3CX integration with a self-hosted AI (EU data protection considerations)

[MrNoName]

Hello everyone,

I’m posting here because I don’t have sufficient permissions to post in the appropriate section — this question likely belongs in the “AI” subforum.

I’d like to ask whether there is a way to connect 3CX not only with OpenAI, but also with a self-hosted AI solution (ideally including a knowledge base / RAG, internal documentation, etc.).

Background: I’m based in Europe, and data protection / GDPR compliance is a major factor for us. The use case involves sensitive data, so we want to avoid sending content to external third-party cloud services where possible.

Is there an official method, supported integration, or recommended approach to integrate 3CX with a self-hosted model (or a privately hosted AI stack) while keeping data within our own infrastructure?

Thank you in advance for any guidance.

Best regards

 

[fxbastler]

Is there an official method, supported integration, or recommended approach to integrate 3CX with a self-hosted model (or a privately hosted AI stack) while keeping data within our own infrastructure?

yes, for transcription and AI receptionist atm:
https://www.3cx.com/blog/releases/v20u8-ai-receptionist/
https://www.3cx.com/docs/transcription-engine-server/

 

[MrNoName]

Thanks for your message.

Yes — it’s possible to use a local AI model for the transcription engine.

But I’m also referring to the “voice agent,” i.e., the receptionist.

In the blog post, it only mentions OpenAI, and the docs also only point to OpenAI.

It doesn’t seem like there’s an option to use a locally installed model there.

 

[fxbastler]

But I’m also referring to the “voice agent,” i.e., the receptionist.

look

yes, for transcription and AI receptionist atm:
https://www.3cx.com/blog/releases/v20u8-ai-receptionist/
https://www.3cx.com/docs/transcription-engine-server/

 

[MrNoName]

It says

Previewing the new AI powered Receptionist leveraging OpenAI for smart call handling and dynamic responses.

But I can NOT(!) use OpenAI due to GDPR reasons.

 

[fxbastler]

But I can NOT(!) use OpenAI due to GDPR reasons.

You already wrote that. I already understood it the first time. No need to ...

As I’ve been following the entire discussions about this for months and as it’s mentioned here and there:

Yes you can, so with one GPU server you can serve all you multi tenant customers and include transcription even AI receptionist if you wish!! We plan to make the GPU server available for multiple PBXs as well, so you can run one for multiple PBX under management.

 

[fxbastler]

Just as a side note, we have exactly the same problem (Dresden, Germany).
The only difference is that we haven’t had time yet for a local instance. We’re still putting it off a bit, also because u8 isn’t in production yet.

 

[MrNoName]

I am also from Germany.

When will this v8 be released?

 

[fxbastler]

I am also from Germany.

I know

When will this v8 be released?

It’s done when it’s done.
Sometimes a bit earlier, too.
To win my second bet: pls not in 01/2026

 

[KyriacosS_3CX]

Hello,

Let me clarify that for update 8 the AI Agents (Receptionist or PA) do not use the local Engine but rely on OpenAI for their function.

 

[fxbastler]

Let me clarify that for update 8 the AI Agents (Receptionist or PA) do not use the local Engine but rely on OpenAI for their function.

As mentioned, we haven’t tried this yet due to a lack of time. I’m fairly sure I’ve read that this will be possible. I couldn’t quickly find it again among the large number of posts that came up in a short time about AI agents and handling, but I’m certain there was a statement about it.
Is this perhaps planned for later on the agenda?

 

[VoIPTools]

It is true, with a very busy 3CX system, the OpenAI costs are not trivial (still exponentially more cost-effective and immeasurably faster than having a human perform the same tasks). However, keep in mind that OpenAI has cost-tiers. The more you use, the lower the per-minute price.

At the same time, we need to be realistic about the hardware requirements of running a local engine. Obviously, the more calls you run through the engine, the higher the hardware demands are, and the greater the computer hardware cost. You cannot run a couple of test calls through your laptop and say "it works great". On a 1024 simultaneous call 3CX, the local engine hardware needs are going to be very expensive requiring multiple high-end GPUs costing many thousands of dollars. Where the break-even number is for your environment is obtainable, but remember to take into account the usage-based discounts you get from OpenAI.

I think running a local engine is a very good idea and has the potential to save a significant amount of money over time. But for anything other than a very small 3CX instance, building a high-performance purpose-built computer for this function is not trivial. This is not something you pick up at the big-box retailer down the street. Don't make assumptions based on a quick test on your laptop, and realize as demands increase you may exceed the capacity of your local engine computer and then you have to start thinking about how to scale...

As a business owner I loved the idea of moving my datacenter to the cloud. Keeping enough people trained and certified on many different platforms -- server blades, iSCSI SANs, load-balanced firewalls, switches, security appliances, Microsoft Exchange, etc., etc. etc. was VERY expensive and hard to scale economically for 24/7. Highly skilled people are expensive, hard to find, and challenging to keep. When you make the business decision to move to a local engine you have to look at the total cost both human, and hardware, and what about redundancy? Can you tolerate being down for a couple of weeks while you try and source a GPU that is in short supply (if obtainable at all)?

I like the idea... but this decision is anything but straightforward.

 

[mwd]

Hi All,

Just wondering if there is any progress on this? We are running a locally hosted LLM for our development team and various background AI processing jobs.

We'd love to integrate 3CX into this. We aren't interested in using a cloud/OpenAI solution as we already have the capacity internally.

So just adding a plus 1 to this. Sorry if I have missed another thread about it.