AI assistant in support: how to introduce it and what to measure
Five stages of introducing an assistant into support, ten topics with a verdict, six metrics, seven rules of a good handover to a person and common mistakes.
In short
An AI assistant in support takes over the repeating questions — where is the order, how to return, what size to choose — and answers them at any hour from your knowledge base, while people get the cases that need judgement. It is introduced in stages: first it helps the team with drafts, then answers the most frequent topics itself, and only then takes on more. Success is measured not by the number of answers but by the share of questions closed without a person, the accuracy of answers, the quality of handovers and customer ratings. The most important part is a good handover: when the assistant does not know or the customer is upset, the conversation goes to a manager with the whole history.
How to introduce it: 5 stages
-
1. Analyse the requests
A few hundred real messages sorted by topic show what repeats and what takes the most time.
-
2. Drafts for the team
The assistant prepares answers, managers edit and send them — and the edits show where it is wrong.
-
3. Frequent topics on its own
Two or three topics with clear rules — order status, delivery, returns — go to the assistant entirely.
-
4. Expansion
New topics are added one at a time, each after a check on real questions.
-
5. Regular care
Once a week — the log: questions without an answer turn into new documents in the base.
Which topics to give to the assistant
Ten typical topics of support with a verdict.
-
Where is my order
AssistantStatus from the system by the order number — the most frequent question.
-
Delivery and payment terms
AssistantClear rules from the knowledge base.
-
How to return or exchange
AssistantExplains the procedure and starts a request.
-
Choosing a product
AssistantSizes, materials, comparison — from the catalogue.
-
Changing an order
With a personThe address — by itself; the composition and the price — through a manager.
-
A defect
With a personCollects photos and details, the decision is made by a person.
-
Refund of money
With a personWithin a set limit — by the rules; above it — a manager.
-
A complaint
A personAn upset customer needs a person; the assistant hands over at once.
-
Disputes and claims
A personLegal consequences — only people.
-
Large wholesale orders
A personThe value of a personal conversation is higher than the saving.
What to measure
Six numbers that show whether the assistant helps or only answers.
| Metric | What it shows | How to count |
|---|---|---|
| Closed without a person | the real relief for the team | dialogues without a handover and without a repeat request |
| Accuracy of answers | whether the answers can be trusted | a weekly sample checked by a person |
| Handovers | where the knowledge base has gaps | share and topics of handovers |
| Customer rating | how people feel the conversation | one question after the dialogue |
| Time to the first answer | speed, especially at night | from the log |
| Price of a closed question | whether it pays off | the bill for the model divided by closed questions |
7 rules of a good handover to a person
-
01
At the first “person, please”
Without persuading to stay with the bot.
-
02
When there is no answer
An honest “I will pass it to a manager” instead of a guess.
-
03
When the customer is upset
Emotion and repeated questions are a signal to hand over, not to explain again.
-
04
With the whole history
The manager sees the dialogue and does not ask the customer to repeat.
-
05
With a short summary
Two lines: what the customer wants and what has been done.
-
06
With an honest time
At night — “a manager will answer in the morning”, not “in a minute”.
-
07
Every handover — a lesson
Topics of handovers show what to add to the knowledge base.
Common mistakes in AI support
-
Everything at once
The assistant answers all topics from the first day, and mistakes in rare ones spoil trust in the rest.
-
Hiding the person
A customer who cannot reach a manager leaves and tells others.
-
An outdated base
Terms changed, the documents did not — the assistant confidently quotes the old ones.
-
Pretending to be a person
The customer finds out and stops trusting the whole company.
-
Counting answers
A thousand answers mean nothing if half the customers write again.
-
Nobody reads the log
The assistant stops improving the day after launch.
Questions about AI in support
What share of questions can an assistant close?
It depends on how repetitive the questions are; the analysis of requests at the first stage gives an honest estimate.
Will customers be unhappy talking to AI?
Not if the answers are fast and correct and a person is always one message away.
Does the assistant replace the support team?
It takes the routine; the team handles difficult cases and improves the knowledge base.
Where does the assistant work?
On the site, in Telegram and WhatsApp, in email — one knowledge base for all channels.
Can it see orders?
Yes, through a connection to the shop or CRM — only the orders of the customer it is talking to.
How long until results show?
The drafts stage helps the team at once; the share of closed questions grows topic by topic.
Online form
An assistant
for your support
I introduce assistants in stages: drafts for the team first, then the frequent topics — with a knowledge base, a handover to a manager and a log of questions. Tell me about your support — I answer within one working day.