How do I tell which overnight calls are actually urgent?

This page may contain affiliate links from which Blur Studio earns commissions.

Overnight calls are stressful precisely because you can't tell what you're dealing with until you've already answered, which defeats the point of trying to protect your sleep in the first place. The fix isn't answering every call yourself just in case, it's deciding ahead of time, calmly and in daylight, exactly what counts as urgent for your specific business, then building a system that sorts calls against that definition without you having to be awake for each one.

The cheap version: one on-call number

The simplest fix costs nothing beyond discipline: publish one number specifically for genuine emergencies, staffed by whoever's on call that week, and let your general business line go to voicemail overnight with a clear message pointing urgent callers to the emergency number. This works well for businesses with a small, predictable set of genuine emergencies (a burst pipe, a lockout, a security issue) that customers can self-identify fairly easily. The greeting wording is covered in what to say on a business voicemail so after-hours callers do not hang up.

Where this breaks down

The one-number approach assumes customers correctly judge their own situation as urgent or not, which doesn't always hold. Some callers escalate a minor issue to the emergency line out of anxiety; others sit on a genuinely urgent problem on the general line because they didn't realize there was a faster path. An AI voice agent, like the one in Calilio, a VoIP business phone system, that actually asks a few questions and routes based on the answer, rather than relying on the caller to self-select correctly, closes that gap. Other options are compared in AI phone answering for a small business.

How to set up AI triage, step by step

1. Define what urgent actually means for your specific business, in concrete terms a caller would recognize, not a vague feeling. "Water actively flooding" is urgent; "my faucet has dripped for a week" is not.
2. Write the questions the agent should ask to tell the two apart. Keep it to two or three sharp questions, not an interrogation.
3. Route anything flagged urgent to a real person immediately, with the call connected live, not just a note sent for later.
4. Send everything else as a structured morning message: caller name, number, what they need, and when they called, so the first person reviewing messages isn't reconstructing context from scratch.
5. Review the whole list weekly at first, specifically checking for anything that got mis-routed in either direction, then adjust the questions.

A worked example

Say a plumbing business sets its urgent definition as "active water damage happening right now" or "no working toilet in the home." A caller at 2am reporting a slow drip gets asked by the AI agent whether water is actively flowing or pooling anywhere in the home; when they say no, the agent takes a detailed message and confirms a callback within normal business hours, and everyone sleeps through the night. A different caller at the same hour reporting a burst pipe actively flooding a kitchen gets the same two questions, answers yes, and the call routes straight through to whoever's on call, awake within a minute rather than finding out at 7am that a home flooded for five hours overnight. The questions are nearly identical in setup time; the outcome for each caller is completely different because the triage actually asked rather than assuming.

What this costs

Verified directly on Calilio's own pricing page: the AI Voice Agent that would handle this triage is free up to 10 minutes per user a month on Standard ($15/mo, $12 annual), then $0.25 a minute, or free up to 100 minutes a month on Premium ($35/mo, $28 annual), then $0.15 a minute. Overnight triage calls tend to be short, mostly under two minutes once the questions are tuned, so even a business fielding a handful of overnight calls a week often stays within or close to the free allowance. A wider set of options is in how to answer phones after hours without hiring.

What I'd actually tell you

Resist the urge to make the urgent definition too broad out of caution. If "urgent" covers too much, you're back to being woken for things that could have waited, which defeats the entire point. Write your urgent definition narrow and specific, test it against real past situations you remember handling, and widen it only if you find a genuine miss, not preemptively. It's far easier to loosen a definition that's too tight than to tighten one that's letting too much through. Checking what came in overnight is covered in how to check after-hours and weekend voicemails quickly on Monday.

Common mistakes with overnight triage

• Defining "urgent" once and never checking it against real calls. A definition that sounded right on paper often misses real edge cases until you actually review a few weeks of logs and see what got mis-routed.
• Letting the on-call rotation go stale. An urgent call routed correctly still fails if it rings through to someone who's no longer actually on call that week because the schedule wasn't updated. The triage logic only works if the routing destination is kept current.
• Treating every non-urgent message the same way regardless of how it's worded. A caller who sounds distressed, even about something that technically fits your "can wait" definition, is worth a slightly different message format than a routine, calm request, so whoever reviews it in the morning can prioritize their callbacks sensibly rather than working through messages in a random order.

A second scenario: a locksmith with a narrower definition

Say a locksmith defines urgent specifically as "locked out and unable to safely stay where you are," which deliberately excludes "locked out but safely at a friend's or at work until morning." A caller at midnight who's locked out of their car in a well-lit, safe parking lot with a working phone gets asked by the AI agent whether they're somewhere safe to wait until morning; a yes routes to a message, a no (say, stranded somewhere with no safe option) routes straight through live. This narrower definition means fewer overnight wake-ups than a blanket "any lockout is urgent" rule would produce, without leaving anyone genuinely at risk waiting till morning. The specificity of the definition is doing almost all the work here, not the technology sorting it.

When to skip the tool

Don't rely on any AI system, however good, as the sole gatekeeper for a true emergency where getting it wrong has serious consequences, a medical situation, a gas leak, anything with real safety risk. Those categories deserve a direct, always-available human path, not a triage question flow, even a well-designed one. Keep a visible, separate emergency number for exactly that category and don't route it through the AI agent at all.

How I'd approach this

Where overnight triage becomes more than a settings page is when the urgent-vs-not decision should actually check real data, is this address already flagged as a recurring problem, does this customer have an active job that makes the call genuinely higher priority, rather than relying purely on what the caller says in the moment. That's a connection to your own job or customer records, and it's a real build, not a script tweak. Get in touch if your triage needs to know more than what the caller tells it.