In short: Running AI on your own servers lets a bank or fund issuer decide exactly where information is processed, which documents the AI can read and who can use it. Before any hardware is ordered, you need to decide six things: the first task, the data boundary, the sources to connect, access rights, the model and hardware, and where people approve results. Get those right and the technology choice becomes much simpler.
Most conversations I have with financial firms about AI start with the same question: where does our data go? For a bank, a fund issuer or an asset manager, that is not a technical detail. Client data, draft product documents, internal policies and pricing logic are exactly the information you cannot send to a service you do not control.
Local, or private, AI answers that question directly. At LabForty we install and configure AI in an environment you control, connect the documents and business software you approve and set up access for your team. We use suitable equipment you already have or supply what is needed.
But "on our own servers" is the start of the decision, not the end. Here is what you actually have to decide.
1. Which task comes first?
Start with one concrete task your team repeats. Not "AI for the bank", but something like turning an enquiry into a proposal draft, answering staff questions from approved internal policies or extracting details from incoming documents into a record.
One task gives you something to measure, a small group of users to learn from and a clear scope for hardware. It also keeps the first approval conversation with your risk and compliance colleagues concrete.
2. Where is the data boundary?
Local AI can work with approved information on systems you control. You decide which software it connects to and what information, if any, may leave your environment.
Write that boundary down before you build:
- Which data classes may the AI read: public documents, internal policies, client data?
- Is anything allowed to call an external service, for example a web search or a hosted model for non-sensitive tasks? On our setups, external services are optional and only the ones you choose.
- Which environment hosts the system: your office, your data centre or a private server you control? LabForty sets up private AI either on-premise, on your own hardware, or on private cloud servers dedicated to you.
3. Which sources does it answer from?
An assistant is only as good as the documents behind it. Our assistants answer from approved guides, policies and technical documents, summarise files, compare information and show the sources your team should check.
For a regulated firm, "approved" matters. Decide who owns the document set, how outdated versions are removed and who signs off when a new source is added. An assistant that quotes last year's fee schedule is worse than no assistant.
4. Who can use it, and for what?
Access rules should mirror the rules you already have. If a team cannot open a folder today, the AI should not answer questions from it for them tomorrow.
Decide the user groups, which sources and tools each group can use and how access is reviewed when people change roles.
5. Which model, on which hardware?
The right local model depends on three things: the work it must do, the languages it must handle and the hardware available. Our AI page shows an illustrative shortlist of open models, checked on 7 September 2026, and the spread is wide:
- Google describes a 16 GB memory setup for local use of Gemma 4 12B.
- OpenAI's reference setup for gpt-oss-120b uses one 80 GB GPU.
- DeepSeek-V4-Flash is a large long-context model that needs a dedicated, high-memory server.
Some models suit capable laptops. Others need dedicated servers. That is why we test your own documents and size the hardware before recommending a setup. For a Bulgarian or multilingual team, language quality on your actual documents is part of that test, not an afterthought.
Check the licence too. The models on our shortlist are published under Apache 2.0 or MIT licences, but terms differ between models and versions, so your legal team should confirm them for your use.
6. Where do people approve?
We build workflows that extract details, prepare records in your business software and route exceptions to the right colleague. Your team approves important results before they move on.
In a financial firm, define those approval points explicitly. Anything that reaches a client, changes a record of account or affects a decision about a person should have a named human approver. The AI prepares the work. Your team decides.
What private AI does not solve on its own
Keeping data on your own servers removes one category of risk. It does not remove the others.
- Accuracy. A local model can still be wrong. Showing sources and requiring review is how you catch it.
- Operations. Someone has to update models, patch servers, monitor performance and handle incidents. Plan support and maintenance from the start.
- Regulation. If you are a financial entity, an AI system and the vendor who runs it sit inside your ICT risk and third-party framework, and the EU AI Act may apply depending on the use case. This is a governance question for your compliance team, not something a deployment model settles by itself.
How we run a first project
We ask you to show us one task your team repeats. We then outline the local AI setup, equipment, connections, costs and next steps. We test with your documents, install and test the system with your team, and only then widen the scope.
If you also want AI to take on multi-step back-office work, see our notes on custom AI agents. For the software around it, such as portals, intranets and CRM, see LabForty Core and custom CRM development.
FAQ
Can private AI run without any connection to external AI services?
Yes. Local AI can work with approved information on systems you control, and external services are optional. You decide what, if anything, may leave your environment.
Do we need to buy new hardware?
Not always. We assess suitable equipment you already own or supply what is needed. Hardware depends on the model, which depends on the task and languages.
Which model is best for a bank?
There is no single answer. We shortlist models by task, language and hardware, then test them on your own documents before recommending one.
Does running AI on our own servers make us compliant?
No deployment model makes you compliant on its own. It gives you control over where data is processed, which supports your own risk and compliance work.
How do we start?
With one concrete task. We outline the setup, equipment, connections, costs and next steps for that task first.
Discuss one task
Show us one task your team repeats, and we will outline what a private AI setup for it would involve. Contact LabForty; we reply within 24 to 48 hours on working days.
