I’m exploring the possibility of integrating AI with SuiteCRM, but there is an important requirement: all CRM data must remain on-premises and must not be sent to any external/cloud AI service.
Is it possible to:
Run an AI/LLM model entirely on-premise alongside SuiteCRM?
Train or fine-tune the model using our SuiteCRM data?
Allow the AI to query SuiteCRM records (customers, contacts, opportunities, cases, notes, etc.) without sending the data outside the local network?
Use an open-source model such as Llama or another locally hosted LLM?
Integrate the local AI with SuiteCRM through its APIs?
What would be the recommended architecture and hardware requirements?
I’m particularly interested in hearing from anyone who has already implemented SuiteCRM + local/on-premise AI, especially for organizations with strict data-privacy or compliance requirements.
A local AI is no different than OpenAI in terms of use with SuiteCRM. Its just another endpoint and or API call.
However, you’ll need a server and be able to install and maintain your own local LLM and have the horsepower to do what you need to do. For an SME you’re looking $10,000 for hardware.
Yes, this is definitely possible. A common approach is to run an open-source model such as Llama locally and put a small AI service between the model and SuiteCRM. The AI service can use the SuiteCRM API to retrieve only the records needed for a particular question, while the model and all retrieved data remain inside the private network.
I’d probably start with RAG rather than fine-tuning the model on the entire CRM database. Keep the structured CRM records in SuiteCRM, retrieve relevant data through the API, and provide that context to the local LLM at query time. Fine-tuning is more useful when you need to change the model’s behavior or terminology rather than simply give it access to current CRM information.
For hardware, the main factor is the model size and whether you want GPU inference. A smaller quantized model can run on a reasonably powerful workstation, while larger models benefit considerably from a dedicated GPU with sufficient VRAM. I’d also make sure the AI service has strict API permissions so the model can only access the SuiteCRM data that the requesting user is authorized to see.
Hi Rsp, yes: it’s definetely possible.
I would not recommend training the LLM using SuiteCRM data as the main architecture. The right approach is a local LLM + SuiteCRM API/tool calling + local RAG + strict ACLs.
I’ve already recorded some demos and built multiple different approaches.
The most important question first:
What are your requirements? What processes do you want to support with AI?
So far, my best experiences are fairly deterministic processes and the interface / interaction is AI enabled.
Here, you’d define business processes like
show me all my meetings for today
summarize a phone call, add action items as and details like sales performance (who talked to how much %, how many questions have been asked, ..)
sentiment analysis
In order to make processes which include AI handing over payloads to other AI processes, try to reduce the payload as much as possible to the barely required minimum of fields and a clean structure, either like this:
and/or build the return values / structure directly in your deterministic processes.
I’ve successfully used AI processes on 8B llama 3.1:
Even more interesting are the MoE models now like Qwen3
try out the quantized models as well - it’s quite a bit of trial and error and it heavily depends on your requirements / use cases.
AI can run fast and easy on standard hardware - even without GPU when you’re using batched / cron / async processes.
My worst experiences so far, are with nonsense MCP wrappers around the standard API and skills that query the API directly and keep everything open to the prompt.
Here, I wouldn’t even waste another minute on anything local - go for the biggest and most expensive models out there in the cloud. Unfortunately, the smaller models are nowhere close to support solid and secure business processes in a non deterministic way.
One alternative approach:
There is an alternative to local: You can just rent / spin up some decent hardware in a data center and selfhost your models there, instead of spending 5 digits on hardware and testing whether things might work out.
