Validator, Not Bottleneck

I watched a senior guy demo his team’s new AI workflow a while back, and he was proud of it, and he should have been. Intake, drafting, and formatting all ran on their own. Hours saved everywhere you looked. Then he showed the last step, almost as a flourish: everything routes to him for final review. Every item. And there’s this pause, because half the room is doing the same math I’m doing, which is that he has just described a machine that produces work faster than any human can approve it, and he has installed himself as the drain.

Three weeks later, I’m told, there was a queue. The queue had a nickname.

Everyone in that room agreed on the principle, because everyone everywhere agrees on the principle: you need a human in the loop. It’s the most reassuring sentence in AI, and I’ve written before that it means nothing until someone says where the human sits. But there’s a version of that question I’ve started putting to people directly, because the stakes turn out to be personal, not just architectural. There are two ways to be the human in the loop. You can be a validator, or you can be a bottleneck. Same seat on the org chart. Completely different futures.

The bottleneck reviews everything. It feels responsible, and honestly it feels good, because every single item passing through your hands is proof you’re essential. The workflow literally cannot proceed without you. That’s the trap in one sentence: the bottleneck confuses being unavoidable with being valuable. And while the queue grows, something quieter happens. Because the bottleneck checks everything the same amount, their judgment never gets sharper about anything in particular. Approval becomes a rubber stamp with a wait time attached. And people notice. The system’s owners start asking which reviews actually change outcomes. The answer is usually a short list. Everything not on the list gets rerouted.

That’s the ending nobody says out loud, so I will: bottlenecks don’t get fired. They get designed around. One workflow at a time, politely, until the loop no longer runs through them, and the essential feeling turns out to have been the queue talking.

The validator plays it completely differently. The validator has done the unglamorous thinking about where mistakes in this workflow are actually expensive, and stands exactly there. A customer about to be told something wrong. Money about to move. Anything about to ship under the company’s name. At those points, the validator’s judgment is slow, careful, and absolutely mandatory.

Everywhere else, the work flows, because the validator has said the sentence that bottlenecks cannot physically say: “this part doesn’t need me.”

Watch a good one work and two things jump out. Their yes is fast. Not careless, fast, because they’ve decided in advance what they’re checking for, so a review takes ninety seconds instead of a meeting. And their judgment is written down. What they look for, what fails, what an automatic pass looks like. Which sounds like the beginning of being automated out of a job, and this is the part I most want to flip for you, because it’s exactly backwards. The bottleneck hoards judgment and gets routed around. The validator makes judgment legible, the criteria get built into the system, and the validator gets handed the next loop, and the one after that.

One of them is guarding a toll booth. The other is becoming the person who decides where toll booths go.

That’s the skill. It has an unsexy name, human-in-the-loop design, and the fringe-looking move is learning to apply it to your own seat. Try the audit sometime; it takes ten minutes and it stings a little. List everywhere work currently waits on you. For each one, two questions. What does a mistake here actually cost? And does my looking at it change the outcome, or just add a day to it? Everything where the honest answers are “not much” and “just the day” is bottleneck territory you’re holding out of habit or fear. Everything where a mistake is expensive and your eye genuinely catches things is your real ground. Stand there. Say the rest out loud: this doesn’t need me.

Right now, being deliberate about this looks fringe. It’s a weird thing to be good at. There’s no line for it on a résumé, and doing it well mostly looks like declining to review things, which takes nerve. That’s what the middle of the arc always looks like. Because the loops are multiplying. Every team is wiring AI into something this year, which means every team is about to place its humans, mostly by accident. The ones who understand placement will do it on purpose, for themselves first, then for everyone around them. There is a year coming, and it’s not far, when “where do you sit in the loop, and why there” is an interview question.

The people asking it will be the ones who practiced while it was still fringe.

Next in the series: Being the Friction — on asking the question nobody has time for.