Thinking out loud

For small tasks, a 15-minute timer works great. Decide what you're going to do with the 15 minutes, commit to finishing it in that window, and you knock it out.

That method falls apart for real deep work. I recently spent three hours on one task, start to finish, no restarts. It felt different than usual, mostly because it's rare now. I'm convinced I was more effective working it straight through than I would have been picking it up and putting it down five separate times.

AI makes this harder to protect. Jumping between AI tools and tasks all day is genuinely useful for a lot of work, but some work still needs a long, uninterrupted block of human focus, and that kind of block is exactly what a day full of quick AI exchanges tends to erode.

Before the week starts, I name a main focus for each day. That way nothing important gets neglected just because it wasn't the loudest thing that day. A big project can get two full focus days while two or three other priorities still get real attention across the rest of the week.

It's basically production scheduling applied to your own time. Lock in the near-term schedule, then treat that lock-in as real. Some cleanup and reactive work will always bleed into a focus day, and true emergencies get to bump the plan.

The part that matters is the bar for what counts as a bump. If things that aren't actually emergencies keep knocking your top projects off the schedule, you're not planning your time anymore, you're just running a reactive department. Protecting the lock-in is what keeps you steering instead of getting steered.

What's actually the bottleneck? It depends on where things are really waiting. If the customer is waiting, that's an operational bottleneck. If operations can clearly handle more, maybe the constraint is sales. And if the business is pointed in the wrong direction entirely, it doesn't matter how hard any one team is working. You can't rerun the business under a different decision to find out, so effort alone can't tell you where the real constraint is.

What can tell you is imbalance. It works like inventory: too low and things break, too high and finance gets upset. Either extreme produces its own kind of friction, and that friction is the signal. The goal is building systems that amplify the under-represented signal instead of letting it go unnoticed until it's a crisis.

This is where AI agents earn their keep. Have one research the biggest risks to an idea ahead of time and pre-plan the pivot for each one. Once you already know what you'd do if a given risk hit, that risk stops being nearly as risky, and that's what actually opens room for bigger bets. Visibility into your real constraints, not just faster execution, is one of the biggest levers in a business.

Dumb AI and smart AI use the same model. The difference is context.

For a long time I wondered why my AI outputs felt generic. Then I realized I was giving it generic inputs. Not because I meant to. Because I hadn't built the structure to give it the right context automatically.

Now I have a memory system. Files that tell the AI who I am, what I'm working on, what good looks like for me, what I care about and what I don't. Every prompt loads the right context without me having to re-explain myself.

The difference between an AI that feels like a tool and one that feels like a collaborator is almost always this. Not the model. The context.

You can't verify what a company does with your data. You can't audit their servers. You can't read their internal policies. You have to trust something. What I trust is incentives.

If a company's business model depends on customers continuing to use and pay for their product, breaking that trust is supposed to be expensive. Except sometimes it isn't. Facebook's user base grew 13% the year after Cambridge Analytica. Revenue went up 49%. The stock fully recovered.

Turns out the incentive only bites when leaving is actually an option. Nobody quit Facebook over data misuse because quitting meant losing every conversation, every photo, every person you'd have to re-find on some other app. The exit cost more than the betrayal did.

For most of the software you actually choose between, switching costs an afternoon, not your identity. That's exactly when incentives start doing real work. Not blind trust. Not naive trust either. Trust that only means something when you could walk away and the company knows it.

Vague goals don't fail loudly. They just quietly never happen. So I changed how I set them.

Before I plan what I'm going to do, I write what I want to be able to say I did. Not goals. Accomplishments. Goals are forward-looking and vague. Accomplishments are past-tense and specific. "Grow the business" is a goal. "Closed our first enterprise customer" is an accomplishment.

Once I have the list, the work becomes obvious. I'm not deciding what to prioritize, I'm reverse-engineering what has to happen for each of those stories to be real.

The uncomfortable part: you have to commit to a story before you know if it's achievable. I've had years where December arrived and I hadn't done what I wrote. But I still got closer than I would have.

The best AI decision rule I've found came from a comment, not a consultant: "If you wouldn't let a new, keen intern do the task, don't rely on AI to do it."

Don't use AI to respond to a frustrated customer. Do use AI to sort and route the incoming tickets. Don't use AI to write the proposal for your most important deal. Do use AI to draft the first version from a standard template. Don't use AI to decide which opportunities to prioritize. Do use AI to pull the data so you can make that call yourself.

The intern isn't the limitation. The intern is the standard. If you'd trust someone new, with good instructions and some oversight, AI can handle it. If you wouldn't, keep your hands on it.

I have an AI note taker in every meeting. I don't read the notes. That sounds like it defeats the purpose. It doesn't. I use it to pull the action items. Most of the time, what actually happened in a meeting matters less than what needs to happen next.

Here's the mental model I use when I train people who've never delegated anything to AI before. Think of it like a smart kid just walked in to help you. They can write, research, work in Excel, whatever you need. But they know nothing about your company, your situation, or what "good" looks like for you. You can outsource your thinking to them. You can't outsource your understanding.

A real example from a couple weeks ago: I needed to schedule two meetings with my manager after a 1:1. I told the AI what the meetings were about, and it checked both our calendars and found times that worked. It didn't just book them. It made me approve each one individually before it touched the calendar. Most of the value isn't in the AI acting alone. It's in how much faster the boring part happens before you make the actual decision.

This is the framework I'd run starting any admin or coordination-heavy role. Not a slow ramp-up. A deliberate push to become hard to replace in the first few weeks.

Join an AI notetaker to every meeting, since nobody likes training you on the same thing twice. When something is ambiguous, ask directly how you'll know if you did a good job, what good actually looks like. In spare moments, work a chatbot for context on the industry and, where you can get it connected, the internal processes too. Spend real time understanding the structure of everything around you. That's what makes humans valuable and what makes you effective at delegating to AI in the first place. Push to get AI connected into every system you touch. Have it read meeting transcripts and start handling the work that's obvious from them. Then keep taking on more.

This works because so much of admin work is coordination and process knowledge, exactly what AI is good at once someone points it in the right direction. The person who does this in their first month looks like a completely different hire than the one who waits to be told what to do.