Two posts ago I told you what AI is: range, memory, speed — a wheelchair. Last post I told you where it goes blind: the pause, the glance, the history, the gap between accurate and true.
Which leaves the only question that matters operationally. What do you actually point it at?
The answer is a division of labor, and it’s not a complicated one.
AI handles what systems should handle. Humans do what only humans can do. The entire art is refusing to let either one do the other’s job.
—
Go back through this manifesto and collect everything I’ve said systems should be doing. The remembering — follow-ups surfaced instead of recalled, context carried across handoffs instead of shredded. The capture that happens inside the work. The routing, the flagging, the four-days-have-passed. The pattern-reading across more records than any person can hold. The early warning that arrives while the problem is still cheap. The reporting that doesn’t require a human to spend Friday assembling it from four sources that don’t agree.
Every item on that list used to require either infrastructure nobody built or discipline nobody sustains. That’s the work AI should be doing — not because it’s menial, but because it’s *systematic*: definable, repeatable, pattern-shaped. It’s the load this entire manifesto has been begging companies to take off their humans’ nervous systems.
And here’s what most organizations miss about that trade. When AI absorbs the systematic load, the point isn’t the cost savings. The point is what it releases.
Because look at what’s been consuming your best people. The rep’s working memory, spent on remembering instead of selling. The ops manager’s week, spent reconciling instead of noticing. The founder’s attention — the single scarcest resource in the building — spent being the router, the translator, the human integration layer between departments. All of Part Two of this manifesto, all that absorption and duct tape and compensation: that’s your humans doing machine work, badly, at great personal cost, because nobody built the machine.
Take that load off, and what’s left is capacity. Capacity for exactly the work the last post said no model can reach. Sitting in the room. Hearing the pause. Untangling the tension that’s relational, not structural. Making the judgment call the data can’t make. Building the trust that makes honesty cheap. Designing the next system — because design is a judgment about humans, and machines don’t get to make it.
That’s the right use of it: AI does the systematic so humans can finally do the human.
—
Most organizations are running this exactly backwards, and the backwards version has a recognizable shape.
They deploy AI at the judgment layer — auto-drafted outreach nobody reviews, scoring models nobody interrogates, chatbots parked directly on the customer relationship — while their humans stay buried in the systematic layer, copying data between tools, reconciling reports, remembering follow-ups. Machine work for the people. Judgment work for the machine. It’s the stair-building problem from 16a with a new coat of paint: technology purchased to avoid the redesign, deployed where it’s blind, withheld where it sees.
The organizations that understand the division will build differently, and it will show up everywhere. They’ll buy AI to eliminate reconciliation, not conversation. They’ll measure it by how much human attention it returns, not how much it replaces. Their people will spend more time in rooms with customers and each other, not less — because the machine finally took the clipboard.
The future isn’t AI-shaped or human-shaped. It’s shaped like a boundary, drawn deliberately, by people who know what each side is for.
Draw it on purpose. That’s the whole instruction.