. But actually delegating a task to it, one that runs while you sleep or sit in a meeting, that’s a different story.

Honestly, I asked myself that question a few months ago, and the answer made me move.

Every morning at 7am, my Mac Mini sends me the brief of my day. Which meetings, which emails to prioritize, which follow-ups still pending, which Granola notes to turn into proper minutes. In the evening, it gives me tomorrow’s reminder. And none of it, not a single byte of my outgoing emails, my meeting transcripts, or my todos, ever leaves the house.

I’m not telling you this to flex my tech credentials. I’m telling you this because I think we have a real posture problem on AI at the executive level.

You still see leadership meetings where AI lands on the agenda, where everyone agrees we need to accelerate, that we’re behind, that we need to bring the teams along, and where, around that same table, very few people have actually delegated a task to an AI assistant in their own week. Me first, for a long time. We hand the topic over to “the AI team”, to a consultancy, to a freshly appointed Chief AI Officer. And we wait for results.

Honestly, I think there’s a prerequisite. Before asking your teams to transform their jobs, you need to have transformed a bit of your own. Otherwise you’re talking about something you’ve never really touched, and it shows.

So, over the past few months, I’ve been building my own digital twin. Basic level at first: CV, reference books, papers I like, prompts for context. And step by step, I’ve been scaling it up.

Here’s what runs on my Mac Mini today. A full analysis, syntactic, linguistic and semantic, of my outgoing emails and meeting transcripts, so it learns how I write (and how I speak, which is not the same thing). Cowork prepares draft emails across my different inboxes, through a homemade Gmail MCP server that technically blocks any actual send: not enough trust yet to let it hit “send” on its own. A daily tech watch. The daily recap of my day and the reminder for tomorrow. The pickup of Granola notes to turn meetings into minutes. And each week, a synthesis of all of the above to enrich its general context.

What’s interesting is that almost all of those tasks, summarizing, extracting, rephrasing, classifying, reminding, run perfectly well on the machine itself. You don’t need a model that writes Shakespeare. You need a model that does one thing well, on data that is yours.

That’s exactly the point of an article I came across recently (unix.foo, Local AI Needs to be the Norm): for the vast majority of use cases, local AI should be the norm and the cloud the exception. Not an ideological stance, an engineering argument. Why turn a simple feature into a distributed system dependent on a third-party vendor, when the silicon in your pocket (or on your desk) can already handle it? For my most complex use cases, I still use Claude through the API. But I’ve gained something: I now know why I’m sending a piece of data out, when I do. The rest, it handles on its own.

And here’s what I want to land on: none of this got built in an hour on a Sunday night. You need to count a half-day a week, minimum, to start picking up real speed. It’s a personal investment, and that’s precisely what makes the conversation credible when you take it back to your teams.

Anyway. I don’t know if my setup is the right one. It’s imperfect, I take it apart and rebuild it regularly. But when I talk about AI to people now, I’m talking about something I’ve lived, not a promise I’ve read.

So, fellow executives, where do you stand on this, personally? Before asking your teams to come along, what did you delegate last week?

PS: post written with Claude, of course :)