In a store, replying to a customer doesn't always take much time. What often takes more is finding the information needed to answer correctly: purchase history, order, invoice, quote, booking, product availability or a previous conversation.
Artificial intelligence can step in at several levels of customer service: finding information, preparing a reply, analyzing recurring requests or, in some cases, automating part of the conversations. But not all of these uses carry the same level of risk or require the same level of control.
Two main uses of AI in customer service
First, two situations need to be told apart.
AI assists the merchant
It works in the background. The merchant can ask it to find information, summarize a history, prepare a reply or look up a document. The customer does not talk directly to the AI.
AI replies directly to the customer
This is the case with a chatbot, a conversational assistant or an automated system able to answer certain requests. This use can suit simple, well-documented questions, but calls for more caution when the answer depends on the customer's particular file or can have a commercial consequence.
This distinction is essential: helping someone reply and replying directly in their place do not carry the same risks. See also the difference between an AI agent and an AI assistant.
Find a customer's context faster
A large part of customer service work consists of looking up information that is already in the store's tools. For example:
“Find this customer's latest purchases.”
or:
“Has he already contacted the store about this order?”
When the necessary data is accessible, AI can help gather the useful elements:
- previous purchases;
- orders;
- quotes;
- invoices;
- bookings;
- recorded conversations;
- information tied to the customer file.
The goal is to avoid manually searching for the same information across several screens or documents. The final reply can then be prepared from data specific to the customer's actual situation, rather than from a generic answer. See also viewing and searching a customer record.
Prepare a customer reply
Once the context is found, AI can help write a reply. For example:
“Prepare a reply to this customer asking where his order is.”
The value is not only in the writing. If the system has access to the right information, the reply can take the context into account: the order concerned, its known status, previous conversations or information available on the file. The merchant can then check the content before sending it.
This validation is particularly useful when the reply is about:
- a complaint;
- a delay;
- a refund;
- an error;
- a commercial discount;
- an unusual situation.
AI can speed up preparation without necessarily making the decision in the merchant's place.
Find quotes, invoices and documents
Customer requests frequently involve documents.
“Find this customer's invoice.”
“Which quotes sent this month are still unanswered?”
“Prepare a follow-up for this quote.”
An AI connected to document data can make it easier to search and to prepare the corresponding actions, relying on the documents already managed in Gillia quotes and invoicing. However, three steps need to be distinguished:
find the document → prepare the action → carry out the action
The first generally carries little risk. The last can have a real consequence and justify validation depending on the context. See also viewing an invoice or viewing a quote. When the invoice is overdue, the approach is different: see our guide on unpaid invoices and customer reminders.
Help manage bookings
For businesses that use bookings, AI can also help with certain requests. For example:
“Find Mrs. Martin's booking for Friday.”
Then:
“What availability do we have one hour later?”
And possibly:
“Prepare moving her booking to 8:30 p.m.”
Here again, looking up information and actually changing a booking are two different operations. The level of control can be adapted to the action taken, in Gillia's booking capability as well as for changing a booking.
Answer frequently asked questions
Some requests do not require access to a customer file. For example:
- what are the opening hours?
- what are the return conditions?
- which payment methods are accepted?
- what are the delivery terms?
- how does in-store pickup work?
This information can be grouped in a knowledge base. An AI can then use this documentation to help the merchant or, when the setup is designed for it, reply directly to the customer. The value of a knowledge base is significant: it provides a controlled source from which to build answers. This reduces the risk of producing an answer based only on the model's general knowledge.
An AI must not invent customer information
This is probably one of the most important rules. If the AI does not know the status of an order, it must not make it up. If it cannot find a return condition, it must not infer one. If an invoice is missing, it must not assume its content.
Good use of AI in customer service should make it possible to tell apart:
- information found in the data;
- information coming from documentation;
- a suggestion written by the AI;
- information that is not available.
A fast but wrong answer is generally more problematic than an answer that requires a check.
Personalize messages without over-interpreting the data
Customer history can help tailor a reply or a communication. For example, a customer who has bought a particular category several times can receive a message about that category. But history cannot tell you with certainty what a person will want in the future. You should therefore avoid turning every data point into a conclusion. AI can use history as context, without claiming to know the customer's preferences precisely.
Detect requests that come up often
AI can also be used on a broader scale. If many requests are about the same topic, it can reveal a problem or information that isn't clear enough. For example:
- several questions about delivery times;
- repeated requests about one item;
- many complaints tied to the same product;
- frequent questions about returns;
- quotes that regularly go unanswered.
Analyzing these requests can help identify recurring problems rather than handling each conversation in isolation. A question that keeps coming back can sometimes be solved more effectively by improving an information page, a procedure or a communication.
Prioritize requests
Not all requests are equally urgent. A question about opening hours and a complaint about a payment should not necessarily be handled the same way. When the necessary data and rules are available, a system can help bring out:
- urgent requests;
- old cases;
- complaints;
- messages requiring human intervention;
- simple questions that can be handled quickly.
AI then becomes a tool for sorting and prioritizing, not just for writing.
How far should you automate customer service?
There is no single rule that holds for every situation. The level of automation can depend on:
- the type of request;
- the reliability of the data;
- the potential impact of a mistake;
- the store's policy;
- the rights granted to the system.
An answer about opening hours can generally be automated more easily than a refund or a major commercial change. A useful approach is to distinguish three levels:
Low risk
Looking up information, summarizing, sorting or replying from validated documentation.
Medium risk
Preparing a message, a follow-up or a change. Validation can be useful before execution.
High risk
Refund, sensitive cancellation, financial change or any other action with a significant consequence. These operations generally require more control.
The more binding the action, the more important validation and traceability become.
Watch out for personal data
Customer service naturally handles customer-related information. The use of AI must therefore take into account the rules that apply to this data: access, retention, purpose, security and individuals' rights.
Not all the information available in a system needs to be accessible to every user or used for every task. The principle remains the same as for the store's other tools: access to data must match the real need and the rights of the person carrying out the action.
Measure the real value of AI
The number of replies generated is not necessarily the best metric. The goal is to improve service. Depending on the use, you can track in particular:
- the average time needed to find information;
- the first response time;
- the handling time;
- the number of requests resolved;
- the reopening rate or rate of new requests on the same topic;
- customer satisfaction when it is measured;
- the share of replies requiring human correction.
This last metric is particularly interesting. An AI that writes very fast but whose replies must be systematically corrected brings less value than a slightly slower but reliable system.
How can Gillia help customer service?
Gillia is positioned mainly as an agent used by the merchant to make use of the information available in their work environment. For example:
“Find this customer's record.”
Then:
“Find the invoice concerned.”
And:
“Prepare a reply to his request.”
Depending on the data and capabilities available, Gillia can also help find a quote, a booking or the history relevant to the situation. The goal is to bring together:
context → analysis → action preparation
from a single conversation. Actions with a real impact can remain subject to validation before they are carried out — the very principle of an AI agent for retail. Gillia can also rely on a knowledge base for documented, recurring information.
From the individual reply to better service
One of the most important benefits of AI appears when you go beyond simple writing. A customer request can be handled faster. But a hundred similar requests can reveal something more useful: why does this question come up so often?
- Maybe some information is missing on the website.
- Maybe a procedure is unclear.
- Maybe a product causes an unusual number of problems.
- Maybe a step in the customer journey needs improving.
AI can therefore serve at two levels: handling an individual request more efficiently; and bringing out the recurring problems that deserve a broader improvement.
It is at this second level that it can truly help improve customer service, rather than simply speeding up the writing of replies. The loyalty actions that follow can be found directly in Gillia's loyalty capability. More examples related to the customer relationship are in our catalog of Gillia use cases: following up on a pending quote or changing a booking.