From DIY chatbot stack to managed agent: a migration checklist
Moving off an OpenAI, n8n and Twilio chatbot stack? This checklist covers inventory, WhatsApp number, templates, data access, prompts, handoff, testing and cutover.
Jesbin James
Engineering
To move from a DIY chatbot stack built on a language model API, n8n and Twilio to a managed agent, inventory what the stack does, connect your WhatsApp number directly through Meta's Embedded Signup, import your templates, expose your data as read-only tools, turn your prompts into agent instructions, wire handoff, test on a spare number, and cut over. The checklist below follows that order.
The stack in the title is one common shape, not the only one. If yours uses another model provider, another automation tool or another messaging provider, the same steps apply. This is the long-form companion to the shorter n8n migration guide, and the DIY LLM chatbot comparison shows when staying with your own build is the better call.
Before you start: decide if you should
A managed agent is a good fit when your goal is answering customers, you have more integration code than you want and no one owns the bot full time. Stay with your stack when messaging infrastructure is your product, when you need channels RoutMe does not have live yet, or when you need broad app-to-app automation. Live channels today are WhatsApp and a website chat widget.
Phase 1: Inventory
- 1List every workflow, the trigger for each and the channel it serves.
- 2List every external system the bot reads or writes, and which credential it uses.
- 3List every prompt, with the date it was last changed and why.
- 4List every template, its category, language and status.
- 5List every account you pay for around the bot: model API, provider, hosting, monitoring.
- 6Note who gets paged when it breaks, and how often that happened last quarter.
Phase 2: The WhatsApp connection
- 1Confirm which business owns the WhatsApp Business Account and who can log in as its admin.
- 2Find out what your current provider needs to release or share the number.
- 3Connect a spare test number to RoutMe first, through Embedded Signup.
- 4Import approved templates from WhatsApp, and submit any new ones now, since approval is the slowest step.
- 5Record opt-in evidence for your contact list, because it travels with you, not with the provider.
You do not need your own Twilio account or another provider to run WhatsApp and website chat with RoutMe. The details are in WhatsApp without Twilio.
Phase 3: Data access
- 1For each system the bot reads, decide between an MCP server, a direct API call from your own service or a read-only database view.
- 2Create read-only credentials, separate from anything an application uses.
- 3Expose narrow, named tools such as get order status, and return the minimum fields.
- 4Identify the customer in code from the conversation, not from text the model produces.
- 5Test each connection from the dashboard and read the tool list RoutMe reports.
The tool-by-tool detail is in connect your systems without glue code. Remember that the chat widget only offers tools marked read-only, while agent steps in flows can call every tool on an attached server, so attach carefully.
Phase 4: Prompts become instructions
Open your prompts and split them into three piles. Policy and tone stays and becomes the agent instructions. Rules about WhatsApp, such as not sending free text outside the window, can usually go because the platform enforces them. Workarounds for tool-calling quirks usually go too. Then add what a prompt could not express: named outcomes and a safe default. See what to put in an AI agent's instructions and what is grounding.
Phase 5: Rebuild the flows
- 1Pick the highest-volume path first, usually order or booking status.
- 2Use the message received trigger, an agent step with the right tools, and a reply.
- 3Use If, Switch and Filter for deterministic rules such as return windows, rather than asking the model to reason about them.
- 4Use wait and delay until steps for follow-ups, and the external event trigger for things your systems announce.
- 5Remember the limits: 60 nodes per flow, no loops, no schedule trigger.
Phase 6: Handoff and the team
- 1Decide which situations always go to a person: disputes, repeated failures, anything involving money changing hands.
- 2Wire those outcomes to the assign step.
- 3Invite the people who will answer, and agree on who watches the Unassigned queue.
- 4Check that a teammate can see the transcript and ask for a summary.
The design notes are in designing the handoff to a human and when an AI agent should hand off.
Phase 7: Test
- A happy path for each flow, checked against the real data.
- An unknown question, to confirm it says so and offers a person.
- An attempt to talk the agent into something out of scope.
- A closed 24-hour window, to confirm a template is used where required.
- An opt-out, to confirm no further messages go out.
- A failure of a connected system, to confirm the customer is not left in silence.
Phase 8: Cut over and clean up
- 1Pick a quiet hour and tell your team.
- 2Move the live number, or point customers at the new one.
- 3Watch the run traces and the inbox for the first day.
- 4Keep the old stack available, not running, until a week of real traffic looks right.
- 5Then cancel what you no longer need: the extra provider, the hosting, the old keys.
Costs and plans are on the pricing page, controls are described on the security page, terms are in the glossary, and every route is on Move to RoutMe. A demo is the fastest way to test your own data.
Common mistakes in these migrations
- Starting with the agent instructions instead of the WhatsApp connection and templates, then waiting on approvals at the end.
- Carrying every old workaround into the new flow, so the new setup inherits the old complexity.
- Giving the agent a broad data tool because it was easier than a narrow one.
- Forgetting handoff until the first confusing conversation arrives.
- Switching off the old stack on the day of cutover rather than a week later.
The fix for most of these is the same. Move one narrow path first, prove it, and add the next. A migration that handles order status well and hands off everything else is better than one that tries to cover everything and does none of it reliably.
- Migration
- n8n
- Twilio
- AI agent
Jesbin James
Engineering
Writes about integrations, agent tooling, and the parts of the WhatsApp Cloud API that only show up in production.
Frequently asked questions
Do I lose my OpenAI prompts?
No. Good prompts translate into the agent step's instructions. You will usually shorten them, because the platform handles rules like the 24-hour window and handoff that the prompt used to carry.
How long does the migration take?
It depends on the number of flows, the systems to connect and how long Meta takes to approve templates. We do not publish a timeline because we have no verified figure. Plan the template approvals first, since they are outside your control.
Can I roll back?
If you keep the old stack running until the checklist passes and cut over one number at a time, yes. Do not delete the old workflows until the new flow has handled real traffic.
Keep reading
- Playbooks
How to move a customer support workflow from n8n to RoutMe
A step-by-step migration of a WhatsApp support workflow from n8n to RoutMe Flows: node mapping, the agent step, connectors and a checklist to cut over safely.
- Playbooks
Do you still need Zapier for WhatsApp customer support?
Probably not for the support part. See what a Zap does in a WhatsApp support setup, what RoutMe Flows cover instead, and where Zapier is still the right tool.
See it on your own data
Connect a system, drop an agent node into a flow, and watch it answer from what is actually true in your systems right now.