Replace a rigid decision-tree chatbot without losing service quality. Audit, rebuild knowledge, test, cut over channel by channel, and tune.
Your legacy chatbot probably still works, in the narrow sense that it answers the questions someone scripted years ago. But every new product, policy or channel means another branch to edit, and customers who step off the path get "Sorry, I didn't understand that." This guide walks through a low-risk migration to an AI agent: what to audit, what to carry over, how to test, and how to switch over without a bad week for your support team.
Why migrate at all, and why carefully
Legacy bots are usually decision trees or keyword matchers. They are predictable, which is their strength, and brittle, which is their weakness. Typical symptoms:
- Staff edit flows by hand whenever a price, policy or product changes.
- Customers who phrase things unexpectedly hit dead ends.
- Each channel (website chat, WhatsApp, Messenger) has its own copy of the flows, and they drift apart.
- Nobody is sure which flows are still used.
The risk of migrating is the opposite problem: an AI agent is flexible, so it needs good source material and clear boundaries. A careful migration treats the old bot as a source of evidence, not something to clone.
Step 1: Audit what the old bot actually does
Before touching anything new, export or screenshot what you have. If your legacy tool offers transcript exports, pull a representative period. Then build a simple inventory.
What to list
- Intents or flows: every topic the bot handles.
- Traffic: which flows are used often, and which have not fired in months.
- Outcomes: where chats end in a resolved answer, an abandoned chat, or a handoff to a person.
- Integrations: order lookups, booking links, lead forms, CRM writes.
- Channels: where the bot is live and who owns each connection.
- Hard rules: things the bot must always or never say (refund wording, legal disclaimers, regulated topics).
Sort into three buckets
| Bucket | Examples | Migration approach |
|---|---|---|
| Knowledge answers | Hours, shipping, returns, product details | Move into a knowledge base; let the agent compose answers |
| Transactional steps | Collect an email, book a call, look up an order | Rebuild as actions or forms |
| Fixed scripts | Mandatory wording, strict menus | Keep as persona rules, or keep a scripted bot |
Most legacy bots turn out to be 70 percent knowledge answers wearing a flowchart costume. That is an observation, not a statistic: your own audit will tell you the real split.
Step 2: Turn flows into a knowledge base
This is the step that decides quality. A decision tree stores answers inside branches. An AI agent needs those answers as readable, current source material.
- Prefer your real documentation. If help-centre articles, a wiki or policy documents exist, connect those instead of retyping bot replies. They are usually more complete.
- Convert orphaned answers. Answers that exist only inside bot flows become short Q&A entries or text snippets.
- Merge duplicates. Three flows that all say "returns within 30 days" should become one authoritative entry.
- Retire the stale. Remove answers about discontinued products or old prices rather than carrying them forward.
- Write down the gaps. Anything customers ask that nobody documented is now visible. Fill it before launch.
A note on tone
Old bots often sound stiff because each reply was written as a standalone string. Use the migration to define one voice in a persona: how formal, how long, what to do when unsure. One paragraph of clear instruction replaces dozens of inconsistent canned lines.
Step 3: Decide what the agent must never improvise
Flexibility needs guardrails. List the topics where the agent should hand off instead of answering: refunds above a threshold, medical or legal questions, complaints, account security, anything where a wrong answer costs real money.
Also decide what "hand off" means in practice:
- Who is notified, and where?
- Who is on shift at what times?
- What does the customer see while they wait?
The last point matters. A customer told "a human will reply in five minutes" when nobody is online is worse off than one told honestly what happens next. Plan the handoff message around your real staffing.
- 01Audit flows and traffic
- 02Build knowledge and persona
- 03Test with real questions
- 04Cut over channel by channel
Step 4: Test with real customer questions
Your old transcripts are the best test set you will ever have. Pull a few dozen real questions, including messy ones: typos, mixed languages, two questions in one message, angry phrasing.
A simple test routine
- Run the questions through the new agent in a private test environment, not on live channels.
- Mark each answer: correct, partly correct, wrong, or should have been handed off.
- For every miss, decide whether the fix is a missing document, a conflicting document, or a persona rule.
- Fix the source, then re-run the same questions.
- Repeat until misses are rare and, more importantly, safe: the agent says it does not know rather than guessing.
Also test the adversarial cases. Ask for discounts that do not exist. Ask it to ignore its instructions. Ask about a competitor. You want to see the agent stay inside its role.
Step 5: Cut over one channel at a time
Resist the urge to switch everything on a Friday afternoon. A staged cutover limits the blast radius.
Suggested order
- Website chat first. It is the easiest to monitor and roll back.
- A second channel such as Messenger or WhatsApp, once the first behaves.
- Remaining channels after a stable period.
Keep the legacy bot reachable until the new one has proven itself. If the old bot sits behind the same widget slot, plan how you will switch back quickly. Note any channel-specific rules: WhatsApp, for instance, has its own messaging-window conventions that affect proactive outreach, so check those before you plan follow-ups.
Step 6: Monitor, review and tune
Launch is the start of the improvement loop, not the end.
- Read conversations weekly at first. Look for wrong answers, awkward handoffs and questions with no good source.
- Use ratings. Thumbs up and down point you to the chats worth reading first.
- Correct, do not just complain. When a human reply is better than the bot's, feed it back as the ideal answer.
- Watch volume and credits. Know what a typical month looks like so billing never surprises you. For example, if a shop receives 300 chats a month and the agent replies several times per chat, count replies, not chats.
- Review the handoff queue. If the same topic always escalates, either write the missing knowledge or accept that it is a human topic.
After a month or two, retire the legacy bot properly: export anything you want to keep, cancel the subscription, and remove old widget code so nobody sees two chat bubbles.
Doing this in AI SNS CHAT
Here is how the steps above map to AI SNS CHAT. Features are listed with their plan requirements where they apply.
Build and rehearse. Create an agent with the guided wizard, set its persona and guardrails, and try 14 starter personas or the prompt generator if you want a head start. Test in Playground before anything goes live. Playground also offers an A/B model compare. Under Agent → Settings → AI you choose the model provider; advanced models and model fallback need Hobby or higher.
Bring in knowledge. Under Data sources you can add Files, Text snippets, Website, Q&A, Notion, and more. Under Channels, connectors cover Confluence, GitBook, ReadMe, Guru, Zendesk Guide, Google Drive, OneDrive, Dropbox, Box, SharePoint and others, so existing documentation can be linked rather than retyped. Wiki connectors sync up to 1.5M characters and 400 pages at a time, and truncation is reported to you.
Keep scripted behaviour where needed. If some flows must stay fixed, scripted bots (fixed question-and-answer, no AI, connected by webhook) are available from the sidebar's "AI Agent LINE Bot" entry.
Rebuild transactional steps. The "Collect leads" action and custom forms capture details; leads appear in Activity → Leads and also create contacts. Built-in actions include Escalations, Stripe, Shopify, Slack, Cal and Calendly, plus read-only lookup tools for systems such as Freshdesk, HubSpot and WooCommerce. Integrations need Hobby or higher, and the number of Actions per agent depends on plan.
Plan the handoff. Under Helpdesk → Scheduling you set shifts per person and timezone, and handoffs go to whoever is on shift. Alerts can reach Slack, Teams, LINE WORKS, WhatsApp, email or a webhook. The customer's "what happens next" message is built from your rota and never invents a wait time. Helpdesk needs Standard or higher.
Cut over by channel. Deploy from Channels: chat widget, WhatsApp, Messenger, Instagram, Telegram, LINE, Slack and many more. See /connections for the current list, including which channels are still being built.
Tune. In Activity → Chat logs you can filter, flag, escalate, resolve and export conversations, and use thumbs ratings. In chat logs and the email Inbox you can take over a chat, and edit a reply into the ideal answer to train the agent. Restore points (up to 40 per agent) let you roll back a change, and Duplicate copies persona, model and training data for safe experiments. Credits are 1 reply = 1 credit; check /pricing. Synthia, the built-in fleet manager, can guide you to any of these screens.
Frequently asked questions
Can I import my old bot's flows directly?
Not as flows. The practical route is to move the answers into the knowledge base as Q&A entries, snippets or connected documents, then rebuild transactional steps as actions or forms. This usually produces a cleaner agent than copying branches.
What if some conversations must follow a strict script?
Use a scripted bot for those, or put the mandatory wording in the agent's persona and guardrails. Scripted bots in AI SNS CHAT are fixed question-and-answer bots with no AI, connected by webhook.
How do I avoid a bad launch day?
Test with real past questions in Playground, then switch on one channel at a time, starting with website chat. Keep the old bot available until the new agent has had a stable period.
Do I need a paid plan to migrate?
You can start building on the Free plan. Integrations, model fallback and custom domains need Hobby or higher, and Helpdesk scheduling needs Standard or higher. Details are on /pricing.


