Who Owns the Decision?

The whole industry agreed this month that every AI agent needs an owner. They're right, and it's the wrong altitude. You don't own an agent. You own a decision — and the thing nobody is owning is what that decision costs.

Open any feed this month and you'll find the same sentence, phrased a dozen ways across newsletters and governance frameworks: every AI agent needs an owner. The work an agent does eventually becomes part of the work, the argument goes, and the work needs one accountable person — not a committee, and not "the AI team" by default. The enterprise governance world arrived at the same place from the security side: treat every agent as a distinct principal with credentials, permissions, a documented lifecycle, and an owner of record in a registry. Different vocabularies, one conclusion: name a human, or the agent drifts.

Good. That's settled, and I'm not going to spend this essay re-litigating it. Yes — assign owners. Of course you should. But watch what happens the moment everyone agrees on a sentence: the agreement becomes a place to stop thinking. "We named owners for our agents" feels like the work is done, and it has barely started — because the agent is not the thing you own. An agent is a vehicle. What it carries is decisions, and decisions, not agents, are where a marketing P&L is actually made or lost.

You own a decision class, not an agent

Here is the move, and it's the whole essay: the unit of ownership is the decision class, not the agent. A decision class is a recurring type of choice a system makes — how to bid, which audience signal to weight, which of a thousand creative combinations to serve, when to expand a budget. A single agent runs many decision classes at once, and they are not interchangeable: they have different objectives, different costs, different tolerances for being wrong. "Who owns the bidding agent" is the wrong question because that one agent is making four economically distinct kinds of decision, and lumping them under one name on a wiki tells you nothing about whether any of them is any good.

So "name an owner for the agent" quietly under-specifies the job. To actually own a decision class is to own five things: its objective (what this class of decision is for), its denominator (the cost per decision), its latency budget (how fast it's allowed to be), its kill condition (the threshold at which a human takes it back), and your name on all four. Ownership without those five isn't accountability. It's a label. And the most common version of the label is the dangerous one, because it satisfies the audit while leaving the only number that matters undefined.

Which number? The denominator. This is the thread from Proof Is the New Flex — And It Has a Unit: proof is a ratio, and the industry forgot the bottom of the fraction. The same blank is sitting underneath "agent ownership" right now. We've gotten very good at naming who is responsible for an agent and completely skipped naming what the agent's decisions cost. An owner with no denominator owns a reputation, not a decision.

An unowned decision is an unpriced decision

Say it plainly, because it's the line the whole argument turns on: an unowned decision is an unpriced decision. Not unmonitored — unpriced. The agent can be logged, compliant, sitting in a tidy registry with a named owner, and still be making ten thousand bidding decisions a day at a cost per decision that no human being has ever computed, with a decision latency nobody has ever budgeted. It passes every check. And it is precisely because it passes every check that it's dangerous: an unpriced decision is silently expensive and silently slow, and the silence is the problem. You cannot manage a cost you've never put a number on.

This is the failure mode the agent-ownership conversation keeps walking past. The governance frameworks are built to catch the agent that misbehaves — exceeds its permissions, touches data it shouldn't, breaks a policy. They are not built to catch the agent that behaves perfectly while making a class of decisions that's quietly, expensively wrong at a scale no human review could ever reach. The well-governed, fully-owned, completely unpriced agent is the one that wrecks the quarter, and it does it without tripping a single alarm.

Governance assigns blame. Decision economics assigns a price.

The strongest objection here comes from the people who've actually built this discipline, and it deserves a real answer. A governance or security veteran will say: we already do all of this. We have an agent registry. Every agent is a principal with an owner. We follow the NIST and CSA agentic standards — bounded autonomy, auditability, accountability, continuous oversight. You've reinvented our framework and stapled marketing words to it.

And the registry is real, and it's good, and you should have one. But notice exactly what it prices. A registry prices risk: who is allowed to do what, who is blamed when something breaks, where the audit trail lives. It does not price decisions. It will record, without complaint, an agent that is perfectly within its permissions and making a class of decisions at a cost per decision no one has computed and a latency no one has budgeted. Governance answers "is this allowed, and who's accountable if it fails?" Decision economics answers a different question entirely: "what does each decision cost, and how fast does it learn?" The two are complementary, not redundant — and the gap between them is exactly where the owned-but-unpriced agent lives. The CISO's registry will tell you who to blame after the P&L erodes. It will not tell you the erosion is happening, because a well-priced disaster and a well-governed one look identical in an audit log.

There's a second objection, and it comes from the operators: this is just management. Assign a person to own outcomes — that's Management 101. And the most agentic shops already do it; the leading agencies run dozens of tools and thousands of logged experiments with clear owners. Daniel Gilbert's Brainlabs brands itself the "most agentic media agency" on a foundation of 32 media tools and 2,500-plus experiments, each presumably owned by someone.

Ownership of people's output is Management 101 — because a person makes a handful of consequential decisions a day, and a manager can review them. That's the assumption that just broke. An agent makes a consequential decision class thousands of times a second, inside a system that also reports its own effectiveness. You cannot manage that by review; there is nothing to review at that cadence, only outcomes the system has already summarized for you. The only way to manage a decision class running at machine speed is to instrument its denominator — watch cost per decision and decision latency move, and trip the kill condition when they go the wrong way. Tool count, in this light, is just inventory. Thirty-two tools is thirty-two more vehicles, and a vehicle with no priced decisions is exquisitely instrumented to be wrong faster. The moat was never the number of agents. It's whether you own the price of what they decide.

Why this is the bill that's about to come due

This would matter in any year. It matters acutely now because the platforms are shipping owned, opinionated agents on both sides of the auction, and each one arrives with its own price for your decisions baked in. On the buy side, Ask Advisor is a single Gemini agent spanning Google Ads, Analytics, Merchant Center, and the Marketing Platform, advising you on how to spend. On the sell side, the newly launched Ask Ad Manager advises publishers on how to sell. The system that spends, the system that sells, and — as I argued in You Can't Let the Optimizer Own the Evidence Layer — the system that grades the spend increasingly share one model family.

When the counterparty's agent makes a decision and prices it for you, you are accepting its denominator by default. Every decision class you leave unpriced is a decision class the platform has already priced — in its own favor, with its own objective, reported back to you as a single tidy outcome number. This is the agentic version of the oldest asymmetry in media buying, except it now runs millions of times a day and grades its own homework. Chamath's warning that companies have roughly 500 days to prove AI spend converts to real revenue is, read correctly, a clock on the denominator: the window in which you can still build your own decision-pricing before "the platform's number" becomes the only number anyone remembers to ask for.

The decision-class ownership card

Here's the Lab version — proof, not takes — and like the best instruments it fits on an index card. Pick one decision class your agents already run, this quarter, and fill in five lines:

One: name the decision class. Not "the PMax agent" — the specific recurring choice. "How we weight first-party audience signals in Performance Max." "Which creative the system serves to returning carts." If your team can't crisply name the class, that's the finding: you've been owning an agent, not a decision.

Two: state its objective in one sentence. What is this class of decision for — and is it your objective, or the one the platform's agent optimizes by default? Where those differ, you've found an unpriced decision wearing a borrowed objective.

Three: compute its denominator. Cost per decision, from your own first-party count of decisions delivered — not the platform's modeled column. This is the number that was blank. The unit should be the decision your P&L actually buys: a member-year, a funded account, a cart.

Four: set a latency budget and a kill condition. How fast is this decision class allowed to act on a new signal, and at what cost or latency does a human take it back? An owner with no kill switch isn't an owner; they're a spectator with a title.

Five: put a name on it — and do it before the surfaces close. One human, accountable for the four lines above. Every quarter you wait, more of the decision moves behind a platform agent that prices it for you, and reconstructing an independent denominator gets harder. You can adopt the platform's agents whenever you like; you cannot retrofit an owned denominator onto a quarter of decisions you let run unpriced. That's the one move here that does not survive being deferred.

What this looks like from the CEO chair

Strip the agent vocabulary and the question a leader is actually answering is simple: when my marketing organization spends money this quarter, who can tell me what each decision cost? "We've assigned owners to our agents" is not an answer to that question; it's an answer to a different, easier one. The harder answer — and the one that translates to the P&L — is a named human who can show you the cost per decision, the latency, and the line they'll pull if either goes wrong. Accountability in an agentic org isn't a list of who owns which bot. It's a map of who owns which decision, priced. The first is a compliance artifact. The second is a way to run a company.

Every agent needs an owner — fine, settled, true. But the sentence is a doorway, not a destination. Walk through it and the real work is waiting: an agent is a vehicle, a decision is the cargo, and the only ownership that changes a P&L is ownership of the decision's price. Pick one decision class this quarter. Name it, state its objective, compute its cost per decision, set its latency budget and its kill condition, and sign your name. Do that and you own the decision. Skip it and you've named an owner for a vehicle while the platform quietly prices the cargo — which is to say, you don't own the decision at all. Somebody does. The only question this whole agentic moment is really asking is whether it's you.