How to run 24/7 support with three people and an AI agent

The classic advice to a small team asking about 24/7 support was simple: don't. A week has 168 hours; a full-time agent covers 40 of them. Before you account for weekends, holidays, sick days and the fact that humans occasionally sleep, honest around-the-clock human coverage needs a minimum of five full-timers — or an outsourcing contract that trades your product knowledge for a script. For a three-person team, the math simply said no.

The math has changed, because the definition of "coverage" has changed. Coverage never actually meant "a human is awake." It meant "a customer at 3 a.m. gets a useful response, and anything urgent reaches a human fast." Those are two different requirements, and only the second one needs a heartbeat. This guide is the operating manual for meeting both with three people and an AI agent.

Audit what actually arrives at night

Before designing anything, pull three months of conversations that arrived outside your business hours and sort them into three piles:

• Answerable from documentation. Password resets, billing questions, how-do-I tasks, plan comparisons, configuration steps. For most SaaS products this is 60–80% of overnight volume.
• Needs a human, can wait until morning. Refund approvals, negotiated exceptions, complex debugging, anything involving judgment about a specific account.
• Needs a human now. Production outages, security incidents, payment failures blocking a purchase. Typically 1–3% of overnight traffic — but this pile is where reputations are made or lost.

The proportions are the design input. If pile one dominates (it almost always does), the bulk of your night shift is a documentation problem wearing a staffing costume.

Build the front line: a knowledge base the AI can actually use

An AI agent answers from what you've written down, which makes knowledge-base quality the ceiling on answer quality. Writing for an AI reader differs slightly from writing for humans:

• One intent per article. "Billing FAQ" containing nine loosely related answers retrieves poorly; "How to change your billing email" retrieves precisely.
• State the conditions. "This applies to annual plans only" saves a wrong answer at 3 a.m. that a human would have caught by glancing at the account.
• Cover the top intents first. Rank last quarter's questions by frequency; the top twenty intents usually cover more than half of total volume. Write those twenty articles before anything else.

Then define what the AI must not answer. Refunds above a threshold, legal questions, security reports, cancellation saves — these route to a human every time. An AI agent is a front line, not a court of final appeal, and customers forgive "I've flagged this for the team, first thing tomorrow" far more readily than a confidently wrong answer.

In Sufox this pairing is native — the AI agent answers from your knowledge base through the same inbox your team works in, and escalates with the full conversation attached — but the design principle is tool-independent: the AI resolves the documented tier, and everything else lands in a queue with context, not a cold start.

The three-person rotation

With the front line handling the documented tier, the human problem shrinks from "staff every hour" to "guarantee escalation response." A rotation that works in practice:

• Stagger working hours if you can. Even a modest spread — one teammate starting three hours earlier, one ending three hours later — stretches live coverage to 14–16 hours on weekdays without anyone working nights.
• One on-call phone, one week at a time. The on-call person responds only to outage-class escalations — the 1–3% pile — routed through an alert, not by watching the inbox. A well-designed week of on-call is quiet; if it isn't, the alert rules are too loose.
• Morning triage as a ritual. The first thirty minutes of each day clear the overnight queue: verify the AI's resolutions, answer the deferred pile, and — this is the compounding step — turn every question the AI couldn't answer into a knowledge-base article. Overnight deflection rises month over month because each gap becomes tomorrow's coverage.

Notice what this rotation doesn't include: night shifts, weekend hero work, or an inbox that anyone checks from bed. The system absorbs the night; humans absorb only the exceptions.

Set expectations honestly

A small team running AI-fronted 24/7 coverage should say exactly that. A status line like "instant answers around the clock; a human reviews every escalation within one business day, outages get paged immediately" is a promise you can keep every week of the year. It also beats the alternative most small teams actually ship — a silent inbox for fourteen hours a day — by a distance no customer will fail to notice.

What you should not promise is live human chat at 4 a.m. That claim belongs to the five-plus-agent world, and pretending otherwise burns the trust the whole model depends on.

The numbers to watch

Four metrics tell you whether the machine is working:

1. Overnight resolution rate — the share of night conversations resolved without a human, with no follow-up or reopen within a few days. Expect it to start near your documentation coverage (often 40–50%) and climb as morning triage feeds the gaps back in.
2. Escalation precision — of the conversations the AI escalated, how many truly needed a human? Too high a false-alarm rate means your routing rules are timid; too low means they're reckless.
3. Morning backlog age — how long the oldest deferred conversation waited. If it regularly exceeds twelve hours, your stagger needs widening.
4. Pages per week — on-call alerts should be rare events. A noisy pager is a misconfigured pager.

On volume: a team fielding, say, 900 conversations a month with 65% overnight-and-daytime deflection consumes well under the 1,000 AI answers included in Sufox's $99 Starter plan — this model is not a premium-tier luxury.

When this model is the wrong answer

Honesty about the edges: contractual 24/7 human response for enterprise SLAs, regulated industries where an AI may not give account-specific answers, and phone-first customer bases all break the pattern — those situations genuinely require staffed shifts or a follow-the-sun partner. And if your overnight pile-three rate is far above 3%, you don't have a support-coverage problem; you have a reliability problem that no rotation will fix.

For everyone else — which is most SaaS teams under twenty people — the sequence is: audit the night queue, document the top twenty intents, put the AI in front with strict escalation rules, stagger three humans, and let morning triage compound the coverage. The 168-hour week stops being a staffing problem and becomes what it should have been all along: a well-designed queue.