AI agents for customer support · Part 1 of 14

How to set up an AI support agent

What it is, what it does with a ticket, and eight steps to an agent your team trusts with real customers.

Aakash William
· 4 min read

The short version

An AI support agent reads your customers’ messages, answers from your own knowledge, and acts in your systems. Eight steps set one up:

  1. Choose the first conversations
  2. Get your knowledge in order
  3. Connect the systems it needs
  4. Decide what it may do
  5. Write the rules for its actions
  6. Plan the hand-over
  7. Test it on real tickets
  8. Roll it out in stages

What an AI support agent is

It handles a customer conversation from start to finish. It reads the message, finds the order, looks up your knowledge, then answers, acts, or hands the customer to a person.

ReadsFindsLooks up DecidesActsReplies the messagethe orderyour knowledge what to dothrough a tooland records check HelpdeskStore, CRMKnowledge ToolsRecord Your rules messageorderanswer changeswrites limits

1 Reads

the message

Helpdesk
↓

2 Finds

the order

Store, CRM
↓

3 Looks up

your knowledge

Knowledge
↓

4 Decides

what to do

↓

Check

against your rules, before it acts

Your rules
↓

5 Acts

through a tool

Tools
↓

6 Replies

and records what it did

Record
Every ticket takes the same path. Setting up an agent means making each step reliable, with a check on your rules between the decision and the action.

What sets it apart from the tools most teams tried before is who acts.

Scripted chatbot AI assistant AI support agent for your team CustomerScriptReply CustomerAIPersonReply CustomerAI agentReply Your systems writesmatches a buttona canned answer writesdrafts a replychecks and sendsto the customer writesreads, looks upto the customer acts within limits off script: stuck

Scripted chatbot

Customer → script → a canned answer

Off script, it gets stuck.

AI assistant for your team

Customer → AI drafts a reply → a person checks and sends

AI support agent

Customer → AI agent reads and looks up → acts in your systems, within limits

A scripted bot can only pick a button. An assistant drafts while a person sends. An agent acts in your systems, within the limits you set.

Step 1Choose the first conversations

Tag a month of tickets by reason. Start with the reasons that come up often, have a written answer, and are cheap to get wrong.

How often it comes up rarely often How cheap a mistake is costly cheap Start here Where is my order? Return or exchange Small refund Change an address Cancel a subscription Complaint needing judgement Dispute or chargeback

Start here: often, and cheap to get wrong

Where is my order?Return or exchangeSmall refund

Next: rarer, or costlier

Change an addressCancel a subscription

Leave with your team: rare and costly

Complaint needing judgementDispute or chargeback
Place your own ticket reasons on these two axes and start in the top right. The positions here are an example.

Step 2Get your knowledge in order

Write down what your team knows but never wrote. Remove contradictions, and give each answer an owner. Keep rules out of it: they belong in step 5.

“Can I get a refundfor this?” AI agentKnowledgeReplyRefund tool how returns workexplains the policygives the refund asks looks up explains wants to refund checked against your rules first

Knowledge shapes the reply

“Can I get a refund?” → the agent looks up how returns work → it explains the policy

Rules decide the action

The agent wants to refund → checked against your rules first → the refund tool acts

Knowledge shapes what the agent says. A rule decides what it may do, and is checked before the tool acts.

An agent missing an answer may invent one. In April 2025 the support bot of Cursor, a code editor, told customers about a one-device policy that didn’t exist.

Cursor (company), Wikipedia; eWeek

Step 3Connect the systems it needs

Give it access to read every system behind your chosen reasons. Access to change things comes only with step 4.

AI agentHelpdeskCRM Store and ordersPaymentsShipping also hands overalso refunds, editsalso refunds reads onlyreads only conversationcustomer, plan order, statuspaid, refundedtracking

Helpdesk

reads the conversation

also hands over

Store and orders

reads the order and its status

also refunds, edits

Payments

reads what was paid and refunded

also refunds

CRM

reads the customer and their plan

reads only

Shipping

reads tracking

reads only
Every line is something the agent reads. The boxes in colour are where it may also change something, and each of those is a decision for step 4.

Step 4Decide what it may do

Sort every action its tools could take into three groups.

AloneAlways a person A person approves first Order statusAddress change before dispatchRefund under its limitCredit within its limit Refund above its limitDiscount above a set % Address change after dispatchDispute or chargebackA case no rule covers

Alone

Order statusAddress change before dispatchRefund under its limitCredit within its limit

A person approves first

Refund above its limitDiscount above a set %

Always a person

Address change after dispatchDispute or chargebackA case no rule covers
An example to start from, not a standard. Your groups depend on your margins, your customers and what your team does today.

Step 5Write the rules for its actions

A rule says how far the agent may go alone, and who decides above that. Count limits per customer across every chat and every agent, and keep rules where the tool checks them before it acts, not in the prompt.

AI agentPerson on the team alone hands it to the team leadaloneteam lead approves first head ofsupport €0€30€50€250 more

AI agent

up to €30
alone
above €30
hands it to the team lead

Person on the team

up to €50
alone
€50 to €250
team lead approves first
above €250
head of support approves first
An example refund policy as a ladder. The agent’s own limit sits below the team’s: above €30 it hands the case to the team lead, whatever the amount.

What your agent says counts as your company’s word. In Moffatt v. Air Canada (2024), a tribunal held the airline to its chatbot’s answer about bereavement fares, and called the claim that it couldn’t be liable for its chatbot “remarkable”.

Moffatt v. Air Canada, 2024 BCCRT 149

Step 6Plan the hand-over

Decide when the agent hands over, to whom, and with what, so the customer never repeats themselves.

AI agent hands over Hand-over to the team lead The orderAsked forCheckedAlready didWhy delivered 3 days ago, €80a refund for a damaged itemphoto of the damage receivednothing refunded yet above the agent’s €30 limit

Hand-over

to the team lead
The order
delivered 3 days ago, €80
Asked for
a refund for a damaged item
Checked
photo of the damage received
Already did
nothing refunded yet
Why
above the agent’s €30 limit
What the person sees the moment they take over. The customer never has to explain it again.

Step 7Test it on real tickets

Run 100 to 200 past conversations through the agent, with the right outcome written down for each. Then test every limit.

the limit: up to €30 €29.99€30.00€30.01 refundsrefunds hands it over

The limit: up to €30

€29.99

refunds

€30.00

refunds

€30.01

hands it over

Test every limit at the limit itself, and one cent past it.

Step 8Roll it out in stages

Give the agent more only when the record shows its past cases were right. If your team approves nearly every request it sends, the limit is too low.

1 Suggest2 Small autonomy3 Widen The agent drafts;a person sends. The agent sends thesimplest kinds alone. More kinds, higherlimits, with evidence. the agent sends nonethe agent sends somethe agent sends most

1 Suggest

The agent drafts; a person sends.

↓

2 Small autonomy

The agent sends the simplest kinds alone.

↓

3 Widen

More kinds, higher limits, with evidence.

Each square is a kind of ticket; the coloured ones the agent sends on its own. Move on only when the last stage’s record supports it.

Anthropic found that users of its coding agent approved 93% of the permission prompts it showed them, and called it “approval fatigue”: people no longer pay close attention to what they approve.

Anthropic Engineering, “Claude Code auto mode”, March 2026

How to measure it

Track the resolution rate, the reasons for hand-overs, CSAT on the agent’s conversations, reopened tickets, and money given against your rules. And measure resolution, not deflection.

Conversations the agent took Resolved by the agent Customer gave upHanded to a person Deflection Resolution
The conversationDeflectionResolution
Resolved by the agentcountscounts
Customer gave upcountsno
Handed to a personnono
Deflection counts the customer who gave up as a success. Resolution counts only the conversations the agent actually settled.

Next in this series

Part 2: what an AI support agent can and can’t do today. Part 3: build or buy. See all 14 parts.

Where Himmel fits

Himmel is built for step 5. Your team confirms the rules in plain words, and before each action that moves money the agent’s tool checks them with Himmel, which keeps a record of every answer.

Sources

  1. Cursor’s invented device policy, April 2025: Cursor (company), Wikipedia; eWeek.
  2. Moffatt v. Air Canada, 2024 BCCRT 149, decided 14 February 2024.
  3. Anthropic Engineering, “Claude Code auto mode”, March 2026.