AI agents for customer service: what to automate and what to keep human
An AI agent for customer service is best used for four things: sorting incoming requests, summing up long exchanges, finding the answer in your own documents and drafting the reply. An AI agent is a software helper that reads, thinks and hands back a result. Complaints, exceptions and anything about money should stay with a person.
The test: is the answer already written down?
Customer requests come in two kinds. For the first kind, the answer exists. What are your opening hours? How do I return something? Where is my order?
For the second kind, someone has to decide. A refund outside the usual rules. An angry customer. A problem nobody has seen before.
An agent handles the first kind well, because it can look the answer up. The second kind needs judgement, and often an apology that means something.
Request by request: who does what
| Type of request | What the agent does | What the person does |
|---|---|---|
| Sorting new requests | Sorts them by topic and urgency | Checks a few now and then |
| Questions your documents answer | Drafts the reply and names the source | Reads it and sends it |
| A long exchange handed to a colleague | Sums it up | Takes over, fully informed |
| Refund or dispute over a bill | Gathers the facts | Decides and replies |
| Complaint or upset customer | Marks it as urgent | Handles it personally |
Why helping the team is the best place to begin
Many people picture an agent that answers customers by itself. The safer gain comes one step earlier. Every request arrives sorted, summed up and with a draft reply. Your colleague deals with it in far less time, and the customer still gets an answer a person has read.
Nothing wrong leaves the company unnoticed. Over time the team also learns which drafts are always right.
Rules for answers that go to customers
- Answer only from your own documents. If they do not cover the question, the agent says so and passes it on.
- Always offer a person. Give one clear way to reach a human.
- Say that it is an AI. Customers accept an automated answer. They dislike finding out later.
- Never promise. Refunds and delivery dates are commitments, and only people make them.
What to prepare before you start
- Write the answers down. Collect your twenty most frequent questions and write one clear answer for each. The agent works from this document.
- Bring requests into one list. Email, forms and chat should land in the same place, each with a status and a person in charge.
- Say who gets what. Who takes complaints, who takes billing, who covers holidays.
- Keep the answers current. When a rule changes, change the document the same day. An outdated document produces wrong answers.
How to tell whether it works
Watch two things from the first week: how long a customer waits for a first answer, and how many drafts are sent without changes. If drafts are often rewritten, improve the documents first.
How Worknano handles this
In Worknano every request to a team becomes a ticket, one entry with a status, in that team's queue under Requests. It can come from a colleague or from outside through a public form. The Requests Assistant tells you where a ticket stands, sums up a queue and helps write replies. Before it answers a question it searches the workspace, including the team's documents on the Docs page, and reads the matching document. It finds only what the person asking may read. It does not change a ticket's status by itself. Agents can also run as steps in a workflow, for example to sum up a long ticket before it is assigned.
- Tickets and client requests: every ask to a team in one place
- How to run the ticket inbox on a project
Create a free workspace and set up your first support queue with an assistant.