Skip to content
Interface language: English
Audit / Signal

A local AI model: what works without an internet connection?

How to test a local AI setup, work without a network, reconnection and dependencies on additional tools.

Article language: English

Illustration: part of a stone column, with a thin line marking a crack.
In this article

During a demonstration, a local assistant summarises a document while the computer has no internet access. A week later, an employee opens the newly installed application in a room without reception and is asked to download a component. The model is local. Apparently the component has not heard the news.

Both situations can fit the product description. In the first, all necessary files were already on the device. In the second, something needed to start the task was missing. A separate question is whether, when the internet is available, the same assistant uses only the computer or also connected services.

Consider a fictional team, Forma, which summarises equipment instructions. A document moves from a company folder to an application, the model prepares an answer and an employee saves the result in a ticket system. The manager would like to tell a customer, “This process runs locally.” Before saying that, they need to establish which actions the statement covers. Accessing the instructions, generating the answer and saving it to an external system may have three different boundaries.

First establish what must stay on the computer

“Local” needs a noun. A local model, a local application and a local document workflow describe different things. A model may compute on a laptop while the application downloads a document from the cloud. The whole setup may run on one device, but the result may then go to a shared folder. The location of computation alone does not describe the rest of the data's journey.

At Forma, map one specific case: instruction F-18, a question about preparing equipment for transport, an answer identifying passages and a manually saved note. For each step record what reads the file, where the content comes from and where a copy is made. An AI data-flow register helps organise that map. The guide to recording AI data flows (in Polish) describes how to trace the stages. The model name in a chat window is not enough at this stage.

For each storage location, establish how material will be removed after the trial. Closing a conversation may leave an output file, an export or a working copy. In this example, the employee needs to know where the draft note is and whether it is still needed on the laptop after the case is handed over. Assessing locality does not end with the last generated sentence.

Also separate what the company promises the customer from what it needs for work. If the condition is that instructions are not sent to an external service, check the path of their content. If the goal is field work without connectivity, check whether the whole task is available. A solution may meet one condition but not the other. An employee on a factory floor cannot use that distinction when the screen stops responding.

In its offline documentation, LM Studio distinguishes core functions such as chatting with a downloaded model and local document processing from actions requiring a network, including searching for and downloading models, runtimes and updates. This describes the named functions, not every possible combination of add-ons.

In the acceptance record, map those functions to your own configuration. Record application version, selected model, document-reading method and connected tools. A screenshot of the model name alone will not show the rest. If any component changes, an old trial no longer automatically describes the new setup.

When buying, do not reduce the question to “Do you have a local mode?” Ask for a breakdown of what happens on and off the device for your task. This adds a concrete dimension to your assessment of an AI supplier: you can compare solutions against the same process instead of comparing two similar slogans. The guide to evaluating an AI supplier before purchase (in Polish) outlines further questions.

Three trials instead of unplugging once

The first trial concerns a prepared workstation. The person responsible for configuration confirms that components are downloaded and the test document is on the device. Only then is internet access restricted under agreed conditions and the whole task performed. Use a fictional instruction containing a detail the model should not know from general knowledge, such as an invented transport-package designation.

Check more than whether an answer appears. Did the application actually read the chosen instruction? Did it cite the correct passage? Could the result be saved in the location intended for offline work? An answer from conversation memory or a freely composed transport description does not prove the file was read. A fluent paragraph can conceal a missing step.

In a second document put the same question but a different correct answer. Compare both results against a prepared key. If the assistant repeats the first designation, check whether it received the second file rather than merely another question in the same conversation. This pair helps distinguish real reading from apparent work on previously retained content.

The second trial starts by restarting the application under the same conditions. This separates an already-running conversation from the ability to begin new work. Record whether the model, document and answer preparation are available. If the first run used an already-open session, the result after closing it may differ. Do not record one general “offline works” in that case.

Also test a supported file of a different size and material in a format the team does not intend to accept. The first checks the actual working range; the second checks how an unsupported task ends. The application message should identify why it stopped. Failure to read an unusual attachment is not the same as model unavailability. Record these outcomes separately.

Aeneia conceptual photography

The third trial concerns preparing or rebuilding the workstation. Its purpose is different: to identify components and permissions needed before travel or after a computer replacement. Downloading a model may require a network even when later processing does not. Assign responsibility for preparation and a way to confirm readiness. Put that organisational task in the calendar before the trip, not at the station.

For each run record the actual restriction on connectivity. No internet does not necessarily mean no company local network. Disabling one connection does not prove the application could not use another. The person conducting the trial should describe conditions and observations instead of inferring complete isolation from an icon in the system bar.

A connection may return before the task is finished

The next case starts offline and ends after connectivity is restored. At Forma an employee has prepared a local note and indicated an intention to save it in the ticket system. Is the save waiting? Will it run automatically or only after renewed confirmation? Can the employee distinguish a local draft from a note shared with the team?

This is not a minor interface detail. Establish what happens to the deferred operation and document copy when connectivity returns. If the application has no such queue, describe the manual completion route. If it does, check the outcome using test data. Do not assume synchronisation merely because a cloud icon appears beside a file.

During the reconnection trial, change the local draft before sending it. Confirmation should concern the version the employee sees when deciding. Retain identifiers for the input instruction and delivered result in the report. This later helps determine whether an error arose during analysis, manual correction or deferred saving.

The Ollama FAQ describes disabling cloud features, which removes access to cloud models and web search in that product. Its documentation requires a restart after changing the setting. Apply the mechanism to the particular installation; it does not determine the behaviour of other applications or external tools.

Your acceptance should therefore cover the state before the setting changes, after applying it and after restarting. Saving a configuration alone does not prove a running process already uses it. Also distinguish an unavailable feature from a failed task: disabled search may be an intended restriction, not a fault.

Before testing with an unavailable ticket system, establish how the employee will recognise pending cases. With several instructions it is easy to confuse a completed note with a delivered one. The list should connect the output to the document and case numbers, not only the file creation time. Otherwise sorting work after returning to the office may take longer than making the summaries.

An additional tool changes what “local” means

Suppose Forma adds a search of the current parts catalogue. The model stays the same, but one question now requires a connection to another system. That is a sensible extension if consciously agreed. It should not silently inherit the earlier description “everything works without internet”.

Record who performs the search: the application, a separate service or a connected tool server. Then check which data it receives. A part number may be enough to find a designation; sending the whole instruction requires a separate assessment of need. The article on changes to MCP tools (in Polish) explains why the catalogue and meaning of available functions are part of process configuration, even when the user still sees one chat window.

Running locally does not itself define file-reading boundaries. The model needs material supplied by an application or tool, and those may access a wider directory than the selected instruction. Check file and network scope separately in your AI agent sandbox, an environment with defined limits on actions. The article on AI agent sandbox boundaries (in Polish) discusses the file and network limits. One product label cannot replace both assessments.

No connection during the trial shows that the task did not need that connection under the described conditions. On its own it does not prove how the application behaves when the network is available or that every possible operation was checked. If the company needs a stronger statement about data traffic, it must define additional observation and safeguards with the person responsible for the environment.

After an update, repeat tests relevant to the changed component. A new model version may need downloading, an application change may add a feature and a new tool may introduce a network dependency. Identify the changed component and the trials covering it. For other results, record why they still describe the setup in use.

Finally, run a short trial with the person who will work in the field. Give them a prepared workstation, fictional document and clear task. Ask them to identify what is complete, what awaits connectivity and how they will return to pending saves. Your business continuity plan should describe that route as concretely as an external service outage. The guide to continuity during an AI outage (in Polish) discusses how to prepare a plan.

Acceptance might end with a short statement of the scope actually checked: a prepared workstation reads specified local instructions and saves drafts without internet, while transfer to the ticket system requires a connection and employee confirmation. This describes a process another person can take over and test again. In the Aeneia Audit offer, the point of reference is how tools are used in work; the model name is only one element.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear