AI agent for customer service: answer first, hand over what needs a person
This template puts a conversational agent at the front of customer service. It answers from content you approve, and when a customer asks for a refund or an exception it stops, summarises the conversation and passes it to a person.
Last reviewed 7 October 2026
Starting point
The status rests on the conversation workflow, a single agent node in conversation mode, and the governed customer refund approval template, itself a starting point. The conversation workflow has no knowledge sources and no handover, and the refund template does not answer customers. You build the join.
The library has a template, chatflow, MCP server or governed template that covers part of the job. You assemble the rest in Studio.
What this template does
Front-line customer service is mostly the same questions asked in different words: where is my order, how do I reset this, what does your policy say. The owner is the head of customer service, who needs fast answers that stay within policy. By hand, the answers drift between colleagues, and the awkward requests sit behind the easy ones in the same queue.
The template separates the two kinds of message. The agent answers what your approved content covers, and says plainly when it does not know. Anything that commits money or breaks a rule goes to a person with the conversation summarised, so the person starts from the facts rather than asking the customer to repeat them.
The workflow, step by step
Steps marked as approval gates pause the run until a named person approves. Nothing after a gate runs before that decision, and the decision is recorded.
- 01
Customer message arrives
A message from the chat widget or a connected channel starts the run. The flow keeps the message and the thread so far. If the customer is signed in, the account reference comes with it. - 02
Look up the customer
The flow reads the account or order the customer refers to in your CRM or order system. It reads only. What it finds is added to the agent's context, so the answer can be specific rather than generic. - 03
Answer from approved content
The agent answers from the policy pages, help articles and canned answers you gave it, and names the source. When the content does not cover the question, it says so and offers a person. It does not guess at policy. - 04
Spot the request that needs a person
A second check looks for refund requests, cancellations, exceptions, complaints, and anything touching legal or safety matters. The agent stops answering that thread and makes no promise. It prepares a short summary of what the customer wants and what it has already said. - 05
Person decides
The case goes to a named person with the summary and the account facts. They approve, change or decline the refund or exception. The Studio approval step holds the run until they do. Assign different people for different value bands.Approval gate: a named person approves before the next step runs. - 06
Relay the decision
The agent tells the customer the outcome in your tone, using the person's decision and note. If the person declined, the message states the reason they gave. The agent does not add anything the person did not approve. - 07
Close and record
The thread closes with the transcript, the sources the agent used, whether it handed over, the person's decision and the outcome stored together. The lead can then read where answers failed and fix the content behind them.
What it can do, cannot do, needs approval for, and records
Can
- Answer questions from your approved content
- Look up the customer account or order
- Recognise a request that needs a person
- Summarise the conversation for the person
Cannot
- Promise a refund, credit or exception
- Answer from outside the content you supplied
- Change account details on its own
- Give legal, medical or financial advice
Requires approval
- Every refund and exception
- Account changes the customer asks for
- Complaints that name a regulator or a lawyer
Records
- The transcript
- Sources used for each answer
- The handover summary
- Approver, decision and time
- The outcome sent to the customer
What the library holds for this job
| Asset | Type | Covers | Role in this flow |
|---|---|---|---|
| Conversation | Workflow template | Part of the job | One agent node in conversation mode between Start and End; it gives you the answering step but carries a fixed test message and no knowledge sources. |
| Customer refund approval | Governed template | Part of the job | The pattern for a refund that a person approves above a threshold, with a record for each one. |
| Customer Service Response Template | Prompt | Used inside the flow | A starting prompt for tone and structure of replies. |
| Zendesk | Integration | Used inside the flow | Where a handed-over case lands for the person. |
Integrations this flow uses
- Zendesk: Receives the handed-over case and holds the history.
- Slack: Alerts the person on duty that a case is waiting.
- Twilio: A message channel for customers who reach you by text.
- HubSpot: Customer records the agent reads for context.
The full list of tools Studio connects to is on the integrations page.
What you supply, and what this page does not cover
- You supply the approved content: policy text, answers and tone. The agent is only as accurate as that material.
- You decide what counts as an exception and who approves each value band.
- The shipped conversation workflow is a test with a fixed message, so retain-history and knowledge settings are yours to turn on.
- Languages and channels beyond text chat are not covered on this page.
- Voice calls are not covered. The call centre page covers that.
Common questions
- What stops the agent from inventing a policy?
- Two things you control: the content it may answer from, and an instruction to say it does not know when that content is silent. Neither is absolute, so read early transcripts and add content where it fell short.
- How does a handover reach a person?
- The flow creates a case with a summary in your helpdesk and alerts the person on duty, for example in Slack. The approval step then holds the run until they decide. The customer is told that a person is looking.
- Is this the same as the support queue template?
- No. This one answers customers live in a conversation. The customer support template works a ticket queue, where an agent drafts replies for a person to send. Many teams use both, one for chat and one for email.
- Can it work in several languages?
- Models reached through Connect can read and write many languages, but quality varies by model and language. Test the languages you serve with real questions before launch, and keep the approved content in each one.