When to pause an AI campaign

Guide · Market Growth · 5 min read · last verified 2026-07-27

Kill criteria are the conditions, written down before a campaign launches, under which it will be paused or stopped without further debate. They answer, in advance, the question every struggling campaign eventually forces: how bad does it have to get, for how long, before we act? Budget caps put a ceiling on the damage any single day can do - that engineering is covered in How to set budget caps for AI campaigns - but caps say nothing about duration. A campaign can sit politely under every cap for weeks while producing nothing. Kill criteria are the control that bounds how long that is allowed to continue, and they matter doubly when an AI agent is doing the operating, because an agent will keep optimizing whatever it was told to optimize long after a human would have smelled the problem.

Write the exit before the entrance

The reason the criteria must predate the launch is not ceremony; it is that after launch, nobody involved is neutral. Sunk cost argues for one more day. Authorship argues too - the person who built the campaign reads every ambiguous signal generously, not because they are dishonest but because they can see what it was supposed to do. And once money is gone, every conversation about stopping becomes a negotiation over whose judgment failed. Criteria written before launch are written by calmer people who do not yet have a side.

The pre-connection discipline in Campaign readiness: what to check before connecting ad accounts requires that stop conditions exist before an account is connected. This piece is about what those conditions should actually say. They come in two families that behave differently and must not be blended.

Spend without signal: the measured triggers

The first family accumulates. Its form: spend has passed a multiple of what one expected signal should cost, and no signal has arrived. Everything in that sentence has to be defined per campaign, before launch.

First, define the signal for this campaign's stage. For a conversion campaign, a click is not a signal; a completed conversion is. For an early awareness test, engaged visits might legitimately count. A campaign that is allowed to redefine its signal downward mid-flight - conversions were the goal, but look at all these clicks - has no kill criteria at all.

Second, run two clocks. The spend clock: a multiple of the expected cost of one signal has been spent, and zero signals have arrived. The calendar clock: the evaluation window ends on a date, regardless of spend pace, so that a slow-spending failure cannot hide by dribbling budget. The trigger is whichever clock runs out first. Choose the multiple with the platform's settling period inside it - new campaigns legitimately need time to stabilize - and then believe your own number. The multiple you pick is a judgment call; the control is that it is written down and not renegotiated while the meter runs.

Quality and brand safety: the immediate triggers

The second family does not accumulate. One instance is enough: a generated ad variant makes a claim you cannot support; ads appear alongside content you would never sponsor; the landing page breaks while spend continues; targeting drifts to an audience nobody chose; a named customer or partner complains. No multiples, no windows, no waiting for a pattern.

For this family the usual order of operations inverts. The measured family tolerates investigate-then-act, because the cost of a wrong day is bounded by the caps. The immediate family is pause-first, investigate-second, because the damage is reputational and compounds while you deliberate - and unlike budget, reputation has no daily limit. This distinction is worth writing into the criteria explicitly, so that whoever is monitoring knows which triggers permit a discussion and which do not.

Pause is not kill

The two words get used interchangeably and should not be. A pause is reversible: spend halts, campaign state and accumulated learning are preserved, restart is possible. A kill is terminal: the budget is released to other uses, the thesis is declared rejected, the record is archived for the postmortem.

Authority should scale with reversibility. Pause authority belongs to everyone monitoring the campaign - a wrong pause costs a little delay and unwinds in minutes, so the threshold for granting the power should be low. Kill authority stays narrow: the budget owner, after reading the record, because killing a campaign is a business judgment about a thesis, not a threshold event. And watch for the limbo state between them: a campaign that has been paused longer than it ever ran is a kill that nobody has done the paperwork for. Put a review date on every pause at the moment you make it.

Restart conditions

Restart conditions are written at pause time, while the trigger is fresh, not debated later from memory. The form is a sentence: this trigger fired, this is what changed, and this is why the same trigger will not fire again for the same reason. A restart with nothing changed is a bet that the trigger itself was miscalibrated - which is sometimes true, but it must be said out loud and logged as exactly that bet, because a trigger you silently override is a trigger you no longer have.

The full sequence - which criterion fired, who paused, what the investigation found, what changed, who approved the restart - belongs in the same record as every other action taken on the account, in the shape described in What belongs in an AI action audit trail. That record is also how the next campaign's criteria get better: calibration comes from reading which past triggers fired usefully and which cried wolf.

Give the agent the pause button first

If an AI agent participates in running campaigns, pause is the first authority worth granting, not the last. Pausing fails safe: its worst case is forgone spend, not lost spend. An agent that can halt on written criteria but must ask a human to spend is a safe asymmetry; the reverse - an agent that can spend but only recommends stopping - is exposure wearing a seatbelt backwards. Magrios sequences its gated execution this way deliberately: halt-shaped actions sit below spend-shaped ones on the autonomy ladder, so the safest capability is the first one automated. Kill authority, by contrast, stays human at every autonomy level. Releasing a budget and rejecting a thesis is a judgment about the business, and no threshold crossing makes it for you.