The human router.

Somewhere in your company, a person with a serious title is spending their afternoon as plumbing. They were hired for judgment: to decide what matters, to own an outcome, to be accountable for a call. What they are doing instead is moving text between windows, because the agent that produced an answer and the people who need it have no other way to meet. Nobody assigned this job. It assembled itself out of small, locally sensible decisions. Judgment is the human job; at most companies adopting AI, we have quietly assigned people the machine job instead.

This is the human-router problem: the agents do the substantial work, and a person carries context in, reconciles what comes out, and decides where every result belongs.

It is a descriptive claim, built from watching teams work, including teams I have been part of; anyone can verify it in an afternoon spent next to a team that uses agents seriously. What I cannot give you is a measured coordination tax, and I would distrust the precision of anyone who offered one. The size of the cost is uncertain. Its direction is not: it grows with adoption while the gains per agent do not, and that property, more than its current size, is what should worry us.

The mechanism is not complicated. Each private session accumulates the corrections, sources, and dead ends that made its output good, and none of that is reachable by anyone else in the company. From there the arithmetic takes over. Agents multiply, tools multiply, and every pairing between them routes through a person, because the agents share no environment in which to meet. A company in that position gains real local productivity and, at the same time, more duplicated work, more decisions nobody witnessed, more time spent verifying, and a deeper dependence on particular employees as the points where information crosses. That is productivity. It is not leverage.

Here is the scenario I keep seeing. A product manager copies a ticket into an AI chat, because the agent cannot see the tracker. The answer is good, so she pastes it into the team chat, trimming the parts that would need explaining. Two people read it; one asks a question the agent already answered in a part she trimmed. In Thursday's meeting she explains the finding again for the people who missed it. Meanwhile a colleague asked his own agent a similar question with slightly different context and got a slightly different answer, so now the team runs a reconciliation discussion between two machine outputs, mediated entirely by people. Every step in that chain was locally rational. The sum is a person working as infrastructure.

The obvious objection is that agent-to-agent protocols will dissolve the problem, and it is partly right. Protocols for connecting agents to tools and to each other are arriving, Model Context Protocol among them, and they are necessary. But transport is the easy half. If two agents exchange results without shared context, persistent identity, and a place where people can see and correct the exchange, the router problem does not disappear; it moves up a level, to the person who must now audit an inter-agent pipeline they cannot observe. A second objection says the human in the loop is the feature, not the bug. For approval and judgment, yes, emphatically. The distinction that matters is between loops where a person adds judgment and loops where a person adds latency. Today most teams cannot tell them apart, because both look like someone pasting text between windows.

That people are currently the integration layer for agent work is description, not forecast. The forecast, more confident in direction than timing, is that this coordination overhead, not model capability, becomes the visible constraint on agent returns for most teams within a few years. The recommendation is to route agent work through shared surfaces with attribution and explicit approval points: keep people at the decisions, remove them from the transport.

The point of removing people from the transport layer is not to remove people.

Attention is the scarcest resource an organization has, and the router job consumes it invisibly, in minutes that never appear on any roadmap. What people get back, when the plumbing stops being their job, is the work only they can do: deciding what matters, judging what is true enough to act on, and being accountable to each other for the outcome. That is not a smaller role for people. It is the original one.