Every AI purchase collapses into one sentence in a security review: where does the data go? Executives ask it of the prompt, of the documents behind the prompt, and of the exhaust the software leaves behind. A cloud assistant sends all three somewhere; a device-local assistant keeps the first two on the machine and shrinks the third to a list you can name.
Retention, Training Use & Telemetry
Where Does Your Data Go, and Does
Iternal or the Model Provider Retain It?
Three separate questions, three separate answers — and the places a device-local deployment still leaves one of them open.
Built from real buyer questions in our sales meetings
Direct Answer
On a device-local deployment, your prompt and your documents stay on the machine. Iternal builds AirgapAI to run 100% local, and states that the customer's data never leaves the device. What a user types lands in a local database and stays in chat history until the user deletes it. Encryption at rest is the device encryption you already run, inherited because the application runs on the PC. On a paid license, telemetry collection is off by default and opting in is an in-app setting.
The limit: the question shrinks, and three things keep it alive. Building a data set at volume is a separate step from chatting with one, and volume preparation runs on a server rather than on the device — a server that can sit on your premises, in your own tenant, or on Iternal's hosted API. Some hybrid designs put a component outside your walls: Iternal describes an architecture where roughly 99 percent of the compute runs on-premises and the orchestration server runs in the cloud. And a hosted model makes retention that provider's answer to give.
Ask the three separately, of the chat step and of the preparation step. Iternal describes one outbound behavior in AirgapAI: a periodic update check, which the installer asks permission for, which can be switched off, and which can point at a file server inside your own network. Licensing adds nothing to the wire either: the license key AirgapAI issues is verified on the device against a public key baked into the application, so there is no activation call.
“Local” and “nothing is kept” settle different things. Local execution settles where computation happens. A retention clause settles what a hosted service may keep and whether it may train on the content — normally its own setting, separate from the ownership terms beside it. For more information on audit evidence, visit the security review page.
What Buyers Are Actually Asking
Buyers raise the destination question early, in unglamorous words. Does the data pass through your servers and come back, or does the whole process stay on ours? Underneath sits an admission repeated again and again — nobody knows where data put into an AI goes, or how long the AI keeps it.
Three Legs, Four Questions
Most datasheets bundle retention, training use and telemetry into one reassuring line, and the bundling is where evaluations go wrong. Split the three, then answer each for every leg your material touches. The last row is the one buyers miss:
| Question | On your device | On a server you choose | On a hosted model |
|---|---|---|---|
| Retention | A local database, kept until the user deletes it. | Your own storage policy governs. | Their contract governs. Require a written window. |
| Training use | Nothing is sent, so nothing can be trained on. | Nothing reaches Iternal on your premises or tenant. | Usually a setting separate from ownership terms. |
| Telemetry | Off by default on a paid license, opt-in in-app; logs stay in the user profile. | Whatever your server emits, under your monitoring. | Their logging, theirs to describe. Ask for the list. |
| Where the corpus is built | On device at reduced precision, roughly a page a minute. | Volume work runs here: Blockify on an Intel Xeon server. | Iternal's hosted Blockify API, at cloud scale. |
The last row produced the sharpest question in the record, asked by a buyer facing his own clients: how do we tell clients we never touch their data if the first pass has to be processed in a secure cloud environment? That first pass does not have to run in anyone else's cloud. Iternal states Blockify can be deployed 100% on-premises, inside a customer's own tenant, or on a Xeon box you own. For more information on building a corpus, visit the keeping the corpus current page.
“You Call It Local, but It Still Phones Home”
Buyers pressed harder on one objection than on any other, and in sharp words. If the data goes back to a cloud server someone else manages, the data is still leaving the premises. Metadata counts as data, because metadata is derived from the data. The objection earns a list rather than a reassurance, so here is what Iternal says AirgapAI puts on the wire:
- One update check. The application checks Iternal's update server and prompts the user to download a newer version. The installer asks permission, and the request can be declined at the cost of updates.
- It can be severed. Iternal describes that check as the only phone-home behavior and states it can be turned off, after which the application connects to nothing.
- It can be pointed inward. Customers can host their own update server, so updates arrive from a local file share.
- No license traffic. The license key AirgapAI issues is verified on the device against a public key baked into the application, so no license check crosses the network.
- Telemetry off by default on a paid license. Collection ceases automatically on conversion to a paid license, and opting back in is an in-app setting most customers are expected to decline. During a free trial, analytics are a condition of the zero-fee license — activation metadata, session frequency, feature-interaction events, version and hardware class — and the trial terms expressly exclude the content of any prompt, answer or document.
Left at factory settings, the software talks once, to one endpoint, for one purpose. Turn every line into a written answer:
Pin it down: questions for your evaluation
-
Is the update check on or off by default in the build we receive?Whether the one outbound behavior is a setting you control on day one.
-
Can the update check point at a file server inside our own network?Whether a sealed environment stays current with no internet route.
-
Is telemetry off in the licensed build we receive, and what is in the payload if we opt in?That opting in is your decision, with every field named.
The Cloud Legs a Local Deployment Can Still Have
Hybrid architectures. Not every deployment is a laptop. Iternal describes one architecture where roughly 99 percent of the compute runs on the customer's premises and the orchestration server runs in the cloud; a buyer holding a hard perimeter rule read that design as short of fully on-premises. Both readings hold, because one counts compute and the other draws a boundary. Iternal names the resolution: customers who need the whole perimeter inside their walls can run the fully on-premises experience, and AirgapAI supports a custom provider pointed at a local host address and port.
Hosted models. AirgapAI runs against a local model and can also be pointed at anything speaking an OpenAI-compatible endpoint. Make that choice and a second organization joins the data path. Iternal's own practice shows both shapes of answer: for its internal use of a frontier model, standard no-data-retention terms sufficed and data processing agreements are in place; in a stricter case a separate agreement with the model provider was required for HIPAA or zero-data-retention terms.
Pin it down: questions for your evaluation
-
Name every component of the proposed architecture that runs outside our premises, and what each one processes.Whether the deployment meets a perimeter rule rather than a share-of-compute rule.
-
Which agreement carries the zero-data-retention and training-use terms for our data class?The exact contract to sign, and whether standard terms suffice.
Four Clauses to Get in Writing
Four clauses settle nearly everything above, and each has to name the step it applies to. Require written answers for the chat step and the corpus-preparation step separately:
- Retention window. What is stored, where it physically resides, for how long, and what deletion actually removes.
- Training use. Whether any prompt, document or derived artifact may train a model, stated separately from ownership terms and for every party in the path.
- Telemetry and metadata. Every outbound call in the lifecycle, named individually, with its default state and the way to disable it.
- Where preparation physically runs. The tenant, facility and legal entity hosting the machine your corpus is built on, plus the three answers above for that machine.
Software whose architecture you understand answers all four in a paragraph. Ask for the paragraph before you buy. For more information on jurisdiction, visit the data residency and sovereignty page.
Answered elsewhere
- Which material may not leave your control — see running AI on data that cannot go to the cloud.
- Which regulated categories may enter an AI system — see the regulated data classes page.
- What keeps working with no network — see the disconnected-operation page.
- Which audit evidence a review board expects — see the security review page.
- What a court or a records request can reach — see the records and privilege page.
- Who may open which data set — see the admin console page.
FAQ
FAQ: Retention, Training Use and Telemetry
It follows the inference. On a device-local deployment such as AirgapAI the prompt and the documents stay on the machine, and Iternal states the customer's data never leaves the device. Point the same application at a hosted model and the prompt travels there instead.
Nothing that stays on the device can be trained on, because nothing is sent. On a hosted leg the answer belongs to the model provider, where refusing training use is normally its own setting, separate from the ownership terms beside it.
Iternal describes one outbound behavior in AirgapAI: a periodic check against an update server, which the installer asks permission for, which can be turned off, and which can point at a file server inside your own network. On a paid license telemetry collection is off by default and opting in is an in-app setting; during a free trial, analytics are collected as a condition of the zero-fee license and exclude the content of prompts, answers and documents. Logging stays in the user profile.
It depends on your data class. For Iternal's own internal use of a frontier model, standard no-data-retention terms covered the use and no elevated agreement was needed. In a stricter case a separate agreement with the model provider was required for HIPAA or zero-data-retention terms.
In AirgapAI, anything a user types is saved to a local database and stays in the chat history until the user deletes it. The files sit in the user profile application-data folders, so your endpoint controls and disk encryption cover them like any other file.
Make the Data Path Something You Can Draw
A buyer who can draw the path a document takes — device, preparation server, model endpoint — clears a security review in one meeting. Draw it for the deployment you are buying, mark every box outside your walls, and collect the four clauses for each one. AirgapAI is built so the drawing ends at the device.