Use cases
How to use an AI agent for customer support
An AI agent for customer support reads the customer’s question, looks up what it needs in your systems, such as the order in Shopify or earlier tickets in Zendesk, and answers or drafts the reply. You choose whether it answers on its own or leaves a draft for a person, and every reply it sends can wait for approval.
Updated · 4 min read
What can an AI agent do in customer support?
- Answer questions in the chat on your website, day and night.
- Look up an order, a customer or an earlier ticket before it answers.
- Read a new ticket, find similar solved ones, and draft the reply.
- Leave the draft as an internal note, so an agent on your team sends it with one click.
- Tag, prioritize and route tickets, or open a ticket when a chat needs a person.
It works best on the questions your team answers again and again: where is my order, how do I change my plan, what is your returns policy. Unusual cases still go to people.
Where does the support agent work: in the chat or in the help desk?
Most teams use one or both of two setups.
| Website chat | Help-desk assistant | |
|---|---|---|
| Who talks to the agent | The customer, live | Nobody: it works on each new ticket |
| How it starts | A visitor sends a message | The help desk calls a workflow’s webhook |
| What it produces | An answer in the chat | A draft reply, a note, tags |
| Who sends the reply | The agent, or the agent after approval | Your team, or the agent after approval |
| Best first step | Answer from public information only | Internal notes only |
Who talks to the agent
- Website chat
- The customer, live
- Help-desk assistant
- Nobody: it works on each new ticket
How it starts
- Website chat
- A visitor sends a message
- Help-desk assistant
- The help desk calls a workflow’s webhook
What it produces
- Website chat
- An answer in the chat
- Help-desk assistant
- A draft reply, a note, tags
Who sends the reply
- Website chat
- The agent, or the agent after approval
- Help-desk assistant
- Your team, or the agent after approval
Best first step
- Website chat
- Answer from public information only
- Help-desk assistant
- Internal notes only
The website chat is covered in how to add an AI agent to your website. The rest of this article sets up the help-desk assistant.
How do you set up a help-desk assistant, step by step?
In MyDeepAgents, with Zendesk as the example:
- Connect Zendesk in the console with your Zendesk subdomain and an API token. Connect Shopify too if order questions are common.
- Create an agent and give it Get a ticket, Search tickets and Comment on a ticket, plus Get an order if you connected Shopify.
- On Comment on a ticket, preset
publicto false. A preset always wins over the model’s choice and the model never sees it, so every comment the agent writes is an internal note. - Write the instructions: your tone, your policies, what to look up first, and which questions to leave alone.
- Create a workflow with a webhook trigger and one agent step that passes it the ticket number from the request. Verify the requests with a shared token or an HMAC signature.
- In Zendesk, add a trigger that calls the workflow’s webhook when a ticket is created.
- Read the first drafts. When they are good, let the agent tag tickets, and later reply in public with approval on.
The same pattern works with Freshdesk (Reply to a ticket, Add a note with private preset to true) and Intercom (Reply to a conversation). Browse the help desk tools for others.
How does the agent know your policies and products?
Three ways work today. Put the policies that matter in the agent’s instructions: refund rules, delivery times, tone. Give it tools that read live data, so it looks up the order instead of guessing. And let it search earlier solved tickets, which hold the answers your team already gave.
How do people stay in charge of support replies?
- Drafts first. With public comments preset off, nothing reaches the customer until a person sends it.
- Approvals. When you let the agent reply, each reply can wait for a person’s yes. In the website chat the customer sees that the team is checking the request.
- Narrow tools. An agent that can read orders cannot refund them unless you add a refund action.
- A way out. Tell the agent to open a ticket for a person when a chat is beyond it. It cannot hand over a live chat yet.
- A record. Every lookup and every reply is recorded with its input and output.
Key points
- Support agents answer in the website chat, or draft replies to help-desk tickets.
- Start with internal notes only: preset the comment to private so nothing reaches customers unreviewed.
- A help-desk trigger calls a workflow’s webhook, and the workflow runs the agent on each new ticket.
- Live order and ticket lookups beat guessing; document search is not available yet.
- Replies can wait for a person’s approval, and every call is recorded.
Questions people also ask
Can an AI agent answer support tickets on its own?
Yes, if you give it an action that replies and switch off approval for it. Most teams start with drafts as internal notes and let the agent reply alone only for question types it handles well.
Which help desks can a MyDeepAgents agent work with?
Zendesk, Freshdesk and Intercom among others; the help desk category has more than 15 connectors.
Can the agent look up a customer’s order?
Yes. With Shopify connected it can get an order and find customers by email, and it reads them live each time.
What happens when the agent cannot answer?
It follows its instructions: usually it says so and opens a ticket for your team. It cannot transfer a live chat to a person yet.