Recommendation or Decision? The Veto That Doesn't Scale

The cold open

Here is this week’s thought experiment, posed exactly as it ran:

At what point does an AI recommendation become an AI decision? When humans stop reviewing it? When the money moves? When it’s automated? Or never — a recommendation is always just advice?

It reads like a question about a line you cross — one clean threshold out ahead, our job to mark it before we step over. That framing is the trap. When a question asks you to find a single line, check first whether you are looking at a slope and calling it a step.

The vote

The poll closed with 49 votes. Here is how the room split:

Read the bottom row first, because it is the one almost nobody picked. Barely anyone said “never”; everyone else agrees authority transfers to the machine at some point — they only disagree about which symptom marks the moment. That is the real finding, hiding under the winner. A vote split on when but near-unanimous on whether is the signature of a gradient, not a boundary. A real line makes people cluster; instead they scattered across the slope. The question as posed — “at what point” — is malformed. It assumes a step where the room is describing a descent.

The reframe

So here is the turn. Look at the winner — “when humans stop reviewing,” also where I put my own vote — and notice the tense. It is written in the future, a threshold we are approaching if we are not careful.

In live programmatic, it already happened. Nobody reviews the 100-millisecond bid, the lookalike expansion that tripled the addressable pool overnight, or the budget reallocation the optimizer ran while the buyer slept. And — this is the part that matters — nobody pretends to. The majority described the present and mistook it for the future. The line isn’t ahead of us. It’s behind us, and we drove past it years ago without a vote.

So here is a working definition you can test. A recommendation is advice you can decline; it becomes a decision the instant a named human can no longer meaningfully refuse it — when the veto still exists on the org chart but not in the workflow. Decisions create accountability — and accountability is a human relation software cannot occupy. The test is not “did a human approve it?” It is “could the human have said no, and made it stick?” Everything below measures that edge.

The Veto That Doesn’t Scale

Here is the first mechanism, the engine under everything else. Human review is fixed-throughput — one person, finite attention, seconds per item, and the day ends. Machine recommendation is elastic; it scales with compute. When a DSP emits thousands of bid and audience decisions per second, “approval” cannot mean inspecting them — only rubber-stamping the aggregate after the fact.

So the transition is not an event anyone schedules. It is a threshold crossed by arithmetic: the moment a system’s output rate exceeds a human’s adjudication rate, “human-in-the-loop” degrades into “human-on-the-receipt.” Nobody decided to stop deciding; the clock decided for them. There was no day anyone handed over the wheel — there was a slope, the output rate outran the refusal rate, and the veto went decorative without anyone unbolting it.

The Dissent-Capacity Test

“The human stopped reviewing” is too soft — it sounds like a behavioral lapse a more diligent operator would fix. It isn’t. It is an engineering property you can measure. A reviewer’s “no” is real only if they hold three things at once: an instrument (a way to interrogate the recommendation’s basis), standing (permission to override the system’s own KPI), and time (a review window longer than the action’s refresh cycle). Lose any one and dissent-capacity collapses to zero.

Run it on a real workflow. A trader “approving” a stream that re-bids every 100 milliseconds via a weekly dashboard has no instrument (it shows outcomes, not reasons), no standing (overriding the optimizer means missing the number it exists to hit), and negative time (the action refreshes hundreds of thousands of times between reviews). Dissent-capacity: zero. The approve button is real; the dissent it implies is not. That is how you audit it — ask a vendor to prove the reviewer’s instrument, standing, and time, and watch the recommendation reveal itself as a decision already made.

The Liability Vacuum — and its uglier twin

Now the second mechanism, the one with the moral sting. Automating a decision does not produce an accountable machine. It produces a decision with no accountable party at all. Accountability is a social relation — someone who can be praised, blamed, fired, sued — and software cannot stand in any of those positions. When the human withdraws, answerability does not move to the model. It falls into a gap.

The twin that makes it worse: the surviving signature does not fill that gap. It disguises it. Accountability laundering — the click converts an ungoverned machine output into an “approved human decision” on the record. The audit trail now points at a person who could not possibly have caught the brand-safety breach or the synthetic-audience spend, because their dissent-capacity was zero. The signature absorbs blame the model cannot hold, which lets the org defer real governance — a human reviewed it — while the review was theater. That is why a human in the frame can be worse than honest automation: at least honest automation admits no one is watching. The laundered signature manufactures false comfort, which is more dangerous than acknowledged risk.

The Reversibility Horizon

So where does the decision live in time? Not where the money moves. The 24% who voted “when money moves” picked the receipt, not the choice — and you cannot govern a system by reading its receipts. The decision lives one step upstream, at the reversibility horizon: the last timestamp where a named owner could still intercept the action at acceptable cost — cap the bid, pause the line item, kill the audience — before the cost-to-undo runs away. Everything after is consequence, not choice. The fix follows directly: write a handoff ledger at the horizon — who owned the boundary, the cost-to-reverse at that instant, and the threshold the system was permitted to cross alone. Without it, “the AI decided” is unfalsifiable, because no named human sits on the handoff and accountability evaporates into the model by default.

Three Fridays, one hole

Three weeks, three blind spots — eroding standards, AI-native fraud, and now the recommendation-decision line — one hole seen from three sides. Strip the jargon and the load-bearing word in all three is identical: answerability. Trust is the expectation someone will answer for an outcome. Authenticity is verifiable provenance of who is answerable. A decision is where answerability attaches. The agentic transition is not a capability problem wearing a fraud costume one Friday and a standards costume the next. It is an accountability problem, every time, all the way down.

My vote, corrected

I voted with the majority, and I want to be precise about what that was worth. “When humans stop reviewing” was the right symptom. But the tense was a tell, and it is my own worth confessing: I picked the answer that lets us imagine the line is still ahead — the comfortable vote, the one that assumes we still have a wheel to let go of. We mostly don’t. Most agentic ad workflows are already over the line, dressed as “AI-assisted,” with a dissent-capacity of zero and a signature doing laundry. The real work was never deciding when to hand over control. It is auditing where we already did, and rebuilding a veto that bites.

Because the veto we have today is a smoke detector someone quietly unplugged because it kept chirping. Still bolted to the ceiling, still listed on the org chart as installed, and everyone feels safer for seeing it there. But a detector that can’t sound is just a white plastic disc with a name on it. We didn’t keep a human in the loop. We kept a human in the frame — because a system that can’t be blamed needs someone who can. You discover the battery’s been out for months only when the room is already full of smoke.

So stop asking when a recommendation becomes a decision. Ask who you’d name in the post-mortem — and if the only honest answer is “the model,” the decision was made the moment you accepted an answer you could no longer refuse.

Next Friday

If a decision is just answerability with a name attached, the next question writes itself: when an agent acts inside its authority and the outcome is still wrong, who answers — and is the thing that proves it a protocol, a constitution, or a contract? That’s where this goes. Let’s see how this plays out. My 2c, as always — food for thought for the weekend.

(Answered the Friday after: What Agents Can’t Manufacture — the credential must be conferred from outside the system, and whoever owns that issuance owns the bottleneck.)