The AI-Native Beauty Company

Finance ties Retail Operations for the most owned duties in the atlas at nine, and both of its wave-1 cells get a pod. Accounting built its review gate before any agent arrived: a controller signing a package an agent prepared is the control that was always there. The preparation moves, and the signature stays with the person who gets examined on it.

Fourteen drafts, more than any function in the atlas, and the blueprint targets every cell at drafts and stops there. Strategy is where an agent's first cut is genuinely strong and where accepting it without doing the trade-off thinking is the failure mode: a manager who rubber-stamps a synthesis memo has outsourced the judgment the role exists for. Corporate development carries a sixty percent agent share and still goes in wave 3, because a screening long-list is cheap and walking away from six months of work is the job.

Marketing's twelve human duties are the ones that spend money, sign a claim, or sit with an agency, and none of them move. All three sub-functions land in wave 2 for the same reason: the agent's output goes to an external party or releases live budget, so the gate has to be a named marketer with the standing to reject it. Digital is the one marketing cell where a mainstream tool already drafts; like most cells that already draft, this blueprint declines to advance it further.

Nine owned duties, tied for the most in the atlas, and every one of them is data about a counter. The pod scores planograms from photos, detects shade-level out-of-stocks, holds the tester register, and reconciles sell-in against sell-out, which is the entire scorecard a counter is measured on. A tester logged as replaced is still verified by picking it up, and the position by the escalator is still allocated by a floor manager who has to like you, which is why area managers still spend their week in cars.

All three sub-functions go in wave 1 and none of them earns a pod. The artifact work is dense enough to hand over first, and the owned share never reaches the level where an agent supervising agents beats one analyst supervising three of them. The function's seat in the room survives, and that seat was never about the data.

Five owned duties against twelve human ones, and the blueprint does not fight that. Talent acquisition is where a mainstream tool already drafts and where this blueprint stops: several markets' labour law requires a documented human review before a rejection goes out, so the screening agent's output is a ranked list a named recruiter is accountable for reading. Employee relations goes in wave 3 and drafts only, because the routine-looking leave query that turns out to be a grievance is exactly the case pattern-matching gets wrong.

The split here is physical. Costing models, pipeline trackers, qualification programmes, and consumer-test tabulation go to agents because their inputs are structured and their output format is fixed, and packaging and claims testing is the wave-1 charter that proves it. Formulation gets no team, because the observation that explains a stability failure three weeks later is the chemist noticing the emulsion looked faintly grainy at cool-down, and nobody wrote it down.

The lowest owned count in the atlas at four, and one cell that is nonetheless the second-highest agent share in the whole lattice. Voice-of-customer is a coding and tabulation problem with only three human duties attached, which is why it gets a charter in wave 1. Complaints sits in wave 3 right next to it, because the same inbox carries the adverse-event report that a named person has to decide is reportable.

One function holds both a wave-1 charter and the blueprint's hardest floor, and the line between them is where the liability attaches. Ingredient screening produces a flag list a person reads before anything is filed, so it goes first and it goes wide: an agent runs every market's annex across the whole portfolio faster and more completely than a person working SKU by SKU. Registration gets no team at all, because the law names a responsible person with an address and the safety assessment carries a qualified assessor's signature, which is a legal object rather than an organisational convention.

Two of the three cells here get no agent team at all, the only place in the blueprint where that happens twice in one function. Key accounts and field sales look automatable on the duty count because sell-in reporting and deduction matching get filed under them, and that work actually belongs to Revenue Operations, which is where the pod goes. What is left in a key account is the room, and the atlas prices it at fifteen human duties out of thirty.

Seventeen of thirty duties stay human here, the highest count in the atlas, and almost every one of them is an arbitration between functions whose incentives are built to conflict. The calculation layer goes to agents in wave 2 and the pods draft rather than own, because the number they produce is the number a planner gets held to. Allocating constrained supply during a shortage is left for last, and it is a political act wearing a spreadsheet.