Vapi and Retell are good products and starting on one of them is usually the right call. But there is a point where the platform becomes the ceiling, and the question turns into whether to rebuild. Here is how I decide, and how I run it when the answer is yes.
The only four reasons that justify it
"We want more control" is not a reason. It is the mechanism. If you cannot name which of the four applies, stay where you are.
What you are actually taking on
The prompts, conversation design and tool definitions port over in days. What does not port is everything the platform was doing quietly: worker autoscaling, retry and failure handling, turn-detection tuning, call recording and storage, observability, and telephony configuration. Budget several weeks for that alone. A migration that skips it produces an agent that is cheaper per minute and worse to talk to, which is the worst possible outcome.
The sequence that works
How long and how much
Four to eight weeks and $10,000 to $28,000 for a typical single-agent migration including the comparison and gradual cutover. If your agent is simple and volume is low, I will usually tell you not to bother. The full scope is on the migration service page, or send me your current numbers.