Loading
Against a severity list you wrote, asking the questions that decide urgency, and escalating rather than downgrading whenever the answer is ambiguous.
Published December 9, 2025
Overnight triage is a problem of asymmetric cost. Waking the person on call for something routine wastes an hour and erodes their willingness to answer next time. Not waking them for something genuine costs far more. Most out-of-hours arrangements fail in the second direction because the first failure is the one people remember.
The single most important input is a written list of what counts as an emergency in your business. Not a general principle, a list. It is usually shorter than people expect and it is almost always the first time anyone has written it down, which is itself worth the exercise.
The agent applies that list exactly. It does not form its own view about whether something sounds serious, because a model's intuition about severity is not accountable to anybody and yours is.
Severity is rarely obvious from the first sentence. Someone reporting that something is broken has told you almost nothing. Three questions usually resolve it.
A human taking the call at 2 AM often does not ask these, because they are half asleep and want to get off the phone. The agent asks them every time, in the same order, and writes the answers into the ticket.
This is the rule that makes the rest safe. If the answers do not clearly place a report on the severity list, it escalates. The failure mode becomes a call the duty engineer did not strictly need, which is recoverable, rather than an incident that sat until morning, which sometimes is not.
Teams occasionally want to tune this the other way after a few unnecessary pages. Resist that for the first few months. The unnecessary pages are visible and the avoided ones are not, so the pressure is always in the direction of loosening, and the cost of loosening only shows up on the night it matters.
The agent should read your rota rather than a static list, and it should follow your fallback chain when the first contact does not acknowledge. An escalation that goes to somebody who left the company or is on holiday is the same as no escalation, and this is a surprisingly common failure in hand-rolled arrangements.
The reports that are not emergencies should not simply wait. They should become a ticket with the detail, the timeline, what was already tried and when the contact is available, filed in your system and formatted the way your team expects.
This is where most of the time saving actually comes from. It is not the avoided phone calls, it is the mornings that start with fifteen tickets somebody can work from rather than fifteen voicemails somebody has to call back about before any work can begin.
Confirm what happens next and when. The person who reported a fault at 2 AM and heard nothing again will chase, and that chase costs your team more than the original report did. A single message saying it is logged and someone will look at it by nine removes most follow-up traffic.
On the site
Keep reading