You do not own an AI agent. You own a class of decisions it makes on your behalf, and the moment nobody has priced that class, the decision still gets made, just by whoever set the default.
This page is the operating manual for that problem: the ladder that assigns accountability, the custody chain that keeps your evidence independent of the platform selling you media, and the three ledgers that make both auditable on a Tuesday rather than in a post mortem. Every term below is defined once here and argued at length in an essay.
Each rung answers a question the one above it leaves open. Most organizations have the first and the last, and nothing in between, which is why AI governance documents tend to read well and change nothing.
Ownership of the decision is worth little if the evidence you judge it by belongs to the counterparty. Over the past year the seller has come to generate the counterfactual, define the billable unit, supply an input to the model meant to audit it, and set the clock on how long you keep your own history. None of that requires bad faith on the platform's part. It requires custody on yours.
The instruments are deliberately cheap: pages, not platforms. Together they are one register with three columns, and an organization that keeps all three can reconstruct why a number moved without asking the party who moved it.
Every number that binds, with a named owner and a re-pricing condition.
For every board-deck metric: who defines it, fact or judgment, when the definition last changed, whether history was restated, and the buyer-owned check.
Which rule was relaxed, how far, how long, what it was expected to buy, and afterward, what it actually bought.
The argument, in the order it was built.
Sixteen terms, each defined in one sentence and argued in full elsewhere.
Decision ownership is the practice of assigning accountability for classes of decisions rather than for tools or agents. An organization does not own an AI agent in any meaningful sense; it owns the decisions that agent makes on its behalf, such as how budget is reallocated, what a lead is worth, or when a target gets relaxed. Ownership means a named person, a priced consequence, and a written condition under which the decision comes back to a human. Where those three are missing, the decision still gets made, usually by whoever set the default.
Whoever owns the decision class, which in practice means whoever owns the profit and loss consequence rather than whoever operates the tool. The common failure is assuming the platform, the agency, or the person who clicked the setting is accountable. Platforms ship defaults that make decisions for everyone who has not priced the alternative, operators execute inside constraints they did not set, and the accountability lands wherever it was assigned in advance. If it was never assigned in advance, it lands nowhere, and an unowned decision is an unpriced decision.
Start by naming the decision classes and assigning each one an owner, because everything below depends on knowing who is answerable. Then price the classes, since an owner without a number cannot be held to anything. Then set the thresholds in both directions: the condition under which a human takes the class back, and the condition under which the machine gets it again. Only then take on the evidence layer, because auditing a seller's numbers is wasted work if nobody owns the decision those numbers feed. Most organizations attempt this in reverse, starting with dashboards and governance documents, which is why the paperwork improves and the accountability does not.
By owning the denominator, the unit, and the counterfactual rather than accepting the seller's versions of them. In practice that means computing your own baseline from first-party data, recording who defines each billable unit and when that definition last changed, treating platform estimates of missed growth as quotes rather than findings, and keeping a measurement input that the optimizer does not also supply. None of this requires distrusting the platform's models, which are often better than what a buyer could build. It requires custody: the buyer keeps a record the seller cannot revise.