AI doesn't run in a vacuum. It runs inside a world: the platforms your work lives on, the records that hold the truth, the rules about what happens next. Most AI disappointment isn't a model problem. It's that there was never a world built around it.

Season 1 of Notion in Practice was about AI inside Notion. Season 2 is about building the bigger world that surrounds it: the integrations, automations and structure that decide whether your AI has anything worth thinking about.

Tim Jeffries and Jerwin Parker open the season with the frame they use across every build: a digital operating system has four parts, the world, the work, the players and the rules. Inside that world, work splits into deterministic work, where you know the inputs and the outputs and can write the rule, and judgement work, where something has to think.

This isn't automation versus AI. It's how you build somewhere both of them can do their best work.

Build the world properly and the two compound each other: the plumbing keeps the records true, and the judgement runs on something worth reading. Skip it, and you have a clever model guessing in the dark, and paying a premium to move an email.

In this episode:

  • Why most businesses run an archipelago of disconnected platforms, and what that costs in duplicated truth
  • The two halves of every job (plumbing and judgement), and the three kinds of player that can now do either
  • Onboarding a referral, split task by task, so you can see exactly where the line falls
  • Why deterministic work is cheap, precise and loud when it fails, and why that's a feature
  • A live look at the email worker that reads a whole Google Workspace domain and files every client email against the right project
  • What it means when the system knows more about a job than any individual on the team
  • How small a worker should be, and why one that does five things becomes a mess fast
  • The run log: 10,000 records, and why observability came before the fleet
  • Real numbers: about a dollar a meeting for deep AI follow-up, versus US$1.20 for 741 automation runs
  • What breaks, why bridges collapse, and debugging a worker with a coding agent an hour before recording

Three actionable takeaways:

  1. Break the job into tasks before you choose which player should perform it. Label each task plumbing or judgement, then cast the cheapest player that can do it well.
  2. Keep each worker small and single-domain. Same process, several triggers is fine. Different workflows in one worker is a mess waiting to happen.
  3. Build the log before you build the fleet. If you can't see what ran overnight, you don't have a system, you have a hope.


Resources Mentioned

Coming up this season: the businesses actually running this. Law firms, designers, film people. What worked, what didn't, and what it cost.

Who this is for: founders, operators and consultants who have the AI part working and now need the connective tissue underneath it to be reliable, cheap and visible.

👉 Subscribe wherever you get your podcasts, or on YouTube @NotioninPractice.

Rss Apple Podcaster
Podden och tillhörande omslagsbild på den här sidan tillhör Tim Jeffries and Jerwin Parker. Innehållet i podden är skapat av Tim Jeffries and Jerwin Parker och inte av, eller tillsammans med, Poddtoppen.