Your data leaves the building
Public AI services process your prompts and documents on their servers, under their terms. For case files, patient records or citizen data, that is often not permitted.
Home / Private AI
Private AI for business, running inside your walls.
Private AI means language models that run on hardware, or in a cloud account, that you control, answering from your own documents. Your data is designed to stay inside your own boundary unless you approve otherwise, your team is trained to run the system, and nobody rents the capability back to you by the seat.
- Designed to keep your documents in your building
- No per-seat fees to a third party
- Your team trained to run it
- A written scope before any invoice
The problem
Why do some businesses need private AI?
The easy route to AI is a public service that reads your documents on its servers. For some organizations, that is not a trade they are allowed to make.
Pricing that scales the wrong way
Per-seat pricing looks small for a pilot and grows with every person you add. Hardware you own does not charge per question, though it has running costs of its own.
A tool procurement cannot approve
Where review is strict, a model inside your own boundary can be easier to approve than a public service. It does not replace the review.
How does private AI compare with public AI services and private cloud?
There are three common ways to run a language model for a business. Private cloud is a variant of private AI: the same kind of open-weight model, in a cloud account you control instead of on your own hardware.
| Factor | Private AI on your hardware | Public AI service | Private cloud |
|---|---|---|---|
| Where data is processed | On hardware you own, inside your network | On the provider's servers, under its terms | In a cloud account you control, in a region you choose |
| Who maintains it | Your team, trained at handover, or a support arrangement | The provider maintains the model; you manage accounts and settings | You and the cloud provider share the work: you manage the model and access |
| Model choice | Open-weight models you can inspect and swap | The provider's models, as offered | Open-weight models, and sometimes provider-hosted models under your account terms |
| Cost drivers | Hardware, power, setup and staff time to maintain it | Usage or per-seat fees that grow with people and volume | Cloud compute hours, storage, data transfer and setup |
| Typical fit | Documents that cannot leave your premises, steady workloads, offline use | General tasks on non-sensitive data, and the fastest start | Data that may sit with a cloud provider you already use, and variable load |
The decision depends on where the model should run, which we cover in local AI vs cloud AI, and on the model and vendor, which we cover in OpenAI API or an open-weight model. For connecting a model to the software you already use, see AI integration services.
How does private AI work?
Four steps, all inside your boundary: we size the machine, install it, connect it to your files, and hand it to your people. At the end you own the hardware, the setup and the know-how.
For a small example of the design, see our internal project: a private email assistant that only drafts replies.
Designed to keep data inside the boundary unless you approve otherwise. No rented capability.
Your building
Your documentsPolicies, cases, records
Your hardwareSized, installed, secured
Your teamTrained to run it
1
Size the machine
We match hardware to your documents, your users and your budget, and put the specification in writing before you buy anything.
2
Install and secure
Open-weight models are set up on-premises or in your own cloud tenancy, inside your network boundary.
3
Ground it in your files
Retrieval over your internal documents means answers come from your material, with the source cited. Fine-tuning where it earns its place.
4
Hand over to your people
We train your team to operate it, document how it works and step back. You are not renting a capability from us.
What drives the cost of private AI?
Public services bill by use. Private AI asks for the cost up front and in running the system, and the size of both depends on a short list of factors. We do not publish prices, because they depend on your situation.
You receive a written scope with a fixed price, or a clear rate with a ceiling, before any invoice. The hardware specification is part of it, and you own the hardware.
For a broader look at budgeting an AI project around one workflow, read how much AI automation costs.
Cost drivers
- The model size and quality you need
- How many people use it at the same time
- The volume of documents, and how often they change
- Hardware sizing, especially graphics memory
- Availability needs: spare parts and backups
- Connections to the systems you already use
- Preparing your documents so the model can read them
- Training, documentation and ongoing support
Is private AI more secure?
Not automatically. Running a model yourself gives you control over where data goes and what is logged, and that control only turns into security if the system around it is looked after. A neglected on-premises server can be a weaker position than a well-run managed service.
What decides the result:
- Access control: who can ask the model questions, and which documents each person can reach through it.
- Patching: the operating system, the model server and every library need updates.
- Logging: a record of what the system did, kept without storing more sensitive content than necessary.
- Network design: which machines can reach the model server, and whether it can reach the internet at all.
- Backups and recovery: what happens when the hardware fails.
- Untrusted input: text in emails and documents can try to steer a model, a weakness described in the OWASP guidance on prompt injection.
We design for least access, logging and review, and are open about what private AI does and does not change. The NIST AI Risk Management Framework is a voluntary framework for managing AI risk and a useful checklist for your own review. The same view appears in our posts on local AI vs cloud AI.
What we deliver
What do you own at handover?
Every engagement is scoped in writing. These are the pieces it is built from.
Hardware, specified and installed
Sized to your workload, installed on-site, and secured before it touches your documents.
Open-weight models, deployed
Chosen for your task and licence needs, running on your hardware or in your own cloud tenancy.
Retrieval over your documents
The model answers from your own material and cites where each answer came from, rather than guessing.
Fine-tuning on your data
Where general models fall short on your vocabulary or format, we adapt one to your own material.
Workflow automation
The work around the model: intake, routing, drafting and review, wired into the tools your staff already use.
Training and documentation
A session for your staff, and written documentation for whoever runs it, so it outlasts our engagement.
Who is private AI for?
- Law firms and practices holding privileged files
- Clinics and care providers with patient records
- Agencies and public-sector teams with case or citizen data
- Organizations quoted a per-seat price that scales badly
- Anyone who wants the capability without the dependency
For regulated work, we design the system to support your compliance review. We do not certify compliance and we make no compliance claims on your behalf.
Who it is not for
When a hosted tool is the honest answer.
If your data is already comfortable in a public service and you need a chat window by Friday, a hosted tool is the right call. We will tell you so on the first conversation.
Talk it throughPrivate AI questions, answered plainly.
Private AI means language models that run on hardware, or in a cloud account, that you control, answering from your own documents. It is designed to keep your data inside your boundary unless you approve otherwise.
It depends on the model, the number of people using it and how much material it reads. We size it in writing before you buy anything, and you own it.
Open-weight models, chosen for your task and licence needs. We explain the trade-offs and put the choice in the written scope.
For answering from your own documents, a well-grounded model is often enough. For open-ended work at the frontier, the largest hosted models can be stronger. We will tell you which side of that line your use case sits on.
Not automatically. It gives you control over where data goes, and whether that becomes security depends on access control, patching, logging and network design. See the security section above.
Yes, in your own cloud tenancy, where you control access and where the data lives. We design it so nothing leaves your boundary unless you decide it does.
We design the system to support your compliance review, and we explain what it does and does not do. We do not certify compliance.
Your team. Training and documentation are part of every engagement, so you are not left depending on us.
A fixed price, or a clear rate with a ceiling, in a written scope before any invoice. No discovery charge before you know what you are buying.
Tell us what you are trying to accomplish.
A conversation, then a written scope with a fixed price or a clear rate and ceiling. No discovery invoice before you know what you are buying.