Privacy model
Where your data goes.
Your prompts and documents stay in your own AWS cloud, in the EU region you choose, and are not sent to any AI model vendor. Here is, plainly, who can see what, how our access works and what we sign.
Who can see what
Four parties, two ways to run the model. Either way, your setup, your data and your keys are in your own cloud; the difference is where the model itself runs.
- Managed AWS model service
- Your account calls Amazon Bedrock, the managed AWS model service, in the region you choose. AWS runs the model.
- Dedicated GPU
- A dedicated GPU instance in your own cloud runs the model. Its version stays fixed until you approve a change.
Which one fits depends on the models your use case needs. We recommend one in the assessment; the decision is yours.
| Who | Managed AWS model service | Dedicated GPU |
|---|---|---|
| Your administratorsThe people you give access to your cloud | Everything, on your terms It is your account. Your administrators control who has access, the encryption keys and the logs. Who in your company may see what is your decision. | Everything, on your terms The same. The instance, the storage and the keys are all in your cloud. |
| Usomesos, through a role in your account | What our role allows We work through a role in your account. You can see and change its permissions, every session is time-limited and logged in your CloudTrail, and you can revoke it at any time. | What our role allows The same role, which also looks after the GPU instance and the model version. The version stays fixed until you approve a change. |
| AWSAmazon Web Services and its staff | Runs the model service Amazon Bedrock processes your requests, including the document passages that go with them, in the region you choose, and AWS does not use them to train models. Whether AWS keeps any of them for a while, for example to detect abuse, depends on the model and on a retention setting in your cloud. We set it with you, and tell you before you choose a model that requires it. | No technical means of access, by AWS’s own terms The model runs on a Nitro-based EC2 instance in your account. AWS commits in its Service Terms that its staff have no technical means to access your content on such instances or on the encrypted volumes attached to them. |
| AI model vendorsThe companies that make the models | Nothing is sent AWS runs each model in accounts of its own, which the model’s vendor cannot reach. The vendor does not see your prompts, answers or documents. | Nothing is sent The model’s open weights are copied into your cloud and run there. No vendor is involved when it answers. |
How our access works
We operate your setup, so we need a way in. It is built so that you stay in control of it.
A role you control
We reach your cloud only through a role in your AWS account. You decide what it may do, and you can read and change its permissions at any time.
Time-limited
Each session uses temporary credentials that expire on their own.
Logged in your CloudTrail
Every session is logged in your own CloudTrail, where your team can review it.
Revocable at any time
You can remove the role whenever you want, without asking us.
What AWS commits to
From the AWS Service Terms
On the dedicated-GPU setup, the model runs on an EC2 instance built on the AWS Nitro System. In its Service Terms, AWS commits that its staff have no technical means to read, copy, change or otherwise access your content on Nitro EC2 instances, or on the encrypted volumes attached to them.
Scope EC2 instances on the Nitro System and the encrypted EBS volumes attached to them. Other AWS services, Amazon Bedrock included, come with their own terms.
Read it in the AWS Service Terms, under “AWS Nitro System”What we sign
- Processing agreement
- A GDPR processing agreement under Art. 28 (in German, AV-Vertrag), covering everything we do in your cloud on your behalf.
- Secrecy agreement
- A separate agreement in which we commit to keep confidential anything we see while operating your setup.
- Help with your paperwork
- An AI notice for your users, a register of the model licences in use, and input for your data protection impact assessment (DPIA).
- Your contract with AWS
- Stays directly between you and AWS. As the account owner you are AWS’s customer yourself, and AWS’s own data processing addendum is already part of its Service Terms.
Our principles
No AI vendor in the data path.
Your prompts and documents are not sent to any AI model vendor, on either setup.
Your region.
Your setup is hosted in the AWS EU region you choose, for example Frankfurt.
Your keys.
You own the encryption keys, in your own cloud. Our access is in your hands too.
Bring your data protection officer
In the assessment we go through this page against your setup and answer their questions directly.
Book an assessment