Most small stores run support the way they run the shop: during business hours. Nights, weekends, and the lunch rush are dead zones, and the messages that arrive there sit unanswered until someone opens the app again. The usual fix — point an auto-reply bot at the inbox and let it answer everything — trades one problem for a worse one, because an unattended bot that can send directly will eventually send something wrong at 2 a.m.
A workable overnight setup is not a switch. It is a ladder: each rung handles more of the load, and the rung that decides what actually gets sent stays with a human until morning.
Why the naive version fails
The failure is not that the bot answers badly. It is that nothing separates messages that are safe to answer from messages that carry money or commitment in them:
- Price and stock drift. A canned "yes, in stock, ships today" is correct at 9 a.m. and wrong at 11 p.m. when the last unit sold and nobody updated the sheet.
- Promises become disputes. "Refund approved, expect it in 3 days" sent overnight by a bot is a commitment the store has to honor, made by software that could not see the order history.
- Escalation has no one to escalate to. An "I'll get a human" reply at 3 a.m. with no human until 9 a.m. is worse than silence — it resets the customer's patience clock.
The three rungs
Rung 1 — Acknowledge, don't answer. The overnight job of the bot is to tell the buyer that a person has seen the queue and when they will get a real answer. A first response that says "received, we'll reply by 10 a.m." holds the window open; a wrong answer burns it.
Rung 2 — Draft, don't send. The bot writes candidate replies into a review queue instead of the customer's chat. In the morning the operator opens the queue, reads twenty drafts, edits three, approves seventeen. Ten minutes of review replaces an hour of writing from scratch — and nothing went out unsupervised.
Rung 3 — Hard-block the money categories. Price, refund, compensation, shipping-time changes, and anything touching a complaint are never drafted as final text. They get a short acknowledgement plus a flag. This list is written down once and does not change because the model got better.
Designing the trigger and the cut-off
Two settings do most of the work:
- Active window. Define the hours the bot may send acknowledgements and hold drafts. Outside it, it only collects. If your support is genuinely 24-hour, the window is 24-hour — but then the review queue has to be cleared on a shift schedule, not "in the morning."
- Morning sweep. One fixed pass over the overnight queue, in a set order: flagged money items first, then anything with two or more messages from the same buyer (they are still waiting), then everything else. Buyers who asked twice get answered before buyers who asked once.
The draft-review loop also produces the raw material for a daytime knowledge base. Every draft an operator rewrites is a rule that should have existed; collect those and the overnight queue gets shorter every week.
What to measure
Three numbers, checked weekly: the share of overnight messages answered before the store opens (should climb toward one), the count of drafts approved without edits (the bot's real accuracy signal, not a benchmark score), and the number of overnight messages that needed a manual fix after sending — this should stay at zero, because rung 3 means those were never sent at all.
After-hours coverage is the clearest place where "AI assists, human decides" pays for itself. The volume is real, the risk is concentrated in a handful of categories you can name in advance, and the review cost is bounded by however many messages actually arrived.
Top comments (0)