Can we run AI on our own infrastructure, keep our data where it has to stay, and what do we give up?
For firms whose data cannot leave a boundary they set, Private AI runs an open-weight model and agent runtime on your own infrastructure, so nothing leaves it at all, or on AWS in the region you require, in an AWS account you own. You give up frontier-model quality on open-ended tasks and take on maintenance. It takes six to ten weeks and costs from £15,000, plus infrastructure at cost. It needs someone to operate it.
What it costs and how long
- Pricefrom £15,000
- Infrastructureservers, GPUs or cloud costs at cost
- Durationsix to ten weeks
What it needs from your systems and people
- A data owner
- A hosting decision
- The documents or workflow data
- For a fine-tune
What you get, and what you give up
- A data-boundary page
- A reference architecture
- The running system
- An honest comparison
What it does, and where the human sits
- Data boundaryOne agreed page: what data is involved, how it is classified, which countries it may be in, who may access it, and what your data protection impact assessment (DPIA) needs.
- Reference architectureAn open-source agent runtime, an open-weight model, and where it runs: your premises, your cloud, or AWS in the region you require, in an AWS account you own. Plus access control, logging, backups and a cost sheet. The model is chosen by measurement on your task, not by preference.
- Build and measureThe system is stood up on one use case, such as answering questions from your documents or handling a workflow. The private system and a hosted frontier model are each run on a matched test set your data owner approves. Only synthetic or redacted items that your data owner approves ever reach the hosted model. Quality, cost per item and speed are reported, including where the private system loses.
- Fine-tune, only if it earns its placeFor a narrow task such as classifying, extracting or routing, with enough labelled examples of the task, a small model can be trained further on your data. It is tested on examples held back from training. If it does not beat the base model, it is not shipped, and the report says so.
- HandoverA runbook covering how to operate it, update the model, rotate keys and restore from backup, plus the DPIA inputs, the measured numbers and the operating-cost sheet.
Where the human sits. Your data owner agrees the boundary before anything is built. Every external call the system makes is listed and approved. Your operator can start and stop it.
What it needs from your systems and people
- A data owner who can agree the data boundary, for example your data protection lead or FD.
- A hosting decision: your premises, your cloud, or AWS in the region you require, in an AWS account you own, and a budget for the hardware or cloud costs.
- The documents or workflow data for the first use case, and the material to build a test set.
- For a fine-tune: labelled examples of the task. Scoping says how many.
- Someone to operate it after handover, or the Fractional AI Architect retainer, within its days each month.
- If your other systems have no API, the system works from exports, mailboxes or a read-only copy, as any integration does.
Who does which part, and who checks it
Private AI is led by Sam Naseem, and this section is in his words.
- Done by me: the data boundary, the infrastructure, access control, keys, logging and backups, and anything that touches your network or live systems.
- Or by a trusted engineer, under my review: part of that work may be done by an engineer I have worked with, named in the proposal before they have access to your systems.
- Done by AI agents I have trained and direct, under my review: the evaluation harness, running the comparison on approved synthetic or redacted items, the cost sheet, and first drafts of the runbook.
- Who checks it: I review every design decision and every deliverable. Your data owner agrees the boundary, and your operator can start and stop the system.
- What AI agents helping with the work are never allowed to touch: data inside your data boundary, your infrastructure, and your keys.
- How your data is handled: Nothing inside your data boundary reaches a hosted AI service. The AI agents I direct work only with synthetic or redacted items your data owner approves; inside the boundary, I do the work. Working material is deleted or returned within 30 days of handover.
What it costs and how long
- Price: from £15,000, depending on the hosting and whether a fine-tune is in scope, scoped from a completed Opportunity Map. The price is fixed in the written proposal before work starts. Prices exclude VAT.
- Infrastructure: servers, GPUs or cloud costs at cost. The cost sheet shows the cost per item and per month before you commit.
- Duration: six to ten weeks, depending on the hosting and whether a fine-tune is in scope.
Why the price can be fixed in the proposal and the timeline kept short: AI agents I have trained and direct do part of the research, drafting and testing, under my review, so fewer engineering hours go into each build. They never touch your infrastructure, your keys or data inside your boundary; that work is mine.
These durations are the method as designed, not an average of builds already delivered. Where a figure on this page is measured, it is shown with the date it was measured.
Not included: ongoing operation, training a foundation model from scratch, and a fine-tune that does not beat the base model.
The boundary, in plain words: we deliver a running system, a measured comparison and a runbook. Ongoing operation is your team, or the Fractional AI Architect retainer.
What can go wrong, and what we do about it
| What can go wrong | What we do about it |
|---|---|
| You expect frontier-model quality from a smaller model | The comparison report shows the gap on your own task before you commit. The task decides the model size. |
| GPUs cost too much, or are not available | The cost sheet sets out options: compressed (quantised) models, running small tasks without a GPU, or GPU capacity rented from a cloud provider where it is offered inside the boundary you have set. |
| The system is called private but quietly calls an outside service | Every external call is listed in the architecture and must be approved. A system that needs a call you cannot accept is not called private. |
| A fine-tuned model looks good in training and fails on new items | It is tested on examples it never saw. The gap is reported, and it is not shipped unless it beats the base model. |
| Nobody on your side can operate it | The runbook is written for your operator. If there is nobody, we say so before the work starts, and the Fractional AI Architect retainer is the stated alternative. |
What you get, and what you give up
You get:
- A data-boundary page agreed with your data owner.
- A reference architecture with a cost sheet.
- The running system, on your own infrastructure, or on AWS in the region you require, in an AWS account you own.
- An honest comparison, private versus hosted, on a matched test set your data owner approves.
- An optional fine-tune report, measured on held-back examples.
- A runbook: operate, update, rotate, restore.
- A recorded handover.
You don't get: ongoing operation, a foundation model trained from scratch, or a fine-tune that does not beat the base model.
In return, you have control of the data path, a predictable monthly cost, auditability, and a system you own and can operate.
You give up frontier-model quality on open-ended tasks, and you take on maintenance. Someone has to look after it.
Who it is not for
- Firms whose data can go to a hosted service under its contract. For them, a comparison on a matched test set their data owner approves shows whether a hosted service is cheaper or stronger for the task.
- Anyone who expects frontier-level answers to open-ended questions from a private model.
- Anyone who wants a system nobody has to look after.
- Anyone who wants a foundation model trained from scratch. It is not offered.
- Businesses without a clear first use case. Start with an Opportunity Map.
Frequently asked questions
What's it going to cost us, really?
From £15,000 for the build, scoped from a completed Opportunity Map and fixed in the written proposal, plus infrastructure at cost. Before you commit, the cost sheet shows what it costs per item and per month to run. After handover, the running costs and the time of whoever operates it are yours.
Does "private" mean nothing leaves our network?
It means data goes only where the written data boundary says. Every external call the system makes is listed and approved. If the design needs a call you cannot accept, we will say so and will not describe the system as private.
We're not in the UK. Does that change anything?
Not for the architecture. Your own boundary decides the design: what the data is, where it may sit, and who may reach it. On your own infrastructure, nothing leaves it. Where hosting is involved, scoping confirms that the region you require is available, and what it costs there, before you commit. AI UVD is a UK company and prices in pounds. What we will not do is claim a compliance regime we have not been through: if your obligation is a named scheme, we tell you what the architecture can and cannot meet, and what would still have to be certified by someone else.
Why not just use Copilot or ChatGPT?
If a hosted service's contract and data location meet your obligations, a comparison on a matched test set your data owner approves shows whether it is cheaper or stronger for your task. Private AI is for when they do not, or when that comparison shows that running it yourself costs less per item at your volume.
What happens when it gets something wrong?
Items it is unsure about go to a person, it has a written list of things it must never do, and every action is logged. The comparison report also tells you in advance where the private model is weaker than a hosted one, so nobody is surprised.
Who looks after it once it is running?
Your team, using the runbook: how to operate it, update the model, rotate keys and restore from backup. The handover is recorded. If you do not have that person, we will say so before you start; the Fractional AI Architect retainer is the alternative, within its days each month.
Start with an Opportunity Map
Private AI is scoped from a completed Opportunity Map, which sets out the use case, the data boundary questions and whether private hosting is worth what it costs you. The Map costs £4,500, fixed, and takes ten working days.
Every engagement is architected by me, and I'm the person you deal with.
Sam Naseem
We take on a small number of clients at a time.