Designing the moment it gives up

The failure customers actually remember

Nobody minds a bot that cannot answer. Everybody minds a bot that will not let them past.

The loop is familiar. Three rephrasings of the same non-answer. “I’m sorry, I didn’t understand that.” No visible way out. The customer leaves, and you never learn that they were trying to give you money.

So the handover (the moment the system stops and fetches a person) is not an edge case bolted on at the end. It is the part of the design that decides whether the whole thing helps or harms.

Four triggers, all cheap to implement

Two failures, then stop. Not five. If the assistant has not helped after two attempts, it will not help on the third, and each additional attempt costs goodwill that is hard to get back.

Anything about money or a commitment. A price for custom work, a discount, a refund, a delivery promise, a change to an existing order. Handover on sight.

Emotional signal. Anger, distress, urgency, or the word “complaint”. A person, immediately. A bot handling a upset customer well is a rare event; a bot handling one badly is the review you will read for two years.

On request, instantly. “Human”, “person”, “speak to someone”, “agent”, always works, at any point, no qualifying questions. Anything else is a dark pattern, and customers recognise it as one.

The handover itself

Two rules.

Carry the conversation across. The customer has already typed their order number and explained the problem. Making them repeat it is the moment they conclude the automation was for your benefit rather than theirs. Whatever tool you use must pass the transcript to the person.

Be honest about the wait. “Kemi will reply within four hours: we’re open until six” is a promise you can keep. “Connecting you to an agent” when nobody is there until Monday is worse than saying nothing, because it converts a mild disappointment into a broken promise.

Out of hours

Most small businesses are not staffed at 11pm, and the honest version wins:

We’re closed now. I can answer questions about prices, opening hours and delivery areas. For anything else, leave a message and Kemi will reply by 10am tomorrow.

That sets a boundary, delivers value inside it, and makes a keepable promise. Compare it with a bot that pretends to be a full service at midnight and produces an invented answer nobody will honour in the morning.

The transcript is the diagnosis

Every handover is a small report on where your system falls short. Read them weekly and sort them into three piles.

Should have been answered. A gap in your documents. Write the missing page.

Correctly escalated. Working as designed. Note the volume; if this pile is large, the bot is scoped too narrowly, or too broadly.

Gave up before asking. The saddest pile and the most useful. Look for the last message before the silence. It is usually the same three or four questions, and they tell you exactly what is missing.

The measure that matters

Not deflection rate. Every vendor reports it, and it counts conversations that did not reach a human, which includes every customer who left in frustration. Deflection and abandonment look identical in that number.

Measure instead: how many enquiries turned into what you wanted: a booking, an order, an answered question the customer confirmed. That is harder to count and it is the only one that distinguishes a bot that helped from a bot that got in the way.

Staffing the other side

A handover promise is only as good as the person on the other end. Before you turn anything on, decide three things: who receives handovers, within what time, and what happens when they are on holiday.

If the answer is “the owner, whenever she sees it”, say that honestly in the bot’s wording rather than implying a support desk. Small businesses are trusted for being small. Pretending to be large is the thing that breaks trust, not the four-hour wait.

The transcript belongs in the record

When a conversation is handed over, the transcript should end up wherever you keep customer history: the helpdesk, the CRM, or an email to yourself.

Two reasons. The person picking it up has the context. And if a dispute follows, you have what was actually said rather than the customer’s recollection of what the bot promised. Given that a bot’s statements may bind the business, a record of its statements is not optional bookkeeping.