WhatsApp was our highest-intent channel. When a billing failure on the Meta side froze our WhatsApp Business Account, every inbound lead that preferred WhatsApp went dark. We couldn't wait out a support queue — we migrated the channel to Twilio in three days, live, without losing a conversation.

The constraint that shaped everything

The CRM never talked to Meta directly. Every messaging provider sat behind a single internal interface:

class MessagingProvider(Protocol):
    async def send_text(self, to: str, body: str) -> MessageId: ...
    async def send_template(self, to: str, template: str, vars: dict) -> MessageId: ...

async def handle_inbound(event: InboundMessage) -> None:
    lead = await leads.upsert_from_contact(event.from_)
    await timeline.record(lead.id, event)
    await router.dispatch(lead, event)

Because the rest of the system only knew that abstraction, swapping providers was a matter of writing a new adapter and a new webhook translator — not touching business logic.

De-risking the cutover

We ran both providers in parallel behind a feature flag keyed per phone number. New conversations were routed to Twilio first; in-flight Meta threads were allowed to drain. A normalization layer mapped both providers' webhook payloads into one internal InboundMessage shape, so the timeline and router never knew which provider a message came from.

The template library was the slow part — every WhatsApp template has to be re-approved by the new provider. We submitted them on day one, in parallel with the adapter work, so approval latency overlapped the build instead of following it.

What made it survivable

• A provider seam that already existed. The migration was an adapter, not a rewrite.
• Idempotent inbound handling. Replays and duplicate webhooks during the cutover were harmless because upsert_from_contact and message dedup were idempotent by design.
• Parallel run, not a big-bang switch. We never had a moment where both old and new were off.

The lesson wasn't about WhatsApp or Twilio specifically. It was that the cost of a vendor emergency is set months earlier — by whether you put a seam between your business logic and the vendor.