Journal · The AI Architect · 2026-07-22
Best tools to build: work that repeats, keeps the same shape, and is tedious. Leave the judgment calls as tasks.
Best tools to build: work that repeats, keeps the same shape, and is tedious. Leave the judgment calls as tasks.
Every time I sit down with someone who’s building their first internal tool or automation, the same anxiety surfaces: they want it to be smart. They want it to handle the edge case, the weird exception, the moment where a human would pause and think. And I understand why. That’s the part of the work that feels valuable. The repetitive part feels beneath the effort of building something.
But that instinct, if you follow it, is exactly backwards. The urge to automate the interesting part is what causes most of these projects to stall out or quietly fail six months later. The real tension isn’t “can the tool be smart enough.” It’s deciding, honestly, which parts of the work were never asking for your judgment in the first place.
Find the seam before you build
Most workflows are a braid of two different kinds of effort, tangled together so tightly that you stop noticing the seam. There’s the part that’s the same every time: pulling a number from one place and putting it in another, renaming files, checking a status, sending the follow-up email that always says roughly the same thing. And there’s the part that changes: should we approve this, is this customer upset, does this document need a different tone.
The mistake is treating the whole braid as one thing. When you do that, you either automate too much and lose the judgment calls, or you automate nothing because the judgment calls make you nervous.
Try this: take one workflow you do weekly and write down every single step, as plainly as you can, like you’re explaining it to someone who’s never done it. Then go through the list and mark each step with an R (repeats, same shape, tedious) or a J (this needed a real judgment call). Don’t build anything yet. Just look at how the R’s and J’s are distributed. Most people are surprised at how much of the list is R.
Build for the R’s, leave the J’s as tasks
Once you can see the seam, the design decision gets much simpler. The R steps are what you hand to a tool or a script or an automation. The J steps stay as tasks — meaning, a person still touches them, reviews them, decides.
This isn’t a compromise or a “for now” measure. It’s the right shape for the work. A tool that reliably formats, sorts, checks, and assembles is worth far more than a tool that tries to also guess whether something is a good idea. The guessing is where trust breaks. When a tool guesses wrong on a judgment call, people stop using it entirely, even for the parts it does well.
Try this: in that same workflow, build the smallest possible piece first — just one repetitive step, not the whole chain. If it’s copying data from a form into a spreadsheet, automate only that. Leave the “does this application look complete” decision as a task someone does by looking at the spreadsheet. Notice how much lighter the tool feels to trust, because it’s not pretending to do something it can’t.
Let the judgment calls stay visible
There’s a temptation, once a tool is working well on the tedious parts, to keep pushing it toward the judgment calls too, because it’s right there and the next step seems small. Resist this longer than feels necessary.
Judgment calls are valuable specifically because they’re visible — someone made a decision, and if it’s wrong, someone can catch it, explain it, revise it. When you fold a judgment call into automation, you don’t remove the judgment. You just hide it inside logic that nobody’s reading closely anymore. The decision still gets made; it’s just made silently, by rules that were written for a different day and never updated.
Try this: if you’re ever unsure whether a step is an R or a J, ask what happens when it’s done wrong. If a wrong answer is annoying but easy to spot and fix, it’s probably closer to R than you thought. If a wrong answer could quietly cause harm, confusion, or a bad outcome that isn’t obvious until later, it’s a J. Leave it as a task, even if it’s tempting to close the loop.
This distinction sounds simple once it’s named, and it is simple. What’s hard is doing it honestly, workflow by workflow, resisting the pull to make everything feel more automated than it should. In the book I go further into how to spot judgment calls that are disguised as repetition, and what to do when a task starts to feel more mechanical over time. But the place to start is right here: look at what you do, mark the shape of it, and build for the part that was never asking you to think.
Go deeper. The full method is in The AI Architect. New here? Start with the free companion pack, or explore the series.