Pull, Not Push: The Factory-Floor Idea That Fixes AI

Most AI breaks for a reason that has nothing to do with AI. It breaks because it's built as a push system, and push systems have always been fragile. We just forgot, because the last time most of us thought about it was never.

Let me explain with the thing I spent years on before any of this. I'm a Lean Six Sigma Black Belt. In operations, there are two ways to run a line. A push system runs on a fixed schedule: do step one at this time, step two at that time, regardless of what's actually happening. A pull system runs on real demand: the next step happens when, and because, the situation calls for it. Toyota built an empire on pull. Push systems pile up waste and jam the second reality shifts. This isn't new-age. It's the most boring, proven idea in manufacturing.

Now look at how almost everyone builds AI. "When a new lead comes in, send email A. Three days later, send email B." That's a push system. It fires the same step no matter what the lead actually did, no matter what's going on. And the moment something doesn't match the script, the whole thing either does the wrong thing confidently or breaks and lands back on your desk. Worse, it still burns the tokens and costs you money while it fails.

I built plenty of those when I started. They felt like progress. They were actually just brittle little conveyor belts I had to stand next to.

So I rebuilt for pull. Instead of "always do X at time T," I told my agents how a thoughtful teammate would decide, and let them read the moment. A push system would send every prospect the same day-three email. My team looks at what each one actually did this week and decides what they need. One adds noise. The other takes work off my plate.

---

Here's the receipt. I had a meeting end on a weird note, the kind you can't quite name. I didn't ask anyone to do anything. My chief-of-staff agent read what happened, worked out where it went sideways, found the opening to pivot, and drafted the recovery email. I read it, made one tweak, and sent it. No rule fired that. She read the room.

And this is where it gets bigger than any single agent. When your agents can read the moment and hand work to each other, you get something a push system never produces: a system that routes around its own problems. My marketing lead pulls in my chief of staff when she needs her. They sort it out and bring me the answer, not the question. There's a name for this in the research, a decentralized system where no single node is the bottleneck and the whole thing keeps working even when one part stalls. But you'll feel it before you can name it: the day the work starts moving without you pushing it.

---

The deep point is this: AI doesn't fail in the real world because the model is weak. It fails because the architecture puts all the judgment on you. Push keeps you as the brain the system can't run without. Pull distributes the judgment into the system, so it can finally carry some of the weight. That is the entire difference between AI that exhausts you and AI that frees you.

Try this this week: find one automation or prompt-chain you've built that runs on "always do X." Rewrite it as a decision instead of a rule: - Push version: "Every Monday, summarize my inbox." - Pull version: "Each morning, look at my inbox and surface only what actually needs me today. If something's urgent, flag it now. If it's noise, leave it."

You're not writing a smarter schedule. You're describing how a sharp teammate would judge the moment, and trusting it to judge. That's the whole move. And you don't need a single line of code to make it.

Pull, not push. Read the room, don't run the script. It's the oldest good idea in operations, and it's exactly what your AI has been missing.

-Jonelle