Unanswered

Is data used for training/operating an AI model kept within Tenant?

(0) Share

Report

Posted on by BennyS

Hi Community!

I just heard that the AI Builder functionality is an isolated service within the environment/tenant, so no training/operation data is leaving the tenant.

However, I cannot find any documentation on this. Does somebody know where this is written down?

Best regards!

Benny

Categories:

AI Builder

I have the same question (0)

All responses (5)

Answers (0)

Antrod Moderator on at

Like (0)

Report

Hi @BennyS ,

Indeed, no training data is leaving the tenant's environment unless you decide it 🙂

The training data is stored in Dataverse in the environment where you created your model so it relies on Dataverse's security principles. There's no possibility for someone to access this data without the required rights.

Also the data is never copied or moved outside the environment by any automated or manual process owned by Microsoft. The data is only temporarily copied to Microsoft Azure storages (in the same environment's geo) to train the model. During the transfer, the data is encrypted and is removed after model is trained.

Learn more here: Administer AI Builder - AI Builder | Microsoft Learn (this documentation provides a link to the governance whitepaper, which could be helpful to you).

Was this reply helpful? Yes No
BennyS 8 on at

Like (0)

Report

Okay thank you for that response.

However, there is still the data riks of the "abuse monitoring" in which scenario a MSFT employee can have access to prompts and responses, right? Does this unfortunately impose a risk.

Or isn't this the case for "on your data" training?

Was this reply helpful? Yes No
Antrod Moderator on at

Like (0)

Report

Hi @BennyS ,

Accessing customer's data for a MSFT employee requires customers approval and a very transparent, just-in-time and controlled process, mainly for support cases.

Doing that outside this framework, if ever possible, would expose any employee and the company to huge risks.

Regarding prediction, nothing is stored by default by Microsoft. For example, if you use a model to extract data from a document, we don't store this extracted data.

Hope that helps!

Was this reply helpful? Yes No
BennyS 8 on at

Like (0)

Report

So, just to summarize and recap:
In Microsoft OpenAI Services (e.g. generated responses in Virtual Agents), MSFT employees need the business approval when conducting abuse monitoring deep-dives?

Was this reply helpful? Yes No
Antrod Moderator on at

Like (0)

Report

Hi Benny,

I can't tell for sure the process for data retention regarding Power Virtual Agent. You should ask your question to this community: Power Virtual Agents Community - Power Platform Community (microsoft.com)

Was this reply helpful? Yes No