← Wiki

The Bottleneck Shift

automationfuture-of-workai-agentsconstraint-theory

The mechanism

Automating one step in a workflow doesn’t make the whole workflow faster. It makes the next step the new rate-limiter. The downstream constraint was always there, but it wasn’t visible because everything upstream was too slow to expose it. Remove the upstream friction, and the constraint migrates.

This is different from two related ideas. The Expanding Work Frontier describes new work that never existed before — tasks AI makes possible for the first time. The Jevons paradox describes increased consumption of a resource when it gets cheaper. The bottleneck shift is narrower: it’s constraint migration within an existing workflow. You clear one chokepoint, and the system’s binding constraint moves to the next one in the chain.

Goldratt formalized this in the Theory of Constraints: every system has exactly one binding constraint at any given time. Improving anything that isn’t the constraint doesn’t improve the system. And when you do relieve the constraint, another element immediately becomes the new bottleneck. Operations researchers have observed this pattern in manufacturing and logistics for decades. AI is now reproducing it in knowledge work.

Healthcare: the referral that goes nowhere

Levie describes a healthcare company that automated patient referrals. The old process was brutal — patients spent days on the phone navigating administrative runaround just to get a referral processed. AI eliminated that friction entirely. But the appointment on the other end didn’t get any closer.

“You can automate anything, but if it still is 18 months out before an appointment is available, your ultimate constraint is still the health care institution and the amount of doctors we have.”

The referral bottleneck was hiding the supply bottleneck. When patients couldn’t even get referred efficiently, nobody noticed (or measured) the 18-month wait as the binding constraint. Now they do. The total work in the system didn’t shrink. It shifted from administrative staff handling referrals to the deeper problem of physician supply.

The same pattern shows up in law. AI makes generating legal documents trivially easy. Contracts, briefs, patent applications — all producible at a fraction of the old cost and time. But the bottleneck was never document creation. It’s document approval: court processing, patent office review, regulatory filings, and the judgment calls that only licensed attorneys can make.

Levie’s prediction runs against the displacement narrative: “There are going to be more lawyers in the next five years than we have today, because we’ve made it easy to generate legal content. But it has not gotten any easier to actually get any of that approved by any court system.”

More legal output means more work for the parts of the legal system that can’t be automated. The constraint migrates downstream, and demand for the people who handle that downstream work goes up.

Cybersecurity: more code, more surface

AI-assisted development produces code at rates that were previously impossible. Levie, responding to Anthropic’s autonomous security research, points out the consequence: “AI is going to generate 100X more code, and along with that, there will be an enormous increase in security discoveries.” AI can triage vulnerabilities and flag patterns, but remediation, architectural decisions, and incident response still require human experts.

“Better AI tooling for security will increase the demand for security talent, not decrease it.”

The bottleneck in software security was never finding vulnerabilities — it was always the scarcity of people who could fix them in context. Automating code generation and vulnerability scanning just makes that scarcity more acute.

What static analyses miss

Employment predictions that count automatable tasks and project displacement assume the rest of the workflow stays constant. They don’t account for constraint migration. When the easy step gets automated, the hard step becomes the system’s new rate-limiter, and demand for workers who handle that step increases. A static task inventory can’t capture this dynamic because the bottleneck only becomes visible after the upstream automation is deployed.

This doesn’t mean automation never displaces workers. It does — at the specific step that got automated. But the system-level effect on labor demand depends on where the constraint lands next, and whether that next constraint requires human judgment, institutional processes, or physical resources that can’t be scaled with software.