AI agents for customer service: the skills, the job and the agent
Most support teams use AI one ticket at a time: paste the question, copy the answer, check it against the help centre. Everyone does it their own way, so customers get different answers to the same question.
A skill describes one support task in detail, the way your most experienced colleague does it. Several skills in order make a job, such as clearing the morning queue. The job running on its own, stopping for a person before anything reaches a customer, is an agent. Here is that ladder for customer service.
- 01
6 skills
One task each, described in detail.
- 02
Connectors
Reach into the systems the tasks need.
- 03
One job
The morning queue
- 04
One agent
The morning queue agent
Claude skills for customer service: the 6 to write first
ticket-triageTicket triage
Use when: new tickets have come in and need a category, a priority and an owner.
Reads the ticket, picks a category from the team’s list, sets the priority by your SLA rules and suggests who should take it.
Output: Category, priority and owner for each ticket, with one line on why.
ticket-reply-draftReply draft from the knowledge base
Use when: a ticket or chat needs an answer.
Finds the answer in the knowledge base, writes the reply in your tone with a link to the article, and says so when the knowledge base does not cover the question.
Output: A reply draft with its source, or “Not in the knowledge base” and who usually knows.
escalation-summaryEscalation summary
Use when: a ticket moves to second line, another team or a colleague.
Sums up the customer, the problem, what has been tried, what was promised and the open question, so the next person does not have to read the whole thread.
Output: A five-line internal note.
complaint-replyComplaint reply
Use when: a customer is angry, asks for compensation or threatens to cancel.
Acknowledges the problem, explains what happened from the ticket history, offers only what your compensation policy allows and flags anything beyond it to the team lead.
Output: A reply draft and, when needed, a note to the team lead.
help-articleHelp article from solved tickets
Use when: the same question has come in several times and has no good article.
Collects the solved tickets and writes a help centre article: the question, a short answer, the steps and the screenshots to add.
Output: An article draft ready to review and publish.
contact-reasonsWeekly contact reasons
Use when: someone asks why customers get in touch or what drove the volume last week.
Groups the week’s tickets by reason, counts them, quotes one typical ticket per reason and suggests one fix for each of the top five: an article, a macro or a bug for the product team.
Output: One page with the top five reasons, their counts and one fix each.
What a good skill looks like
An example written for this page, following Anthropic’s guidance for writing skills. Adapt it to how your team works.
---
name: ticket-reply-draft
description: Drafts the reply to a customer ticket from the knowledge base, with its source. Use when a ticket or chat needs an answer. Not for complaints or refunds, which go to complaint-reply.
---
# Reply draft for a customer ticket
## What you deliver
Two parts, in this order:
1. The reply, ready to paste: greeting, answer, steps if needed, link to the help article, sign-off. At most 120 words.
2. One internal line for the colleague on the ticket: the sources you used and anything to check before sending.
## Steps
1. Read the whole ticket and the customer's two latest tickets. Find the actual question. If there are two, answer both.
2. Search the knowledge base. Use only articles updated in the last 12 months.
3. Give the answer first and the explanation after. Steps as a numbered list.
4. Add the link to the article the answer comes from.
5. Use the customer's name and the language they wrote in.
6. Read the reply as the customer: can they solve it without writing back?
## Rules
- Answer only from the knowledge base and the ticket. If the answer is not there, write no reply. Write "Not in the knowledge base" and who in the team usually knows. Never guess.
- Never promise a refund, a credit, a delivery date or a bug fix. Hand those to complaint-reply or the team lead.
- No "We apologise for any inconvenience". Say what happened and what happens next.
- If the customer is angry or mentions cancelling, say so on the first line of the internal part.
## Example
Hi Anna,
You add a new user under Settings > Team > Invite. They get an email with a link that works for 7 days.
1. Go to Settings > Team.
2. Click Invite and enter their email address.
3. Choose a role: Admin or Member.
Full guide: help.acme.example/invite-users
Best regards, Jonas at Acme AB support
Internal: source "Invite users" (updated in March). The customer is on Basic, max 5 users. Check before sending.
- The description also says when not
- A reply draft and a complaint reply look alike to Claude. Sending complaints to the other skill keeps angry customers out of the friendly template.
- A source, or no answer
- A confident wrong answer costs more than a slow one. The skill writes “Not in the knowledge base” instead of filling the gap, and the gap tells you which article to write next.
- One example with real steps
- The example shows the length, the numbered steps and the link. Every draft looks like it, whoever is on shift.
From skills to a job: The morning queue
Today the first hour of the day goes on reading the night’s tickets, sorting them and looking up answers. With the skills and the right connectors, one person has the queue sorted and the easy tickets drafted in twenty minutes, in Claude or ChatGPT.
- 01Ask the AI for the tickets that came in to Zendesk since yesterday afternoon.
- 02Run ticket-triage on them for category, priority and owner.
- 03Run ticket-reply-draft on the ones the knowledge base can answer.
- 04Read each draft and add what only a person knows, such as a promise made on the phone.
- 05Send the replies from Zendesk, urgent tickets first, and hand complaints to the team lead.
From job to agent: The morning queue agent
Every morning at 07:00, an hour before the team logs in.
- 01Finds the new tickets that came in since yesterday afternoon. via Zendesk MCP
- 02Reads each ticket with its comments and the customer’s earlier tickets. via Zendesk MCP
- 03Searches the knowledge base for the answer to each question. via Confluence MCP
- 04Runs ticket-triage, then ticket-reply-draft, on each ticket.
- 05Adds the suggested priority and the reply draft to each ticket as an internal note. The customer sees nothing. via Zendesk MCP
- 06Posts the queue in the support channel: how many tickets, which are urgent, which the knowledge base could not answer. via Slack MCP
Whoever picks up a ticket reads the draft in the internal note, fixes it and sends the reply from Zendesk. The team lead takes the urgent tickets and the complaints. Nothing reaches a customer before a person has read it.
At 08:00 the night’s tickets have a suggested priority, the easy ones have a reply draft waiting, and the urgent ones are listed in the support channel.
In Claude, make it a scheduled task: describe the job once, choose daily at 07:00, and it runs with your skills and connectors even when your laptop is closed. It works through the queue at a set time, not the second a ticket arrives, so it suits the night’s backlog rather than live chat. Scheduled tasks are on every paid Claude plan. In ChatGPT Work, a scheduled task does the same on the same daily cadence.
With Deployed OS, the skills and connectors belong to the company, not to one person on the team. Everyone gets the same reply skill in Claude, ChatGPT and Copilot, a fix to the skill reaches the whole team the same day, each person connects Zendesk with their own login, and every call the agent makes is logged.
Where it goes wrong
- An out-of-date knowledge base
- The agent answers from what is written. If the help articles are two versions old, so are the drafts. Update the five most used articles before you build anything.
- Angry customers in the standard template
- Complaints, refunds and cancellations need a person from the first line. Send them to complaint-reply and the team lead, not to the everyday reply draft.
- Drafts without a source
- If a draft does not show where the answer came from, the team has to check everything or nothing. Make the skill name the article every time.
- Drafts sent unread
- If the team is measured on handling time alone, drafts go out unread. Look at what people change in the drafts instead: every fix is a rule the skill is missing.
AI agents for customer service: common questions
An AI agent for customer service takes the recurring work around every ticket: sorting the queue, drafting replies from the knowledge base, summing up a ticket before it is escalated, and turning repeated questions into help articles. It works behind the team, not in front of customers, and stops for a person before anything is sent.
Ticket triage and the reply draft. They run on every ticket, so within a week you see what the skills get wrong. Then the escalation summary, then the help article.
Not through these connectors. In Zendesk it adds internal notes, and in Intercom it adds notes only teammates see. It does not send public replies, close tickets or change status, priority or assignee. In Gmail and Outlook it only creates drafts. We recommend a person reads every reply before it goes out. With Deployed OS, admins decide which groups may let the AI write at all.
Both. Skills are written in plain language, and the same connectors work in Claude, ChatGPT and Copilot through Deployed OS. Pick the assistant the team already uses.
Where it already is. The agent can search and read Confluence, Notion, Google Drive and SharePoint. What matters more is that each article is current and answers one question. In Notion the agent searches page titles, so a title that states the question makes a difference.
It depends on the size of your team. Get in touch and we will walk you through it.
Start with one job.
Tell us which task eats your team’s week. We’ll show you the skills that do it, and the agent that runs it.
