Private AI

What it actually is

Private AI is a complete AI system — models, retrieval, interface, access control and audit logging — installed where your data already lives. Your staff get the thing they keep asking for, a chat assistant that knows the business, without anyone sending a contract, a patient record or a price list to a public API.

It is built on open-weight models you can run yourself, wrapped in an interface your team will actually use, and grounded in your own documents. The difference from a public AI subscription is not the chat box. It is that nothing leaves.

Who this is for

Businesses where a data leak is not an inconvenience but an existential problem, or where someone has already said no to AI because of exactly that risk:

• Regulated work — medical practices, clinics, law firms, accounting and financial services, where client confidentiality is a licensing condition rather than a preference.
• Manufacturers and engineering firms with proprietary specifications, tooling drawings or formulations that must not end up in somebody else's training corpus.
• Any company with a signed NDA covering the documents staff would most like AI help with.
• Teams already paying per seat for public AI and finding the per-head cost climbs faster than the value.

How deployment actually works

There is no mystery to it, and we do not disappear for three months. The shape is consistent across deployments:

1. Assessment. We look at what your staff are trying to do with AI, what documents matter, and what your existing network and identity setup looks like. This is where most of the decisions get made.
2. Model and hardware selection. Model choice follows the work, not the benchmark leaderboard. A document-retrieval assistant for forty people does not need the same hardware as an engineering-spec assistant for four hundred.
3. Knowledge ingestion. Your documents, policies and procedures are indexed into a private vector store. Retrieval is scoped, so the warehouse team does not get answers from HR files.
4. Access control and audit. Permissions map to your existing groups. Every query and every document retrieval is logged, with retention rules you set.
5. Rollout and training. We get the first group using it properly before expanding. Most of the value is lost when people are handed a login and no method.

What it runs on

Private AI is deliberately flexible about infrastructure, because the right answer depends on what you already own:

• On-premises. A GPU server in your own rack. Highest control, highest up-front cost, no per-token spend.
• Private cloud tenant. Your own isolated environment with a provider you already trust. No shared inference, no training on your data, and nothing co-mingled.
• Hybrid. Sensitive retrieval stays inside; non-sensitive general work can route to a commercial model when that is cheaper. The routing rules are yours and they are explicit.

Worth saying: We do not have a preferred answer we sell to everyone. If you already have capable hardware sitting idle, the honest recommendation is usually to use it.

What determines the cost

Three things, and none of them is a per-seat licence that grows every time you hire:

• Where it runs. Owned hardware is capital spend and near-zero marginal cost. A private cloud tenant is the reverse.
• How much knowledge it holds. Indexing ten thousand documents is not the same job as indexing ten.
• How deep the integration goes. A standalone assistant is straightforward. One that writes back into your ERP is a project.

A bounded version of Private AI ships inside the AI Business Starter Pack at a fixed price, which is the cheapest way to find out whether this belongs in your business before committing to a full deployment. For engine-level token economics we publish what we pay at AI engine pricing.

The objection we hear most

"Isn't a self-hosted model just a worse version of the frontier model?"

Sometimes, and it depends entirely on the job. For open-ended reasoning the frontier models are genuinely ahead. For the work most businesses actually want — find the clause in this contract, summarise this intake form, tell me what our policy says about this, extract these fields from this invoice — a well-configured open-weight model grounded in your own documents performs the task, and it performs it on data you would never have been allowed to send anywhere else.

The comparison that matters is not private AI versus the best public model. It is private AI versus the AI you are currently not allowed to use at all. We wrote that comparison out properly in Private AI vs ChatGPT Enterprise.