How to use AI to manage your company's internal knowledge securely, without data leaving your infrastructure
To use AI to manage your company's internal knowledge securely, run the whole system on servers you own: the model that writes the answers, the search index built from your documents, the access controls and the audit log, with nothing sent to an outside service. Mickai does this as an AI operating system (SIOS) that installs on your own hardware, runs fully offline, turns your company's documents into private knowledge bases you can scope by department or sensitivity, and records every action it takes in a signed, tamper-evident audit log you can check yourself. Whichever route you choose, prove it by unplugging the network: indexing, search, answers with their sources and audit logging should all keep working.
What does "data never leaves your infrastructure" actually require?
It requires every stage of the system to run on hardware you control, not only the part that stores files. Private AI over documents has several moving parts, and any one of them can quietly send data out. Check each of these: the language model that writes answers; the embedding step that turns text into searchable vectors; the index that stores those vectors; the conversion of PDFs and scanned pages into text; the logs of questions and answers; licence checks and software updates; and any remote support access. A hosted embedding service or a telemetry feed is enough to move content off site, even when the model itself runs locally. The simplest proof is physical. Disconnect the network and confirm that indexing, search, answers and audit logging all still work. If anything stops, something was calling out.
What are the realistic ways to do it?
There are three broad routes, and each suits a different organisation. The first is to assemble your own system from open-source components on your servers, following self-hosting guides. That gives full control, but your team owns the integration, access control, security hardening, audit logging and upkeep. The second is a private or on-premise deployment from an established AI vendor. Several offer deployments where the vendor cannot see your data, which is a strong baseline, so check which components still depend on a connected environment. The third is an operating system built for the job, which is where Mickai sits. The AI runtime, private knowledge bases, access control and a signed audit record arrive together, installed on hardware you own and designed to run with no network at all. Choose on three questions: who maintains it, what still connects out, and what evidence it produces.
How does AI answer questions from internal documents on your own servers?
Most private knowledge systems use a pattern called retrieval-augmented generation. Documents are converted to text, split into passages with their structure kept, and turned into vectors by a model running locally. The vectors go into an index on your hardware. When someone asks a question, the system searches that index for the most relevant passages the person is permitted to see, and a local language model writes an answer that cites them. The model does not need to be trained on your documents. That matters: remove a document from the index and it stops informing answers, because nothing about it was built into the model itself. In Mickai, your records become private knowledge bases, which Mickai calls brains, built and indexed on your own hardware and scoped by department or sensitivity so access is controlled at that level. The documents, the models and the questions asked of them stay in the building.
What hardware and deployment do you actually need?
Sizing depends on three things: how many people ask questions at once, how large the document collection is, and how capable a model you want to run. The system can run on processors alone, a suitable graphics card adds headroom for larger models, and heavier work such as reading scanned pages with an AI model needs one. Mickai runs on customer-owned hardware in three shapes: a single workstation for the most sensitive work, a server in your own server room serving departments over your office network, and a fully air-gapped setup with no network path at all. It scales by adding machines you own. Plan for the work around the hardware too: indexing the existing collection, keeping the index in step with source systems, backing up the index and the audit record, and deciding how updates reach a machine that is offline. A pilot on one collection is the most reliable way to size a wider rollout.
How do you keep it secure and prove what the AI did?
Security for internal knowledge comes down to three controls. Access: answers must honour the same permissions as the source documents, so a person never receives a summary of a file they cannot open. Keys: encryption and signing keys should be generated and held on your own hardware, with no vendor escrow, which is how Mickai handles them. Evidence: you need a record of what the system did that an auditor can check without trusting the vendor. Mickai signs every action before it executes and seals it to the Open Audit Record, an append-only, hash-chained log. The record is tamper-evident: alterations to past entries can be detected afterwards by someone who does not have to trust Mickai. Signatures use ML-DSA-65 (FIPS 204), a post-quantum standard, and verify offline with the operator's public key, with no call to Mickai. Outbound actions, such as sending an email, happen only when the operator authorises and signs them, and each is recorded before it leaves.
What do UK data protection rules expect?
Running AI on your own servers does not make it compliant on its own, but it removes several hard questions. UK GDPR and the Data Protection Act 2018 still apply to personal data in your documents and in the questions staff ask. What changes is the shape of the work. If the supplier has no access to the data, there is no outside AI provider processing it on your behalf, no international transfer to assess when the servers are in the UK, and no vendor holding copies of prompts. You remain the controller, so lawful basis, retention, subject access and erasure all apply, including to what sits in the search index. The ICO's guidance on AI and data protection says most AI use involving personal data is likely to need a data protection impact assessment. The NCSC's guidelines for secure AI system development, written with international partners, cover secure design, development, deployment, and operation and maintenance. FCA-regulated firms should also map the system against their operational resilience and third-party risk obligations.
How should you evaluate or pilot a private AI knowledge base?
Run a short pilot on one bounded collection, such as a policy library, a procedures manual or a contracts archive, with a named group of users and success measures agreed in writing before it starts. Then put every option through the same six tests. Disconnect the network and confirm everything still works. Ask a question as a user who lacks access to a document and confirm the answer does not reveal it. Check that each answer cites passages you can open and read. Delete a document and confirm it no longer informs answers. Export the audit record and verify it offline. Finally, confirm how you would take your documents and index elsewhere if you stopped using the product. If your documents are scattered or poorly organised, fix that first: Mickai's AI Readiness engagement covers that step. To try Mickai itself, request a briefing or apply to join the Mickai SIOS beta.
Frequently asked questions
Can we use AI on our internal documents without uploading them to the cloud?
Yes. The model, the search index, the access controls and the logs can all run on servers you own, so documents are processed inside your building and nothing is sent to an outside service. The test is simple: disconnect the network and check that indexing, search and answers still work. Mickai is built this way, installing on customer-owned hardware and running fully offline.
Do we need to train an AI model on our company documents?
Usually not. Most private knowledge systems use retrieval: the system searches an index of your documents for relevant passages and a language model writes an answer that cites them. Your documents stay in the index rather than being built into the model, which keeps updates and deletions straightforward. Remove a document from the index and it stops informing answers.
Will staff see answers from documents they are not allowed to access?
They should not, and it is worth testing before you commit. A well-built system applies access rules when it searches, so material a person is not cleared for is never used in their answer. In Mickai, knowledge bases can be scoped by department or sensitivity, so access is controlled at that level. In a pilot, ask the same question as two users with different access and compare the answers and sources shown.
What hardware do I need to run a private AI knowledge base on-premise?
It depends on how many people use it at once, how many documents you index and which model you run. The system can run on processors alone, a suitable graphics card adds headroom for larger models, and reading scanned pages with an AI model needs one. Mickai runs on a single workstation, on servers in your own server room, or in an air-gapped setup, and scales by adding machines you own.
Does keeping AI on our own servers make us UK GDPR compliant?
Not by itself. If the supplier has no access to the data, it removes the outside AI processor and, with servers in the UK, international transfers, but your organisation remains the controller. Lawful basis, retention, subject access and erasure still apply, including to content held in the search index. The ICO's guidance says most AI use involving personal data is likely to need a data protection impact assessment, so plan one alongside the technical rollout.
How can we prove to an auditor what the AI system did?
Keep a signed record that can be checked without trusting the vendor. Mickai signs every action before it executes and seals it to the Open Audit Record, a tamper-evident, hash-chained log in which alterations to past entries can be detected afterwards. Signatures use ML-DSA-65 (FIPS 204) and can be verified offline with the operator's public key, with no call to Mickai.
Does Mickai need an internet connection to work?
No. Once installed, Mickai is an operating system that runs on hardware you own and is designed to work with the network physically disconnected. Documents, models, questions and the audit record stay on your machines. Where an organisation chooses to connect a workload, an outbound action happens only when the operator authorises and signs it, and it is recorded before it leaves.
How do we start a pilot with Mickai?
Pick one bounded document collection and a small group of users, and agree success measures in writing before anything is installed. Then request a briefing to scope the pilot on your own hardware, or apply to join the Mickai SIOS beta. If your documents are scattered or need organising first, the AI Readiness engagement covers that step before the pilot begins.
Related reading
- What sovereign AI means: AI that runs on hardware you own, fully offline
- How the Mickai operating system is built, from your hardware to the audit record
- Mickai security architecture: operator-held keys, attestation and the air gap
- The Open Audit Record: signed, hash-chained and verifiable offline
- Verify an Open Audit Record chain in your browser
- AI Readiness: getting your documents and data ready for AI
- Have Mickai build a private system on your own records
Request a briefing to scope a pilot of Mickai on one of your own document collections, or apply to join the Mickai SIOS beta.