When a UK business asks us for a private LLM, it rarely means it wants to train its own model. It means it wants to use AI on client or company data without that data being used for training, seen by people who should not see it, or processed somewhere its policy does not allow. There are three ways to get there, and the cheapest is often enough.
What "private" actually has to mean
The word is used loosely, so it helps to break it into four questions. Is your data used to train the provider's models? Who, at the provider and inside your own firm, can see the prompts and the answers? Where is the data processed and stored, and for how long? And can you show all of this to an auditor, a regulator or a client who asks?
A setup that answers those four questions to your satisfaction is private enough, whatever it is called. A setup that cannot answer them is not private, even if the model runs on your own server.
Option 1: a business plan from a major provider
The business tiers of the big AI providers do not train on your content by default, and they come with admin controls and contractual terms. For Claude, that means the Team or Enterprise plan or the API rather than personal accounts, which are opt-out for training. We compare those plans in our guide to Claude for business in the UK.
This is the right answer for most internal use: drafting, summarising, research, working through your own documents. What it does not give you is UK processing. Anthropic's own service does not currently offer a UK or EU processing location, so personal data sent to it is an international transfer that needs the usual UK GDPR safeguards.
Option 2: a model in your own cloud account, in the UK
The large cloud platforms let you call well-known models from inside your own account, under your existing cloud agreement, with your own logging and access controls. The detail that matters is the region, and it varies by model.
On Amazon Bedrock, as checked on 9 October 2026, Claude Sonnet 5.5 can be called in-region in Europe (London), so requests are processed in the UK. Claude Haiku 5.5 is only available from London through EU or global routing, and Claude Opus 5.5 through global routing only, which means requests can be processed outside the UK (AWS: regional availability by model). The same check is worth doing on any platform: pick the model, then confirm it supports in-region processing where you need it.
This is usually the best fit for automated processes on client records, such as reading documents, extracting fields or drafting reports, where you need to show exactly where the data went. You pay per use, and the build work is in the pipeline around the model: masking, logging, human review and error handling.
Option 3: an open-weight model you run yourself
Open-weight models such as Mistral, Llama or Qwen can be downloaded and run on your own servers or in a private cloud, each under its own licence. Nothing leaves your environment, and there is no per-request bill to a model provider.
The trade-offs are real. You need suitable hardware, usually GPUs, and someone to keep it patched, monitored and secure. The strongest open-weight models are capable, but for complex reasoning and long documents the leading hosted models are often still ahead, so it pays to test on your own documents before committing. And "on our own server" does not on its own make a system compliant: access control, logging and retention still have to be designed.
Self-hosting makes sense when a client contract or policy rules out any third-party processing, when the volume is high and steady enough to justify the hardware, or when the task is narrow enough for a smaller model to do well, such as classifying or extracting from one document type.
Do you need to train or fine-tune your own model?
Almost never at the start. Most "we need our own model" requirements are really "the model needs to know our documents". That is usually solved by retrieval: the system looks up the relevant passages from your own files and gives them to the model with the question, so the answer is grounded in your material and the data stays in your storage. Fine-tuning is worth considering later, for a narrow task with a lot of good examples, once a simpler design has been shown to fall short.
How to choose
Staff using AI for everyday work, with no regulated client data going in: option 1, on a business plan, with a short internal policy on what may not be pasted in.
An automated process on client records, where you must show where data is processed: option 2, in a UK region, with a model confirmed to support in-region processing there.
A contract or policy that rules out any third-party processing, or a narrow high-volume task: option 3, after testing an open-weight model on your own documents.
Whichever you choose, the same groundwork applies: decide what data is allowed in, remove what is not needed, keep a log of what the system did, and keep a person in charge of the decisions that carry risk. If you want a second opinion on which route fits, our AI readiness assessment covers exactly this, and the UK GDPR guide goes deeper on the data protection side.
Frequently asked questions
What is a private LLM? A large language model set up so your data stays under your control: not used for training, visible only to the people who need it, processed where your policy allows, and auditable. It can be a business plan from a provider, a model in your own cloud account, or an open-weight model you host yourself.
Is ChatGPT or Claude a private LLM? Not on a personal account. On business plans and the API, neither trains on your content by default, which covers much of what most firms mean by private. Neither provider's own service currently guarantees UK processing, so for that you need a UK cloud region or self-hosting.
Can a private LLM keep data in the UK? Yes, if the model is self-hosted in the UK or called in-region in a UK cloud region. On Amazon Bedrock, for example, Claude Sonnet 5.5 supports in-region processing in London, while some other Claude models are only available from London through routing that can leave the UK.
Do we need to train our own model? Usually not. Retrieval over your own documents covers most needs, with less cost and risk than training or fine-tuning.
Sources, checked on 9 October 2026: AWS Bedrock regional availability by model, Anthropic data residency documentation, Anthropic pricing.