Private AI
What “air-gapped AI” actually means, and what it does not
Vendors use “private AI” for very different things. For a regulated organisation, the differences decide whether a deployment is acceptable at all.
Four ways AI can be deployed
Most of the confusion comes from treating these as one thing. They are not. The table below compares them on the questions a risk team normally asks first.
| Model | Where processing runs | Network connection | Who operates it |
|---|---|---|---|
| Public AI service | The provider's shared infrastructure | Required | The provider |
| Your own cloud tenant | The provider's infrastructure, in an account and region you configure | Required | Shared between you and the provider |
| On premises | Hardware you own, inside your network | Usually connected to your network, and sometimes to the internet | You |
| Air-gapped | Hardware you own, on a network with no connection to the internet | None | You |
All four can be described as “private”. Only the last one removes the internet from the picture entirely.
What air-gapped changes in practice
- Nothing goes out. The system has no route to external services, so a question or a document cannot be sent anywhere by accident.
- Updates arrive on media you control. New software, models and reference material are carried in and checked, not downloaded.
- Logging and monitoring stay local. There is no vendor dashboard watching the system. Your team sees what the system does, on your own tooling.
- Support changes shape. Remote troubleshooting is off the table unless you choose to open a session, so documentation and handover matter more.
What it does not mean
Air-gapped is a property of the network. It is not a guarantee about everything else.
- It is not automatically safe. Removable media, insiders and misconfiguration remain real risks. Moving files across the gap needs a controlled process.
- It does not make answers correct. A model on a closed network can be just as wrong as one in the cloud. Showing the source of each answer still matters.
- It does not make an organisation compliant. It helps with one set of concerns, such as data location. Governance, accountability and oversight still need their own work.
Questions to ask any vendor
- Where exactly does each step run: documents, questions, answers, logs?
- Does anything leave our network, including telemetry or crash reports?
- How do updates reach us, and how do we verify them before installing?
- Which models are used, where do they come from, and under which licences?
- What access do you have after deployment, and who starts it?
- What happens to the system and our data if we stop working with you?
When cloud is the better choice
If your policy allows cloud services and you want to scale quickly, a cloud-hosted service may suit you better. An air-gapped deployment makes sense when policy or risk appetite requires processing to stay on your own premises, and when you are prepared to operate it.
If that describes you, the private AI pilot is designed for one department, on your hardware, in four to six weeks.
James Ma runs Maandoh, which deploys private and air-gapped AI for regulated organisations in Singapore and Malaysia. He has more than 15 years of experience in fintech and secure software engineering and is the author of three books.