Blog · Chat
When an AI assistant should hand over
The question that stops projects is not “how well does it answer”, it is “what if it gets it wrong?”. A trustworthy assistant does not know everything: it knows where its knowledge ends. The four situations where it must stop, and what a good handover looks like.
The question that stops most automated support projects is not "how well does it answer", it is "what if it gets it wrong?". It is the right question, and the answer is not "it will not".
A trustworthy assistant is not one that knows everything. It is one that knows where its knowledge ends, and hands over before that line rather than after it.
The four situations where it has to stop
When the data is not there. If an answer needs information that is not in the store's data, the only correct answer is to say so and hand over. The worst case is not silence, it is the plausible invented answer: a customer acts as though it were true, and the damage surfaces later.
When there is a complaint. A parcel that arrived broken, an order never dispatched, a wrong charge. This is not about technical capability: an angry person wants a human, and routing them through two rounds of automated questions makes a bad situation worse.
When money outside the price list is involved. Refunds, exceptions, unplanned discounts, special terms. Those are commercial decisions, and an automated commercial decision is a decision nobody made.
When the same person has asked twice. If somebody rephrases the question it is because the first answer did not help. Pressing on with a third variant of the same answer is the surest way to end the conversation.
What a good handover looks like
The handover matters almost as much as the decision to make it.
Do not make them repeat anything. Whoever picks up the conversation must see everything said and the order in question. Asking again for the order number after the customer already typed it is the moment a person realises they are talking to a disconnected system.
Say what happens now. "A person will reply within X" is different from silence. Even when the wait is long, knowing it reduces irritation more than shortening it would.
Do not pretend. An assistant posing as a person works until it is found out, and from that moment everything it said before becomes suspect.
The number to watch is not how many it deflects
The standard metric is the share of conversations handled without a person, and on its own it pushes the wrong way: you get a very high number simply by never handing over, with customers leaving instead of complaining.
It has to be read with two others. How many resolved conversations had a follow-up in the days after, which is the signal the answer had not actually resolved anything. And how many people explicitly asked to speak to somebody, which is the most honest measure of trust in the system.
What gets prepared beforehand
Most avoidable handovers are avoided by writing down the exceptions, not by improving the model. Shipping terms to islands, the thresholds where something changes, what happens if a product arrives faulty, which categories have a different returns policy.
These are all things somebody in the company knows by heart and that are written down nowhere. While they stay in that person's head, the assistant will hand over every time somebody asks, and most questions are exactly those.
How Uptonica handles this
Lia answers by reading orders, availability and the store's real terms, and when a question goes beyond that it hands the conversation to a person in the same inbox: whoever picks it up sees everything already said.
See what Chat doesFrequently asked questions
What is a reasonable share of conversations handled without a person?
It depends entirely on how repetitive the questions are and how complete the information is. Where questions are few and recurring you get high; on catalogues with many exceptions the share is naturally lower, and that is fine. A high share achieved by never handing over is not a good result, it is a risk.
Should you say upfront that it is an automated assistant?
Yes. Somebody who knows calibrates their expectations and asks the right things, and at handover they do not feel deceived. The cost is nothing, the benefit is trust when something goes wrong.
How do you stop it making things up?
By tying it to the store's data rather than its general knowledge, and by making "I do not know" an acceptable answer. A system forbidden from admitting ignorance will always produce something, and sometimes that something will be false.
Do I need a person available at all hours?
No, you need the wait to be stated. "We will reply tomorrow morning" with the conversation saved works; silence does not. What to avoid is a handover to nobody.
Out of hours, better to answer badly or not answer?
Answer on what is known, say the rest comes tomorrow. Night is also when the assistant is worth most, because there is no alternative: provided it does not try to cover what it cannot know.