A private AI stack is an AI setup that runs in a workspace your business owns: the chat interface, the documents, the indexes and the logs stay in systems you control. When the stack calls a model, whether a provider's API or one you run yourself, the choice and the accounts are yours, not a reseller's. The difference with ChatGPT Enterprise is who holds the data, who chooses the models, and what you are trusting: a contract with a provider, or your own environment.
What the stack is made of
The word stack is doing real work in that sentence, because there is no single product called a private AI stack. A typical one has four layers:
- An interface your team actually uses, usually a chat application that handles logins and conversation history. LibreChat, the open-source platform I build on, is one example: one interface in front of several providers.
- Model access, either external models through API keys in your name, models on hardware you control, or a mix, with routing decided per task.
- Your information: the document store, the search index and any automation state, on machines whose location and backup you know.
- Operations: logs, updates and access rules, in systems you or your team can inspect.
Open-weight models such as Llama, Mistral or Qwen run on hardware you own or a server you control, served through tools like vLLM, Ollama or llama.cpp, so the prompts and documents never leave your infrastructure. The trade is real: you carry the hardware or server cost and the work of running and updating it, and quality and context length against the frontier hosted models vary by task. A mix works too: an open-weight model for the documents that must stay local, a hosted API for the rest.
ChatGPT Enterprise gives you the first layer and administers the rest inside OpenAI. A private stack puts all four under your roof, and puts the work of running them there too.
How ChatGPT Enterprise is private, and where it stops
OpenAI's own enterprise page states that business data, including conversations, files and outputs, is not used to train its models by default, and that the plan provides encryption, retention and compliance controls, with admin settings over who can use which tools.
That is privacy by contract and by provider controls. The checks to run before you rely on it for regulated work are the ones I list in what to verify in OpenAI's data processing agreement.
A stack you own differs on three points. The models are chosen per task, not per vendor, so a cheap model can do routine work while a stronger one handles the hard calls, under your own accounts. The data path is visible: you can see which component calls which service and read your own logs. And the decisions about access, retention and residency are yours to make and change, because the pieces are yours.
What private does not mean
Private is a property of the architecture, not of the pricing page. Private hosting alone does not prove that no data leaves. If a component of your stack calls an external model, that data leaves your systems, whatever the hosting label says. Data residency depends on the architecture: where the models run, where the indexes live, where backups land, what gets logged and who can read it.
So before calling any setup private, I check where each step of the work actually runs. I wrote the checklist for the document side of that question in mapping the whole document path before calling a workflow private. The same discipline applies to chat: the honest answer to "can our data stay put?" is "here is exactly where it goes, step by step", not a fixed table of assurances.
Can a small business run one?
Yes, without an internal AI team. The realistic shape is one interface, one or two recurring tasks, external models billed to your own accounts, and an open-weight model only for the work that genuinely has to stay local. My own setup, which I have run every day for two years, is built that way. I worked through a full cost example in a self-hosted versus API cost comparison.
What I build through CPLT follows from that: I scope one process with you, agree the models, the scope and the human checks in writing, build the stack in a workspace you own, test the handover with your team, and leave. Your team operates it afterwards, with documentation, not a dependency on me.
Frequently asked questions
Is ChatGPT Enterprise a private AI stack?
Only by contract. OpenAI itself ships an open-weight model, gpt-oss, separate from ChatGPT Enterprise: if your data path itself needs to stay inside your systems rather than inside a contract, that is the kind of model you would run yourself, not ChatGPT Enterprise.
Does running an open-weight model locally guarantee my data stays private?
No. Running the model locally keeps that one call inside your systems, but the rest of the stack, OCR, embeddings, logging, backups, can still call an external service. The same discipline applies: check where each component actually runs, not just where the model sits.
What is LibreChat?
An open-source AI platform you can host yourself. It gives your team one interface in front of several hosted providers, including Anthropic, OpenAI and Azure, or a model you run yourself. I use it as the interface layer of the stacks I build.
If the question is your data
If your team keeps asking where the documents actually go, that question is worth answering with an architecture, not a reassurance. Bring it to a free 45-minute call, and I will tell you plainly whether a private stack, a provider plan, or a mix fits the work you do. The private AI hosting service describes how such a workspace is built and handed over.