What a night shift actually looks like
The interesting thing about an unattended shift is not that software ran overnight. Cron has done that for forty years. It is that at 08:00 there is a short list of decisions with money attached, instead of a dashboard you still have to interpret.
Here is a real shape of one, from a 412-ASIN home and kitchen account. Ten entries between 22:00 and 08:00, spread across the seven agents on the roster. Zero alerts sent. One escalation.
The first four hours are inventory and waste
Velocity was up 18% week over week. That alone is not a decision — it could be a returns spike inflating sell-through. Supply checked returns first, found them flat at 3.1%, then read inbound shipments and lead time from the last three POs before putting a date on it. The date is the decision. The velocity is just the reason.
The obvious objection to this one is attribution lag, and it is the right objection. Spend ruled it out because the window closed fourteen days earlier. An ads tool that reports the number without ruling out the lag is handing you homework, not a decision.
The middle of the night is where the money is
Contribution per unit went negative after a size-tier change. Not a fee correction that would reverse — a permanent tier move. The campaign pause was staged and costed, then left for a human, because this account runs in Recommend. What each permission mode does overnight is worth settling before you switch anything on.
This is the entry that pays for the shift. Advertising a product that loses $0.87 a unit is the kind of mistake that survives for weeks precisely because every individual dashboard looks fine: spend is on plan, ACOS is acceptable, the listing is healthy. Only contribution after fees tells you it is wrong, and nobody checks contribution after fees at 03:42. The same failure, followed through a whole agency for five weeks, is written up in the fee tier that turned six SKUs negative.
One event, not three alerts
At 02:40 a competitor cut price 12%. Margin recomputed the break-even ACOS at the new competitive price. Spend proposed holding the bid rather than matching, because matching would spend $1,900 to defend $340. Health confirmed the listing was clean, so the share loss was price and not suppression.
Four findings. One event. The work is deciding they describe the same thing, and then sending one decision rather than four notifications that each look urgent on their own. The hand-off that assembles them is drawn step by step on its own page, and three accounts it has run on are in the case files.
Account data in this piece is illustrative and anonymised.