Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
Using AI SNS CHAT

Getting Started with AI SNS CHAT: From Sign-Up to Going Live

AI SNS CHAT EditorialPublished 7 min read

Set up AI SNS CHAT step by step: create an account, build an agent, add knowledge, test it in Playground and publish on web chat or WhatsApp.

Most teams do not struggle to sign up for an AI support tool. They struggle with what comes next: what to teach the agent, how to know it is ready, and where to put it first. This guide walks through the whole path, from creating an account to publishing your first agent on a channel such as website chat, WhatsApp or Messenger, and shows how to keep the first launch small and safe.

Decide the scope before you click anything

A first agent works best when it has one job. "Answer everything for everyone" is how a pilot turns into a pile of vague replies. Pick a narrow scope you can describe in a sentence, for example: "answer shipping, returns and opening-hours questions for our online shop, and pass anything about an existing order to a person."

Write down three things before you start:

  • The audience. Website visitors, existing customers on WhatsApp, or people who message your Facebook page.
  • The top questions. Pull 15 to 30 real questions from your inbox or chat history. These become your test set later.
  • The boundaries. What the agent must never do: quote prices you have not published, promise refund outcomes, give legal or medical advice.

Create the account and set up your workspace

Account creation is the quick part. Use an email address that will outlast any one employee, such as a shared support mailbox, because the first account becomes the owner. Turn on two-factor authentication straight away; an owner account that can edit every agent deserves more than a password.

Then think about structure. If you run more than one brand, region or client, separate them into workspaces from day one. Moving things later is possible, but a tidy structure at the start saves you from untangling permissions when the team grows. Decide who needs what access: people who edit the agent, people who only read conversations, and people who only answer escalations are three different roles.

Build the agent: persona first, knowledge second

The persona is the agent's job description. It covers tone, language, how it introduces itself, what it should refuse, and when it should hand over to a human. A good persona is short and specific:

  1. Role: "You are the support assistant for a small online homeware shop."
  2. Tone: friendly, concise, no jargon, no slang.
  3. Rules: answer only from the knowledge you have been given; if unsure, say so and offer a human.
  4. Escalation: pass over complaints, anything involving payments, and anyone who asks for a person.

Starting from a ready-made persona and editing it is usually faster than writing from scratch, as long as you replace every generic line with something true about your business. The rule "say so when you do not know" matters more than any other. An agent that admits a gap is far less risky than one that improvises.

Give it knowledge it can actually use

The quality of answers is capped by the quality of the sources. A few principles keep the first version strong:

  • Prefer fewer, cleaner sources. One accurate returns page beats five overlapping, outdated ones.
  • Write for questions, not for documents. If customers ask "Can I change my delivery address?", make sure a source answers that in plain words.
  • Use Q&A pairs for the awkward cases. For topics where the exact wording matters, such as opening hours over a holiday, a direct question-and-answer entry is more reliable than hoping the agent finds the right paragraph.
  • Remove what you would not say aloud. Internal notes, draft pricing and staff-only procedures should not be in a source a customer-facing agent can read.

Sources fall into a few families: uploaded files, pasted text, your website, hand-written Q&A, and documentation tools you already use. Pick whichever is closest to the truth you already maintain, so you are not creating a second copy that drifts out of date.

  1. 01Create account
  2. 02Build persona
  3. 03Add knowledge
  4. 04Test and publish
The path from sign-up to a live agentIllustrative example

Test hard before anyone sees it

Testing is where most launches are won. Use the 15 to 30 real questions you collected and run each one. Then run a second round of awkward ones: misspellings, two questions in one message, questions in another language, and requests the agent should decline.

Score each answer simply: correct, partly right, wrong, or should have escalated. For example, if you test 20 questions and 14 are correct, 3 are partly right, 2 are wrong and 1 should have escalated, you know exactly where to spend your next hour. This is a hypothetical split, shown to illustrate the method rather than a benchmark.

Correct
14
Partly right
3
Wrong
2
Should have escalated
1
Example: scoring a 20-question test round (hypothetical)Illustrative example

Fix causes, not symptoms. A wrong answer usually means a missing or conflicting source, not a broken agent. Add or correct the source, then re-run the same question. Repeat until the test set passes, and keep it, because you will want to re-run it after every significant change.

Choose where to publish, and plan the human handoff

For a first launch, publish on one channel and keep the audience small. Website chat is the lowest-risk start because you control the page and can place the widget only where it makes sense, such as a help or shipping page. WhatsApp and Messenger reach customers who already message you, but they come with the expectation of a quick, personal reply, so make sure the handoff works before you switch them on.

Handoff deserves as much thought as the answers:

QuestionWhy it matters
Who receives escalations?A handoff nobody sees is worse than no handoff.
When are they available?Customers should be told honestly what happens next, not given invented wait times.
How do they take over?The person should see the full conversation, so the customer never repeats themselves.

Also set a welcome message that tells people they are talking to an AI assistant and how to reach a human. Being upfront builds trust and reduces frustration when a question is out of scope.

The first week after launch

Treat week one as a supervised trial. Read conversations daily, not weekly. Look for three patterns: questions the agent could not answer, answers customers rated poorly, and chats that should have been escalated sooner. Each one points to a concrete fix: a new source, a corrected Q&A entry, or a tighter persona rule.

Keep a small change log: what you changed, why, and what you expect to improve. When something goes wrong, you will know which edit to suspect. Once the agent is stable, widen the audience, add a second channel, and only then consider extras like lead capture or outbound messages.

Doing this in AI SNS CHAT

Here is the same path mapped onto the product.

1. Sign up. Go to /signup and register with email, Google or phone SMS. The Free plan includes one agent, one seat, 30 MB of storage and sign-up credits, which is enough for a first test. Before you can create an agent, you are asked to sign a memorandum on handling Japan's My Number (特定個人情報) data. Compare plans at /pricing when you outgrow it. Turn on two-factor authentication (TOTP) in your account security settings.

2. Create the agent. Use the guided wizard. You can start from one of the 14 starter personas, then edit the persona and guardrails to match your business. Under Agent → Settings → AI you choose the model; advanced models and model fallback need Hobby or higher, and lower plans use the provider's standard model. If you experiment and regret it, restore points let you roll back, and Duplicate copies an agent's persona, model and training data (not its conversations) so you can try changes safely.

3. Add knowledge. Open Data sources in the sidebar and add Files, Text snippets, a Website, or Q&A entries. Notion is also available there, and connectors under Channels cover tools such as Confluence, GitBook, Google Drive and Zendesk Guide. Wiki connectors read up to 1.5M characters and 400 pages per sync and tell you if content was truncated.

4. Test in Playground. Run your question set in Playground. The A/B model compare view lets you put two models side by side on the same question. If you get stuck on a screen, ask Synthia, the built-in fleet manager: she can answer by voice or text and can navigate to the screen and highlight the control.

5. Publish. Open Channels in the sidebar and choose Chat widget, WhatsApp, Messenger or any other listed channel. Widget studio controls the widget's look, and the welcome message is auto-translated into every language the widget speaks. For your own address, custom domains (Agent → Settings → Custom domains) need Hobby or higher. Setup guidance for each channel is at /connections, and the full capability list is at /features.

6. Hand off and review. Helpdesk → Scheduling assigns handoffs to whoever is on shift and sends alerts to Slack, Teams, LINE WORKS, WhatsApp, email or a webhook; Helpdesk needs Standard or higher. Conversations from non-email channels appear in Activity → Chat logs, where you can take over a chat, edit a reply into the ideal answer to train the agent, and review thumbs up/down ratings. Track credits per agent on the Usage page; one reply uses one credit. Workspace settings → Activity records who created, edited or uploaded what.

For security details, see /security.

Frequently asked questions

Can I try AI SNS CHAT without paying?

Yes. The Free plan includes one agent, one seat, 30 MB of storage and sign-up credits. Some features, such as integrations, custom domains, API keys and advanced model choice, need Hobby or higher, and Helpdesk needs Standard or higher.

Which channel should I publish on first?

Use the one where your customers already contact you. Website chat is the safest default because you control where it appears. WhatsApp and Messenger work well once your handoff to a human is set up and tested.

How do I know the agent is ready to go live?

Run a real set of customer questions in Playground, including awkward ones, and fix the sources behind any wrong answer. When the set passes and a handoff test reaches the right person, start with a small audience and read the Chat logs daily during the first week.

What happens when the agent cannot answer?

It should say so and offer a person, which you reinforce in the persona. With Helpdesk scheduling, the handoff goes to whoever is on shift, and the customer message reflects the rota rather than an invented wait time.

Can I undo a change that made the agent worse?

Yes. Restore points let you roll an agent back, and you can duplicate an agent to test changes on a copy before touching the live one.

#getting started#setup#AI agent#website chat#WhatsApp

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.