Air-gapped AI: what works without a connection and what does not
- Updated
- By
- The MagnolIA team
- On this page
- 11 min read
Short answer
AI works without an internet connection when the model runs on a machine in your own network and does not depend on any external service. Reading documents, drafting or summarising stays available; web search and remote monitoring do not. MagnolIA tested a real installation with the internet cut off: it kept reading documents and answering.
What does it mean for AI to work without the internet?
It means nothing needed to produce an answer sits outside. The model is loaded on a machine in your network, the prompt arrives over the local network and the answer goes back the same way. There is no call to anyone else's server in between.
"Air-gapped" strictly describes a network that is physically isolated from the internet. There is a more common middle ground: an office with a normal connection whose AI does not depend on it. If the line drops, the AI carries on.
It follows directly from on-premise AI: cloud AI services, by definition, do not work offline.
What is the difference between on-premise AI, offline AI and air-gapped AI?
They are three degrees, and they are often confused. The strictest has an official definition. The NIST glossary defines an air gap as "an interface between two systems at which (a) they are not connected physically and (b) any logical connection is not automated (i.e., data is transferred through the interface only manually, under human control)". An office with Wi-Fi whose AI sits on its own server is not that, and does not need to be.
| Aspect | Connected on-premise | AI with no route out | Air-gapped network |
|---|---|---|---|
| What it is | The AI is in your network and the office has internet | The AI machine has no route to the internet, although the office does | The whole network is separated from the internet |
| What leaves | The minimum: with MagnolIA, monitoring of the machine's status | Nothing from the AI machine | Nothing, from any system |
| How an update gets in | Over the network | In an agreed visit, on the machine itself | By hand, on physical media and under a procedure |
| If the line drops | The AI keeps answering | It is unaffected | There is no line |
| Who it is for | An office with no special isolation requirements | Anyone who wants to be able to say the AI talks to nobody outside | Anyone who already runs an isolated network by regulation or contract |
What works without the internet, and what does not?
Whatever runs on the machine in your network stays available. What stops is whatever, by its nature, lives outside.
| Function | Needs the internet? | Why |
|---|---|---|
| Chat, drafting and summaries | No | The model runs on the machine in your network. |
| Reading PDFs, Word, Excel and images | No | The document is read on that same machine. |
| Audio and meeting transcription | No | The speech model runs on your server. |
| Document anonymisation | No | The rules and the model that detect the data are in your network. |
| Search over your internal documentation | No | The documents and their index are in your installation. |
| Coding with the server's models | No | The editor talks to your server, not to the internet. |
| Web search | Yes | It queries pages on the internet. |
| Remote monitoring by MagnolIA | Yes | It is the only thing that travels outside; with no connection, it cannot. |
| Access from outside the office | Yes | Remote access, if enabled, runs through an outbound tunnel that needs a connection. |
| Connections to Google, Microsoft 365 or Notion | Yes | They are internet services. |
| Third-party models, if they were enabled | Yes | They run in their provider's cloud. |
Has it really been tested with the internet cut off?
Yes. In September 2026 MagnolIA installed a machine in the office of a company in the insurance sector that handles patients' health data. Before handover the internet was cut, and the machine kept reading documents and answering.
- The model runs on the machine in the office. The document is read and answered there.
- It does not depend on any external service.
- It is not connected to MagnolIA's cloud or anyone else's: it was delivered without being registered with any MagnolIA service.
- The connection between each computer and the machine is encrypted, with a certificate issued within the installation itself.
- Each person signs in with their own username and password, not a shared account.
The full account, with what goes in and what comes out, is in AI use cases with sensitive data.
Who needs AI that does not depend on the internet?
- Organisations with isolated networks, by regulation or internal policy, where no system may call out.
- Anyone handling special-category data, such as health data: a system that works with no route to the outside is the easiest argument to explain.
- Offices where AI is already part of daily work, and a dropped line cannot stop the whole staff.
- Anyone who does not want to depend on an outside provider's availability, prices or changes of terms.
What leaves the network and what does not in a standard installation is detailed in on-premise AI security. The office tasks that keep working are covered in AI in the office.
What should you bear in mind in an isolated environment?
Isolating AI comes at a price, and it is better to know it before deciding.
- With no remote monitoring, the provider cannot see the machine's status. Oversight has to be organised some other way.
- The model knows what it knew when it was trained. Without web search it cannot look up recent information; it can use the internal documentation you give it.
- Integrations with cloud services are out. Email, calendars or storage on the internet are not reachable from an isolated network.
- Updates do not arrive by themselves. How models and software are updated in an isolated network has to be agreed before installation.
That is why the middle ground is the usual choice: the AI does not need the internet to work, but the office keeps its connection for monitoring and for whatever each person chooses to connect. Tell us about your case and we will see which one fits.
What has to be settled before you isolate an AI system?
A lot of software assumes the internet is there, and you find out the day it is not. These are the points worth going through with any provider before installation, not after.
| Matter | What happens offline | What to agree first |
|---|---|---|
| New models and software patches | They do not arrive by themselves | Who brings them, on what media, how often, and how you check the file is the original and has not been altered |
| Urgent security patches | They do not arrive by themselves either | A fast track outside the normal schedule |
| Licences validated online | A program may stop working if it cannot validate its licence | Ask, component by component, whether any needs to call its vendor |
| Downloads at start-up | A component that fetches something from the internet on first run fails or hangs | Test a full cold start with the cable unplugged |
| System time | With no outside time server the clock can drift; logs and certificates depend on it | Where the machine takes its time from |
| Certificates | They are not renewed against an outside service | Who issues them and who renews them before they expire |
| Backups | They cannot go to a cloud service | Where they are kept inside, and who checks they can be restored |
| Monitoring and alerts | The provider cannot see the machine's status | Who watches load and errors inside, and how a fault is reported |
| Support | There is no remote assistance | Whether support is on site, by phone or over a temporary, supervised connection |
If you build the AI yourself with open tools, go through the same list for each one, starting with where it gets its models from. What each tool does is covered in the software a self-hosted AI server is built with.
How do you verify that nothing really leaves?
The provider saying so is not enough, and neither is this page saying so. You test it, and whoever runs the office network can do the test.
Ask for the list of outbound connections
In writing: which addresses the system calls, what for, and what happens if it cannot reach them. An empty list should be in writing too.
Unplug the cable and work
Not for five minutes: a full day of real tasks. Reading documents, asking questions, transcribing, whatever you intend to use.
Restart offline
Switch the machine off and on with the internet cut. Whatever only fails at start-up is invisible once the system is running.
Watch what tries to get out
With the connection back, the office firewall or DNS log shows which addresses the AI machine tries to call. Whatever appears should be on the list from step one.
Go over what has been connected
Web search, email or storage accounts, third-party models, remote access. Each is a way out, even with the model inside.
Repeat after every update
A new version can bring a new connection.
One way out remains that no technical test closes: people. If someone copies a document and pastes it into an outside service from their browser, it has left all the same. That is known as shadow AI, and it is dealt with through house rules and an internal alternative that works.
When is fully isolating the AI overkill?
Isolation is a measure, not a badge. There are cases where it costs more than it protects.
- If the aim is for documents not to leave, connected on-premise AI already achieves it: the only thing that travels out is the machine's status, not the content.
- If the team needs recent information or its email and calendar accounts, AI with no route out cannot provide them.
- If nobody in the organisation is going to look after updates, an isolated system ages without anyone noticing. An unpatched system is not safer for being disconnected.
- If there is no isolated network around it, isolating only the AI leaves everything else as it was: the computers it is used from are still connected.
When isolation answers a formal security requirement, such as a category under Spain's National Security Framework (ENS), the decision is not one of convenience: the rules set it and the security officer assesses it. What this site currently says about the ENS is in who the ENS binds.
What should you ask a provider of AI for a network with no internet access?
- Does everything work with no connection at all, including installation and start-up, or only once it is running?
- Which outbound connections does the system make by default, and which can be switched off?
- Does any component validate its licence online?
- How do new models and patches get in, and how is it verified that they are the originals?
- Which functions stop being available offline?
- Who watches the machine if there is no remote monitoring, and how is support given?
- Has it been tested with the connection cut, what exactly was tested, and can I repeat the test?
To the last one, MagnolIA answers with what is on record: a real installation, the internet cut before handover, document reading and answers. The rest is verified with you during the pilot.
Frequently asked questions about offline AI
Does the AI work if the internet goes down?
Whatever runs on the machine in your network keeps working without the internet. In the installation we describe as a real case, the machine was tested with the internet cut off before handover and kept reading documents and answering: it does not depend on any external service. Without a connection there is, naturally, no remote monitoring or web search.
What is the difference between on-premise AI and air-gapped AI?
On-premise AI runs on a machine in the company's network, which can stay connected to the internet. Air-gapped AI runs in an isolated network: no physical connection to the outside, with data coming in only by hand, under human control. Between the two sits a middle degree: an AI machine with no route to the internet in a connected office.
How do you update AI on a network with no internet?
On a network with no internet, new models and patches for an AI system do not arrive by themselves: someone has to bring them in, on physical media or in an agreed visit, and check the files are the originals. Who does it, how often, and what happens with an urgent patch should be agreed with the provider before installation.
How do you check that an AI system sends no data to the internet?
Whether an AI system sends data to the internet is something you test rather than assume. Ask the provider for the list of outbound connections, work for a day with the cable unplugged, restart the machine offline, and look at the firewall to see what tries to get out. Repeat the test after every update.
How do people reach the AI from outside the office?
Inside the office, each workstation connects directly to the MagnolIA machine through a single TLS-encrypted entry point. Remote access is optional: if enabled, it goes through an authenticated outbound tunnel, so no port is opened towards the internal network.
What is private web search?
Private web search is MagnolIA's own search engine, which gives the models access to current information without sending queries to third-party services. The AI can answer with recent data without what your team searches for ending up with an outside search engine or provider.
Sources
External figures on this page link to their original source.
NIST (US National Institute of Standards and Technology). Computer Security Resource Center Glossary: "air gap" (definition from CNSSI 4009-2022, from IETF RFC 4949 Ver 2). csrc.nist.gov/glossary/term/air_gap
AI that does not depend on your connection
Tell us what your network looks like and what cannot leave it. We will tell you what would work offline in your case and what a pilot would look like.
Request a pilot


