Playbook
How to migrate from a rule-based chatbot
By the RoutMe team. Last updated .
Migrate from a rule-based chatbot in stages rather than switching overnight. Audit what the old bot does, keep the content and the rules that still work, connect the AI agent to live data, run both side by side on a slice of traffic, and keep a fallback you can switch back to. The goal is to replace dead ends with real answers, not to throw away what your team already learned.
What a rule-based bot is good and bad at
Decision-tree bots follow scripted paths: menus, buttons and keyword triggers. They are predictable and cheap to reason about, which is why some businesses keep them. They fail when a customer phrases something unexpectedly, asks two things at once or needs a fact that lives in another system. Those failures show up as loops, "I did not understand" messages and customers typing "agent". An AI agent handles free-form language and can call tools, which addresses those failures but needs different controls. For a wider selection checklist see how to choose an AI customer support agent.
Step 1: Audit the existing bot
Export or review every flow and record:
- The intents it covers and the volume each gets, if you have analytics.
- Where customers drop out or ask for a person.
- The answers written into it, and who owns each.
- Integrations it uses, such as an order lookup or a form.
- Any compliance wording that must be preserved.
The fallback and "talk to an agent" statistics are especially valuable: they show what the bot could not do.
Step 2: Sort what to keep
Divide the audit into three piles.
| Pile | Examples | Action |
|---|---|---|
| Content to reuse | Return policy, opening hours, shipping rules | Move into knowledge sources, in plain prose |
| Logic to keep | Verification before order lookup, escalation rules | Rewrite as instructions or flows |
| Scaffolding to drop | Menus, keyword matching, "press 1" paths | Retire |
Clean the content as you move it. Old bots often contain contradictory or outdated answers, and an AI agent will repeat them faithfully. Writing guidance is in what to put in an AI agent's instructions.
Step 3: Connect live data
The biggest gain comes from answering questions the tree could only deflect: where is my order, what is the status of my return, is this in stock. Expose those through read-only tools, as described in how to connect your REST API to an AI agent, and see how an agent looks up order status safely. Start with a small set and confirm each works with real, messy inputs.
Step 4: Rebuild the handoff
Old bots often hand off crudely, dropping the customer into an empty queue. Use the migration to improve it: define triggers, pass the transcript and any lookups, tell the customer what happens next and assign an owner for the inbox. Human handoff best practices covers the design.
Step 5: Test before customers see it
Turn your audit into a test set. Take real historical questions, including the ones that failed, and run them through the new agent. Check answers against your policies, look for invented details and try cases that should be refused. The approach in what grounding is and how to test it applies directly. Have support staff review the results, because they know which answers are subtly wrong.
Step 6: Run in parallel
Do not flip everyone at once. Route a small share of conversations, or a single channel or topic, to the new agent while the old bot handles the rest. Compare using definitions you set in advance, such as conversations resolved without a person, handoffs and repeat contacts. How to measure AI support quality provides definitions. Widen the share as confidence grows, and keep the ability to route back.
Step 7: Cut over and keep improving
When results are steady, move the remaining traffic and retire the old flows. Keep reviewing handed-off conversations weekly: they reveal missing documents, tools and rules. Communicate the change to staff so they know how handed-off conversations will arrive.
Channel notes
If the old bot lived on WhatsApp, keep the same number and check template and opt-in status carry over, since those rules are unchanged by the migration. The WhatsApp compliance page explains how RoutMe handles them. If you are moving from a widget, the new one is typically a script snippet on your site.
How RoutMe approaches this
RoutMe agents run on WhatsApp and a website chat widget, read live data through MCP, APIs or databases in read-only mode, and hand off to a shared inbox with context. You can bring existing content in as knowledge sources and rebuild your handoff rules as instructions or flows. See pricing and the developer documentation.
Sources
Explore the product: RoutMe for Shopify, solutions and the documentation.