I created an assistant in VAPI and set up a transfer_call_tool with 2 destination numbers. Using GPT 4.1 and GPT 4.1 mini, the LLM consistently reads the tool line instead of executing it:
{
"role": "bot",
"time": 1745073902354.002,
"source": "",
"endTime": 1745073917664,
"message": "I'm having trouble finding Midwest. Hang on while I connect you with someone who can help. Okay. I see Midwest Automation in our client list. 1 moment while I connect you. Tool code, transfer call tool REDACTED.",
"duration": 14329.9990234375,
"secondsFromStart": 26.900002
}
It is consistent 100% of the time so far rather than the 20% someone else was having problems with. Do you know what I should do to prevent this?