How to Automate Customer Support Step by Step

To automate customer support, start with one pilot queue, give an AI agent approved answers for it, connect a handoff to your team, test the edge cases, launch on a few pages, and measure what changes before you expand. This tutorial walks through each of those steps in Chatsy, with the exact settings and plan limits as of November 12, 2026. The screenshots come from a local test workspace with sample data. That workspace had no AI model key or email provider configured, so they show setup screens and the human handoff rather than AI answers or sent emails.

This page is about implementation. If you are still deciding what to automate, how to audit your queue or how to build the business case, read the customer support automation strategy guide first, then come back here to set it up.

Before you start

This tutorial assumes you have:

A list of the 10 to 20 questions your team answers most, with real examples. The strategy guide's queue audit explains how to build it.
Approved, current answers for those questions: help articles, policy pages or internal notes you are happy to show customers.
One person who owns the pilot and can change content during the first weeks.
A clear list of requests the AI must never handle alone, such as refunds, account changes, legal or safety issues.

You do not need code. You need to paste one script tag into your website.

Step 1: Pick one pilot queue

Choose one group of questions that has written answers and low risk if the answer is incomplete. Good first pilots:

Product and policy questions answered in your help center
Shipping times, return windows, opening hours
"How do I" setup questions with documented steps

Poor first pilots:

Anything that needs account or order data (order status, billing changes)
Complaints and disputes
Questions where the answer depends on judgment

Write the pilot scope in one sentence, for example: "The agent answers return-policy and shipping-time questions on our help pages and hands everything else to the team."

Step 2: Choose a plan that fits the pilot

Chatsy charges per AI reply, in credits. Each reply uses 1 to 30 credits depending on the model you pick. Seats are capped per plan and include the owner.

Plan	Price	Monthly credits	Seats	Live chat, mailbox, help center
Free	$0	40	1	No
Hobby	$40/mo ($400/yr)	500	1	Yes
Standard	$150/mo ($1,500/yr)	4,000	3	Yes
Pro	$500/mo ($5,000/yr)	15,000	5	Yes
Enterprise	Custom	100,000	10	Yes

Estimate credits with this formula: expected chats per month x AI replies per chat x model credits per reply.

Chatsy blocks a reply that would make usage reach the monthly allowance. There are no top-ups, overages or rollover. So Hobby covers at most 499 replies on a 1-credit model, or 49 on a 10-credit model. Pro covers at most 14,999 one-credit replies or 499 thirty-credit replies.

Free is enough to build and test the agent, but it has no live chat or mailbox. For a pilot with real handoff, start on Hobby. If more than one person will answer escalations, you need Standard (3 seats) or Pro (5 seats). See pricing and message credits.

Step 3: Create the agent and load approved answers

Create an account and follow the onboarding wizard: Start with your website, Name your agent, Set its tone of voice, Style the widget, then Ask it a real question. The first agent tutorial covers each screen.

Then open the agent's Knowledge page and use Add knowledge to load only the content in your pilot scope:

Add website, then One page or List of links (up to 500 URLs) for specific help pages. Whole site crawling needs Hobby or higher.
File uploads (TXT, HTML, PDF, DOC, DOCX, MD or CSV, 25 MB each) for policy documents.
Write an answer for short facts that must be exact, such as "Returns are accepted within 30 days of delivery."
Import help articles if you publish a Chatsy help center (Hobby or higher).

Leave out drafts, old policies and internal notes you would not show a customer. The agent can only be as accurate as its sources. Chatsy does not re-read websites on a schedule, so recrawl a source whenever the page changes.

Step 4: Write the agent's rules

Open the agent's Settings, then the AI tab.

Model. Start with a 1-credit model such as GPT-6 Luna or Claude Haiku 5.5. Move to a 2-credit model such as Claude Sonnet 5.5 or GPT-6.1 Sol only if testing shows weak answers. Avoid models their providers have deprecated, such as GPT-5 and Claude Sonnet 4.5, for a new pilot.
Response instructions. State the pilot scope, what to refuse and when to hand off. For example:

Answer only questions about returns and shipping times, using the
supplied sources. If the sources do not contain the answer, say so
and offer to connect the customer with the team. Never approve
refunds, promise delivery dates, or change orders. Hand off at once
if the customer asks for a person, is upset, or mentions a charge.

Response controls. Keep Response variation low so repeated questions get consistent answers.
Knowledge retrieval. Leave Use knowledge sources on.

Chatsy does not score sentiment or apply a confidence threshold. The agent escalates when your instructions tell it to. Separately, the widget offers to talk to the team or leave an email when a visitor asks for a person, an answer has no supporting source or says it is unsure, the visitor marks an answer unhelpful, or a question repeats. The wording of your rules still decides most handoffs.

Step 5: Turn on handoff, tickets and email follow-up

Open workspace Settings, then Modules (Hobby and higher):

Live chat adds an Escalate to Human tool and a Talk to a person option in the widget. Escalated chats wait on the Live chat page, where a teammate can claim, release, resolve or return them to the agent. See human handoff.
Inbox and tickets lets you forward a support address into Chatsy. Verify the address, forward it and send a test email. The agent gets a Create Support Ticket tool. In the agent's General settings, choose the Primary mailbox. See connect your support email.

In the agent's General settings, set When to ask for email so visitors can leave an address when nobody is available. Chatsy then emails the workspace owner to follow up, when email delivery is configured.

Plan around two limits. There is no staff online or away status, so escalations reach the queue whether or not anyone is watching. And Chatsy does not email staff when a chat escalates; alerts can go to a connected Slack workspace. Agree who checks the queue and when.

Do not promise actions in other systems. Chatsy's provider connections, such as Stripe, Slack and Pipedrive, can be set up in the dashboard, but agents cannot run actions through them yet. Zendesk cannot be connected today. A custom API tool is possible for your own endpoints; test it end to end before customers can trigger it.

Step 6: Test before customers see it

Open the private playground from the agent overview (Test the answers, then Open playground). Record every answer in a sheet as correct, wrong or declined.

The test set

Run at least 30 questions:

15 in-scope questions phrased the way customers write them, including typos
5 in-scope questions with tricky details (exceptions, dates, regional rules)
5 out-of-scope questions your sources do not answer. The agent should say it does not know and offer the team.
5 requests that must go to a person: "I was charged twice", "cancel my order", "I want to talk to someone"

Fix the source or the instructions for every wrong answer, then rerun the whole set. Do not launch until the out-of-scope and must-escalate questions behave correctly.

Handoff tests

Turn on Allow public chats under Settings, Sharing, and install the widget on a test page. Then, from a normal browser window:

Ask for a person and confirm the chat appears on the Live chat page.
Claim it, reply, and return it to the agent.
Ask a question that should create a ticket and confirm it lands in the right mailbox.
Leave an email with nobody watching and confirm the follow-up email arrives.

Step 7: Launch to a small slice of traffic

Chatsy has no traffic-percentage switch, so control the rollout by where you paste the script. Copy it from the agent's Deploy panel or Settings, Sharing, Link and embed, and paste it before </body>.

Week 1: help center or FAQ pages only. Read every conversation.
Week 2: add the contact page if week 1 had no wrong policy answers.
Week 3 and later: add more pages, one group at a time, only when the previous group is clean.

The widget install guide has platform-specific steps. To take the agent offline fast, turn off Allow public chats; the script stops loading.

Step 8: Measure what changed

Write down your baseline before launch: weekly ticket or email volume for the pilot topics, and your current first reply time. Then use what Chatsy reports:

Agent Analytics (7, 30 or 90 days): conversations, messages, average first reply time, top questions and knowledge gaps labeled "Rated not helpful", "No answer found" and "Handed off".
Agent overview: helpful-rating share (thumbs up or down) and handoff rate.
Knowledge page: "Questions it struggled with".

Chatsy does not report CSAT, containment or resolution rate. To see whether the pilot reduced work, compare weekly ticket volume for the pilot topics in your inbox before and after launch, and sample conversations to check that customers did not simply give up.

Step 9: Review weekly and expand

Spend 30 to 60 minutes a week on this loop:

Read the knowledge gaps and every handed-off conversation.
Fix or add the source for each gap. Recrawl changed pages.
Rerun your test set in the playground.
Note wrong answers that reached customers, and why.

Expand to the next question group only when the current one has had two weeks without wrong policy answers. Each expansion repeats Steps 3 to 7 for the new scope.

A worked example

A store gets about 1,200 support emails a month. It pilots the agent on its help pages and expects about 400 chats a month, averaging 3 AI replies each.

AI replies: 400 x 3 = 1,200 a month
On a 1-credit model: 1,200 credits. Hobby (500) is too small. Standard (4,000) fits with room to spare.
On a 2-credit model such as Claude Sonnet 5.5: 2,400 credits. Still within Standard, which allows up to 1,999 two-credit replies.
Two people will answer escalations, so the store needs at least 2 seats. Hobby has 1; Standard has 3.
Plan: Standard at $150 a month, or $1,500 a year.

This tells you what the pilot costs. Whether it saves money depends on how many emails it removes and how much time your team spends on handoffs, which you will only know after Step 8. Do not budget savings before you have measured them. Use the support cost calculator to model your own numbers.

Common implementation mistakes

Loading every document at once. The agent then answers outside the pilot scope from old or internal content. Load only what the pilot covers.

Skipping the out-of-scope tests. An agent that answers everything confidently is the most dangerous kind. Test what it should refuse.

No one watching the queue. Without presence status or email alerts, an unwatched queue means customers wait. Assign hours, or rely on email follow-up.

Underestimating credits. Test chats count. A larger model on a small plan can run out mid-month, and replies stop until the reset or an upgrade.

Promising actions the agent cannot take. If the agent cannot look up orders or issue refunds, its instructions must say so and hand off.

When this setup is the wrong fit

You have fewer than about 50 support requests a month. Better help articles and a contact form may be enough.
Your support is spread across several inboxes and chat tools. Consolidate into one inbox first, or results will be hard to measure.
Most requests need live account or order data and you have no tested API for it.
You need SOC 2, a signed HIPAA BAA or SSO. Chatsy does not offer them; see the security overview.
You need the widget interface in languages other than English and Spanish.

Frequently Asked Questions

How long does it take to automate customer support with Chatsy?

The setup steps take a day or two if your help content exists. The pilot, testing and first weekly reviews take a few weeks. Writing missing content is usually the longest part.

What should I automate first?

Questions with written, stable answers and low risk: policies, shipping times, product details and documented how-to steps. Keep refunds, disputes and account changes with people. The strategy guide covers how to rank candidates.

Which Chatsy plan do I need?

Free is enough to build and test. A pilot with live chat or a mailbox needs Hobby or higher. If more than one person answers escalations, you need Standard or Pro for the extra seats. Check the credit formula in Step 2 against your expected volume.

How do I know if automation is working?

Compare weekly ticket volume for the pilot topics before and after launch, track the handoff rate and knowledge gaps in Chatsy Analytics, and sample conversations each week for wrong answers. Chatsy does not report CSAT or resolution rate.

Can the agent take actions like refunds or order changes?