AI not forwarding calls to call queues

[gbtinc]

Speaking to the AI for basic support based on knowledge sources works perfectly but once asked to transfer the call, the AI confirms the caller but is stuck after that. The call never transfers and sometimes the AI will ask if the caller is still there.

this seem to be the most common error in the logs:
2026-03-03 15:05:20.846|WARNING|{'call': 59, 'dir': 'pbx'} Endpoint PC lost connection; forcing idle & closing bridge

openai model for real-time calls set to gpt-realtime
openai model for text set to gpt-5-mini
i've test each variable as well and get the same results.

on the open ai side, the only limit i've set is the $20 monthly budget. i've set the allowed models to the ones defined by 3cx; https://www.3cx.com/docs/ai-agent-configuration/ and there are no rate limits on each model other than default.

 

[jim@valley-tel]

I am having a similar issue. When asking it to transfer to a sales or service queue it says it is not listed. Tried to make a dummy extension that is forwarded to the queue but it detects it's not available. Is there a workaround for this?

 

[nikolascx]

Hello Dears
@gbtinc It could be a number of issues.
Without full logs its hard to understand.
What are you transferring to? (Is it a long wait for the transfer to complete?)

These are the things I have in mind
1. Realtime API session expires when audio passes during transfers
2. Transfer target not reachable
3. for some reason the AI Agent disconnects because of a failed tool call (I suspect your 3CX Version plus mixing realtime model with a mini 5 model)
4. Model mismatch (this is very common especially with your current 3CX installation version)

Can you replace the text model to a 4o model example gpt-4o-mini or gpt-4.1-mini.
Let me know how this works

@jim@valley-tel Dear Jim, your issue is definitely the result of a small model (Small models are not understanding all the destinations properly)

We are making fixes for this in update 9 so stay tuned. In the meantime, let us know and you can contact me in private so I can inform you how to get debug logs.. (The process is a bit tricky but this is also fixed in update 9)

 

[nikolascx]

Dears an additional update

SP8 was tested mainly with this combination:
gpt-realtime and 4o.
This works 100% (We also needed a baseline to understand which are the best models for successful evaluation)

It's excellent to experiment with other cheaper models and I am here at your disposal to help. But let's agree to do the following:

If you have a scenario that fails, immediately switch to the gpt-realtime and 4o pair to see if the problem occurs or not. Like this we can safely exclude a model/scenario incompatability issue.

I suspect, that the tool calling on small models in update 8 fails (this is an open ai issue and we reported it here
https://community.openai.com/t/gpt-...tly-seems-to-be-an-error-in-the-model/1373936

Moving forward, we are preparing Update 9.
Update 9 will work and will be fully tested with the cheapest models

gpt-realtime mini and gpt-5-mini ($0.25 $0.025 $2.00) or
gpt-realtime mini and gpt-5-nano ($0.05 $0.005 $0.40)

This is the plan - let me know about your cases.

We are trying to push a release of update 9 really quick (2-3 weeks)..

 

[gbtinc]

Hi @nikolascx

your guide recommended GPT 5 mini https://www.3cx.com/docs/ai-agent-configuration/

I changed to gpt-realtime and 4o and still having the same results. I will send you my 3cxai.log file directly.

 

[nikolascx]

Do you have a custom prompt?
Prompt looks very big. I have a feeling it might be larger and if it is too large, the model will "forget" parts of it.
Im afraid from your logs, it looks it "forgot" that it has a tool that it can use to transfer. this is why you hear silence.

ILl send you another prompt in private. thanks for the logs - they totally helped.

 

[gbtinc]

I've tried modifying the existing agent with the new prompt, it stops responding after asking for a transfer. I then created a new agent with the same prompt and it stops responding after asking for my name and reason why i'm calling. send you the logs now.

 

[nikolascx]

@gbtinc rate limit issue
'realtime.response', 'id': 'resp_XXXXXXXXXXX', 'status': 'failed', 'status_details': {'type': 'failed', 'error': {'type': 'tokens', 'code': 'rate_limit_exceeded', 'message': 'Rate limit reached for gpt-4o-realtime in organization org-XXXXXXXXXXXXXXXX on tokens per min (TPM): Limit 40000, Used 31179, Requested 14691. Please try again in 8.805s.

To the rest - this is what happens when rate limits exceed.

We will make a blog post about this shortly after the release of update 9 and we will explain how to go around it. In this case, this pbx is too large for the standard tier of openai (40000 tokens) and we need to increase the limit.

Limit 40000, Used 31179, Requested 14691

This is when the silence occurs

See if you can increase the limits or ask openAI to put you on a higher tier (then you inherit higher rate limits)

Keep me posted

 

[gbtinc]

Hi Nikolas,

Our account was on tier 1 when the issue was happening. After being bumped up to tier 3, there is no longer an issue.

thank you for your help!

 

[gbtinc]

although, now i've encountered a new issue where the transcription service no longer works. both the openai and 3cx cloud transcription does not work. it does not work when manually pressing the transcribe call under call logs and under recordings and it does not work under the department settings.

i've tried deleting the api key for openai and recreating a new one. switching back and forth from openai to 3cx cloud and restarting the AI service.

would the logs be the same 3cxai.log?

 

[nikolascx]

although, now i've encountered a new issue where the transcription service no longer works. both the openai and 3cx cloud transcription does not work. it does not work when manually pressing the transcribe call under call logs and under recordings and it does not work under the department settings.

i've tried deleting the api key for openai and recreating a new one. switching back and forth from openai to 3cx cloud and restarting the AI service.

would the logs be the same 3cxai.log?

Hi dear @gbtinc
You are definitely missing some setting.
If you do not have an ENT + (plus) license key, you cannot use 3CX Cloud.
Stick to openai in this case.
You can use GPT4o Transcribe for STT model and gpt5 mini
Put in the API Key
ANd save.

In departments,. you set "Select what to transcribe"

In the extensions, go to Options > Transcription and set "Use Department Settings"

Then you should be ready to go.

The log file that will show you the progress is 3cxSystemService.log

Let me know how it goes. You will manage.

 

[KyriacosS_3CX]

Hi All,

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