<!-- Markdown-versie van https://outdeliver.ai/en/blog/data-residency, gemaakt door tools/bouw.mjs -->

Blog

# Data residency in AI: where does your data go?

The question is not only whether data leaves the organisation. Map the full
route for each model, region, storage feature and access right.

"Does our data stay in-house?" is a fair question, but a simple yes or no does
not answer it. An AI solution consists of several steps. An employee asks a
question, the application may retrieve a document, a model processes the text
and the result goes to another system. There are logs, identities, storage and
sometimes a safety check. Each step can have its own rules for location,
retention and access.

If you only ask whether the solution runs "in Azure", you miss those
differences. A good supplier can map the route for each application, model and
setting. Only then can you decide whether the processing fits your own privacy
rules and contracts, and whether the risks are acceptable.

## Map the data route for each task

Consider an agent that handles customer questions. The route might start with a
question in a portal. The agent then reads customer and order data from a CRM
and an order system. The relevant passages are sent with the model call. The
result comes back as a draft reply. An employee approves it, or the application
places the reply in a queue. The processing leaves a task ID and an outcome in a
log.

For each step you want a clear answer. What data is sent? Where does it come
from? In which region does processing take place? Is content stored, and if so,
for how long? Who has access? Does any part go to an external service or to the
open internet? And where can a user check that this is what actually happens?

This is not box-ticking. A system can give technically excellent answers and
still be unsuitable for contract files if conversation history is kept longer
than you want or the chosen deployment processes data in a different geography.

## Be precise about Microsoft Foundry

Microsoft Foundry offers several models and deployment types. For models sold
by Azure, prompts and responses in standard inference are processed within the
chosen geography. Global and Data Zone deployments have different processing
boundaries. Optional features for conversation history, files or vector storage
are also designed to store data. So the feature you choose matters as much as
the name of the model.

For models sold by Azure, Microsoft states that prompts, responses and training
data are not available to other customers or to the model provider, and are not
used to train generative AI foundation models without permission or instruction.
Safety and abuse monitoring, on the other hand, has its own processing rules. If
misuse is suspected, Microsoft may, for example, review a limited set of data
under the applicable terms.

A blanket claim such as "our data goes nowhere" is too vague to base a decision
on. Ask for the exact combination of model, region, deployment type, storage
feature and safety setting. Also check whether a preview feature has different
terms from a generally available service.

## Ask for evidence, not reassurance

A supplier's privacy statement is a starting point, not evidence. Ask for the
resource and region the solution runs in, the chosen deployment type, the
retention period for files and conversations, and the roles that grant access.
Use a request ID or a similar trace to find one processing run in your own logs.
That does not prove the full data route on its own, but it does give the
conversation about data residency something you can check.

Test the route first with a document that holds no sensitive information. Check
which source is consulted, which log line is created and which data remains
after the test. That shows you early on whether the set-up matches what was
agreed. The processing may also call for a Data Protection Impact Assessment.
That assessment needs the actual data route, not a general product brochure.

## Access is part of privacy

Data is not automatically protected just because it sits in your own cloud
environment. Identities and permissions also decide who can use an agent, who
can change a source and who can read logs. Give an agent only the permissions
its task needs. An agent that reads an order status has no need to change
prices. A service employee who reviews a reply has no need to manage the whole
AI environment.

Once every processing step, storage location and access group is visible, the
question changes from "may we use AI?" to "which data may take which route,
exactly?" That is a question you can base a real decision on.

Three ways to start

## Want to check the data route of an existing or planned AI process?

Start small, with one process that comes up every week. Choose how:
from half an hour behind the scenes to a session at your offices.

### Look behind the scenes

Free

Half an hour, by invitation

See everything we build running in our sandbox, the Proeftuin. No sales pitch, unless you ask for one.

[Request an invitation](https://outdeliver.ai/en/microsoft.html#toegang)

02 Recommended

### Introductory call

Free

Half an hour on Teams

You tell us what you already do with AI. We tell you what fits, or whether it is still too early.

Call with Marcel Logtenberg

[Book a call](https://outdeliver.ai/en/book-a-call.html)

### Get started

From €495

At your offices, excluding VAT

An AI quick scan, an AI assessment or an AI assessment plus inspiration session, at a fixed price.

[Get started](https://outdeliver.ai/en/get-started.html)
