AI Call Routing Fails

[Allen_Miller]

I am learning by setting up a Receptionist Agent.
In the Call Routing screen I have one test extension set up and in the Description I have a few phrases separated by commas like "support, tech support, customer service"
The agent appears to recognize my test call asking for tech support and says it will transfer my call and then there is silence. If I say "hello" it comes back and says it was unable to do the transfer.
I found the 3CXAI.log file (which is not in the same folder as a lot of other log files.
/var/lib/3cxpbx/Data/logs vs
/var/lib/3cxpbx/Instance1/logs

I am running Version 20.0 Update 8 (Build 1109 Release)

I found this error for my test call
2026-02-18 10:14:45.665|WARNING|{} resolve_contact_destination: analyzer failed for destination='AMCC Technical Support': Error code: 400 - {'error': {'message': "Unsupported parameter: 'temperature' is not supported with this model.", 'type': 'invalid_request_error', 'param': 'temperature', 'code': None}}
2026-02-18 10:14:45.666|DEBUG|{} resolve_contact_destination: returning fallback result for destination='AMCC Technical Support': {'resolution': 'ambiguous', 'extension': None}
2026-02-18 10:14:45.667|DEBUG|{} intent_to_transfer_call_to_ext (confidential): resolver result for destination='AMCC Technical Support': {'resolution': 'ambiguous', 'extension': None}
2026-02-18 10:14:45.667|DEBUG|{'call': 23} wait_until_idle: resumed (AI is idle)
2026-02-18 10:14:45.742|INFO|{'call': 23, 'dir': 'ai'} [OpenAI] conversation.item.added > None role=None id=item_DAdYDY4Fkw3HBCaMfwKx3
2026-02-18 10:14:45.744|INFO|{'call': 23, 'dir': 'ai'} [OpenAI] response.created > resp_DAdYDRX5Oq7r9OxQ1q6gk
2026-02-18 10:14:45.806|INFO|{'call': 23, 'dir': 'ai'} [OpenAI] response.done > resp_DAdYDRX5Oq7r9OxQ1q6gk
2026-02-18 10:14:45.807|DEBUG|{} Audio rate: 0 tokens over 0.770s = 0.0 tokens/min (accumulated=415)
2026-02-18 10:15:16.148|INFO|{'call': 23, 'dir': 'ai'} [Caller] input_audio_buffer.speech_started > None (buffer started)
2026-02-18 10:15:17.192|INFO|{'call': 23, 'dir': 'ai'} [Caller] input_audio_buffer.speech_stopped > None (buffer stopped)
2026-02-18 10:15:17.193|DEBUG|{'call': 23} send system message using lan:"English"
2026-02-18 10:15:17.194|INFO|{'call': 23, 'dir': 'ai'} [OpenAI] conversation.item.added > None role=user id=item_DAdYiiUiFBCxdcDyiTckp
2026-02-18 10:15:17.245|INFO|{'call': 23, 'dir': 'ai'} [OpenAI] conversation.item.added > None role=system id=item_DAdYjDUJnKqjjjqcYwKUp
2026-02-18 10:15:17.273|INFO|{'call': 23, 'dir': 'ai'} [OpenAI] response.created > resp_DAdYjFJgmGqcBtiRN8Acy
2026-02-18 10:15:17.962|INFO|{'call': 23, 'dir': 'ai'} [OpenAI] conversation.item.added > None role=assistant id=item_DAdYjZzPyH3SpDTMKIeLZ
2026-02-18 10:15:18.038|INFO|{'call': 23, 'dir': 'ai'} [OpenAI] output_audio_buffer.started > resp_DAdYjFJgmGqcBtiRN8Acy (AI started speaking)
2026-02-18 10:15:18.039|DEBUG|{'call': 23} speaking: START

Does this have something to do with the OpenAI Model(s) selected in the OpenAI API Configuration?

Thanks

 

[Allen_Miller]

UPDATE

I changed the OpenAI model for text from gpt-5-mini to gpt-4o-mini and my initial calls are getting routed.

Allen

 

[KyriacosS_3CX]

Hello,
gpt-5-mini works for me.

Can you confirm what the Real-Time model you use is? Some additional config details, if something else in the Agent config is not on defaults. Did you wait a few minutes after adding the agent before making the call, multiple calls were made I would assume?

If you switch back to gpt-5-mini, does it stop working again?

 

[Allen_Miller]

The real-time model is gpt-realtime. all are defaults, except the change I made in text model. When I changed the real-time model back to default the problem returned. A little more detail about the experience. The test caller responded by asking for customer service when prompted by the receptionist agent. Agent responded that the connection would be made. then silence for 15 or 20 seconds when caller said - "hello". With that the call was put through. We made another test call and this time waited a minute before saying "hello" and then the call was put through. I changed back to gpt4o-mini, waited several minutes, made test calls and they performed as expected being connected to the queue right away.

 

[nikolascx]

@Allen_Miller Hi How are you? Good to see you!

Indeed - you have stumbled on a problem - we are aware of this.
The problem is that gpt-realtime-mini suffers at calling tools. whilst the gpt-realtime does not.

We have an open case with open ai on this. For now use gpt-realtime (not the gpt-realtime mini) and it will work.

Update 9 has a fix for this. (coming soon within the next 1-2 weeks)

Fix: Use gpt-realtime for now.
When Update 9 is out, you can switch to the cheaper gpt-realtime-mini and the problem will go away.

Send me a pm - next week and Ill give you early access to update 9 Cheers!!!

 

[Jinko]

@nikolascx i would like to have a early acces too.

do i need to change the text to gpt-4-o mini?

 

[KyriacosS_3CX]

@nikolascx i would like to have a early acces too.

do i need to change the text to gpt-4-o mini?

You can change to any model you wish to balance model performance and cost to your needs. The most widely tested is 4o.

Early access is not for general use, your PBX might face other issues.

 

[Jinko]

Can you tell when the version 9 will be released? Seems 8 have a lot of bugs

 

[nikolascx]

@Jinko What models are you using?
Can you send the 3CXAI.log file?

 

[nikolascx]

@Jinko you have 2 models - the realtime and the text (chat completion)
The AI Agent needs them both
Set the realtime to gpt-realtime
Set the chat model to 4o (These 2 in update 8 were the frontier models at the time and the most tested)

If you have problems, send me the logs and the exact scenario. (What you do and what you say and what you expect)

 

[KyriacosS_3CX]

Hi All,

We just published a Blog post about avoiding rate limiting for the AI Agents. You can read it here.