Custom solution · scoped per organisation
Use AI without sending your data outside
Many finance, security and logistics teams still do not use AI. Their files are confidential: client contracts, payment records, customs files and internal reports. With a public AI tool, these files go to another company’s servers.
In-house AI changes where the AI runs. The AI model runs inside your own environment: on your own servers, or in a private setup that you control. Your team asks a question. The answer comes from your own documents. Nothing is sent to an outside AI company.
Where the AI can run
Choose the level of control you need
We agree the setup with your IT and security team before we build. Each option keeps your data out of public AI tools. They differ in where the hardware stands.
On your own servers
The AI model runs on servers in your own data centre. Questions, documents and answers stay in your building and on your network.
Data stays: in your building
Fully offline
The same setup, with no connection to the internet. This is called air-gapped. Updates arrive through a controlled, checked transfer, so it takes more work to run.
Data stays: in your building, with no network link out
Private cloud in your region
The AI runs in your own sealed account at a cloud provider, in the country or region you choose. It is private, but it runs on the provider’s hardware, under a contract.
Data stays: in your account and your region
The controls around it
Controls your security team can check
Keeping data inside is the first step. These controls decide who can use the AI, what it may answer, and how you check its work.
1. Guardrails
Rules for what the AI may and may not answer. For example, it does not show account numbers, and it does not give investment advice.
2. Access rights
The AI shows each person only what that person is already allowed to see. A branch employee does not get board papers by asking the AI.
3. Audit log
A record of who asked what, when, and what the AI answered. Your auditors can review it.
4. Human review
A person checks the AI’s work before it reaches a client or affects a decision.
5. Security audits
Independent testers try to break the system before it goes live, including by trying to trick the AI. Testing repeats after major changes.
6. Updates and upkeep
New models and security patches are tested and approved before rollout. Your IT team or a partner under contract does this work. We plan it with you from the start.
Where teams start
Internal work that is strong today
Search your own documents
Ask a question about policies, procedures or contracts. The answer shows the document it came from.
Summaries
Short summaries of long reports, email threads and meeting notes.
First drafts
Drafts of reports, letters and replies. A person reviews and sends them.
Checks on files
Pull fields from KYC files, invoices or customs documents, and flag what is missing. People make the decision.
Answers for customers, and decisions such as credit approval, need more controls. We start with internal work.
It is already real
Large organisations already keep AI inside
Public examples. They are not Summerhill clients.
about150,000
staff at Germany’s savings banks use an AI assistant. It runs in the data centres of their own IT provider.
Finanz Informatik, FI-Magazin 1/2026BNP Paribas
is developing on-premises AI with Mistral AI for sensitive work, such as KYC checks.
BNP Paribas press release, 26 May 2026German armed forces
are building a cloud in their own data centres, fully cut off from outside networks. It is planned for the end of 2027.
Bundeswehr, May 2025Honest limits
What in-house AI does not do
It is not free. It needs hardware, setup work and people to keep it running. There is no list price. The cost depends on the number of users, the tasks, the model size and where it runs.
It can still give a wrong answer. Answers show their source, and a person stays in charge of decisions.
It does not remove every risk. Wrong access rules, insider misuse and hidden instructions inside documents are still risks. Access rights, guardrails and security audits reduce them.
It does not fit every task. Models you can run yourself are close to the top cloud models for everyday work. The hardest reasoning tasks still run better on the largest cloud models.
Start small
Start with one team and one task
Tell us which confidential work takes your team the most time. We will help you choose a safe first use case and the right place for the AI to run.
Questions people ask
Does any of our data leave the company?
In a setup on your own servers, or a fully offline setup, questions, documents and answers stay in your environment. In a private cloud setup, data stays in your own account and chosen region, on the cloud provider's hardware under a contract. We agree the setup with your IT and security team before we build.
Is a model we run ourselves good enough?
For everyday work such as search, summaries, drafts and checks on documents, it is good enough in most cases. For the hardest reasoning tasks, the largest cloud models are still ahead. We start where models you run yourself are strong.
Does it meet our regulator's rules?
Rules differ by country and sector, and we do not give legal advice. In-house AI makes it easier to show control: where the data is, who used the AI and what it answered. Your compliance team confirms what your regulator needs.
Can we start small?
Yes. Start with one team, one task and a small group of users. Measure the result, then decide what to add.