How to answer customers 24/7 without a night shift

Almost every support tool promises 24/7. Very few explain what running it actually involves: which questions are safe to answer while nobody is watching, what it costs to reply at nine in the morning instead of two in the night, and what has to be decided before your team goes to sleep. This is that guide.

Start by counting your own nights

Before you automate anything, spend ten minutes with the inbox you already have. Take last week and pull out every message that arrived outside your working hours, then sort it into three piles: the ones you answered in the morning and the customer was still there, the ones you answered and the customer had already gone quiet, and the ones nobody ever answered. The third pile is what 24/7 is actually being bought to fix, and it is usually larger than anyone on the team expects, because an unanswered message leaves no trace in a revenue report.

That count also settles the comparison worth making. The choice is rarely between an AI agent and the status quo; it is between an AI agent and paying people to sit through the quiet hours. A week has 168 hours, so covering every one of them with staff is more than four full-time shifts before you add holidays, sick days and the handover conversation at the start of each one. For most businesses the overnight volume does not justify that, which is exactly why those messages go unanswered.

The overnight clock that actually costs money

On WhatsApp there is a second reason to answer at night, and it is on the invoice. When a customer messages you, Meta opens a 24-hour service window, and every reply you send inside that window is free. Once it closes, reopening the conversation needs an approved template message, and Meta bills for it.

So follow one message through. A customer writes at 7:40pm on Friday. If the agent answers at 7:40pm, the reply is free and the customer has their answer while they still care about it. If it waits for Monday at 9am, the window shut on Saturday evening, and getting back in touch is now a paid template sent to someone whose question is three days stale. Answering overnight is not only the better service, it is the cheaper of the two options. We break Meta's rates down in WhatsApp API pricing explained.

Instagram and Messenger have their own messaging windows set by Meta. Telegram and the website chat widget have no per-message fee at all, which is a decent reason to have one of them on while you are still deciding about the rest.

Decide what the night is allowed to answer

This is the whole job, and it takes about twenty minutes. The rule to hold in mind is that a wrong answer at three in the afternoon gets corrected in five minutes, and the same wrong answer at three in the morning has eight hours to sit there. Overnight, the line between answer and hand over should sit further towards handing over than it does in the day.

Write that list down before you go live rather than discovering it from a bad transcript. Everything in the top half is what removes the morning backlog; everything in the bottom half is what keeps the agent from embarrassing you while you sleep.

Write the three o'clock sentence before you need it

The handover line that works at 2pm does not work at 2am. "Let me get a colleague to check that for you" is true within minutes during the day and quietly false at night, and a customer who waits an hour for the colleague who was promised is more annoyed than one who was told the truth.

Say the hour instead. Something close to: "Our team opens at 8am and will pick this up first thing. I have saved everything you have told me, so you will not have to explain it again." It sets a real expectation, it is honest about there being nobody at a desk, and it tells the customer the repeating is over. Conversations the agent hands over are flagged in the shared inbox with the whole thread and whatever it already collected, so the person who opens the queue at 8am starts with the order number, the photo and the question rather than with "hi, how can I help".

Decide what is worth waking someone for

Round-the-clock coverage fails in two opposite directions. Nothing reaches a person until the morning, and a genuine emergency sat in a queue all night; or everything reaches a person, the phone buzzes six times before dawn, and within a fortnight the alerts are muted and it may as well be the first case.

Pick one narrow category that can wake someone. For most businesses it is anger, a safety issue, or a named account that has earned it. Set escalation alerts for that and let the rest accumulate in the shared inbox for the morning. Automatic follow-up messages are rate-limited per customer as well, so a quiet overnight conversation cannot turn into a string of nudges while somebody is asleep.

Your night is someone else's Tuesday afternoon

Overnight traffic does not look like daytime traffic, and the biggest difference is where it comes from. The hours your team is asleep are working hours somewhere else, so the share of messages in a language nobody in the office speaks is higher at 3am than at 3pm. The agent detects each customer's language and answers in it across 50+ languages from one knowledge base, rather than a separate setup per language, which is the only version of this that is maintainable.

The format changes too. Late-night messages skew towards voice notes from people who cannot be bothered to type and photos of receipts, screenshots and delivered items. Both are handled: voice notes are transcribed and answered like any other message, and images are read, so a photo of the wrong item sent at midnight becomes a logged complaint with a picture attached rather than a mystery waiting for the morning.

The morning routine is the part people skip

Automating the night creates a small new habit, and it is fifteen minutes. Open the night's conversations in order, including the dull ones, and look for exactly three things.

1. Questions that were answered badly. These go into the knowledge base today, not on a backlog. Overnight gaps repeat, because the people asking at 2am tonight are much like the people asking at 2am tomorrow.
2. Conversations that were handed over and did not need to be. Every unnecessary escalation is a morning task you created for yourself. Tighten the rule or fill the gap in the knowledge that caused it.
3. Questions that were answered well and came up repeatedly. This is the most valuable output of the night and the one everybody ignores. It is a free, unfiltered read on what customers actually want, and it belongs in front of whoever writes your website copy.

Replio also proposes answers it has learned from real conversations for you to approve with one tap, so the knowledge base improves from the night's traffic without anyone rewriting it.

One night, hour by hour

9:40pm: a WhatsApp message asks whether you deliver to a suburb on the other side of the city. The knowledge base knows, the answer goes out in seconds inside the free service window, and the order is placed before the customer falls asleep. 12:15am: a voice note in Portuguese, transcribed, answered in Portuguese, about opening hours over a public holiday. 3:50am: a photo of a damaged item with the word "unacceptable" in the caption. The agent does not try to fix it, logs the complaint with the photo, replies that a person will handle it when the team opens at eight, and, because you set anger as the one thing worth a phone buzzing, the duty manager sees it. 6:20am: an appointment booked straight inside the chat for Thursday. 8:05am: someone opens a queue with four conversations in it, three already finished and one that needs a person and arrives with the photo, the order number and the customer's own words already attached.

What 24/7 does not mean

It does not mean nobody needs to be reachable. If your business can have an outage, a stuck delivery or a safety problem at night, you still need a human on call; the agent's job is to make sure that person is woken for the right thing rather than for everything. It does not mean the agent knows anything you have not told it. And it does not mean your team's hours change: every message a customer sends is answered, but the ones that need judgment are waiting, organised, for the morning rather than being decided by software at 4am.

What it costs

Starter is $49 a month and covers one of WhatsApp, Instagram or Messenger plus Telegram and the website widget, with 2 seats and 8,000 AI replies a month. Scale at $149 runs WhatsApp, Instagram and Messenger together in one shared inbox with 5 seats and 30,000 AI replies. Pro at $349 has 15 seats and 80,000. Customer messages are unlimited on every plan, and replies typed by your own team are free and unlimited too, which matters here: handing a conversation to a person costs you nothing. On WhatsApp, Meta charges its own per-message rates directly to you at its published prices, with no markup from Replio. The free plan covers Telegram and the website widget, and the trial runs 7 days without a card.

Frequently asked questions

Related

Last updated 2026-09-23. Meta sets the messaging windows and per-message rates described here and may change them, so please check Meta's current documentation alongside this guide.