Article · Architecture

Keep your prompts in Sweden: Claude on AWS Bedrock with zero data retention

"Our data stays in Europe" is a configuration, not a slogan. Here is how I set up Claude through AWS Bedrock so nothing is stored and nothing leaves the region, and the two places where "EU" quietly stops meaning EU.

Half the companies I talk to have a rule that says "no customer data to US AI services". The other half have the same rule and break it every day, because the approved tool is hosted somewhere nobody checked. This article is the setup I use when the promise has to be "prompts never leave the region", and the verification that proves it.

Why it matters

GDPR Chapter V. Sending personal data to a model hosted outside the EU needs a transfer safeguard: an adequacy decision, standard contractual clauses, a transfer impact assessment. Keeping the processing inside the region removes the question. That is worth a lot of lawyer hours.

The setup

Region first. AWS Bedrock runs Claude in EU regions. Stockholm (eu-north-1) is the obvious pick for a Swedish promise, Frankfurt (eu-central-1) for a European one. Model availability differs per region and changes over time, so check the current availability table for the exact model you want before you promise anything. Write the region into the contract, not into a slide.

Zero data retention. Bedrock does not store your prompts or completions after the request, and does not use them to train models. That is the default, not a setting you have to find. Model invocation logging is opt-in. If you turn it on, it lands in your own S3 bucket or CloudWatch, in your region. Keep it off, or keep it in-region and behind your own retention policy.

Least privilege. One IAM role per application with only the invoke permission on the specific model. No admin keys "to be safe". A VPC endpoint for Bedrock so the traffic never touches the public internet.

Point your tools at it. Claude Code, for example, runs against Bedrock with an environment flag and a region setting. The developer experience is identical. The data path is not.

The two traps

Cross-region inference profiles. Bedrock offers "geo" inference profiles (the ones prefixed eu.) that can route a request to another region inside the EU for capacity. That is fine if your promise is "Europe". It is not fine if your promise is "Sweden". Use the single-region model ID, not the eu. profile, when the region matters.

The rest of the chain. The model is one hop. Your logs, your backups, your observability vendor and the SaaS tool that calls the model are the other hops. I have seen a perfectly configured Bedrock setup feed a US-hosted logging tool with every prompt in plaintext. Map the whole chain, not the model.

The proof

Turn on CloudTrail and query the InvokeModel events. Every one should show the region you promised. Screenshot that. It is the only slide anyone actually needs to see.

"Our data stays in Europe" is a configuration, not a slogan. Show the CloudTrail line.

Do this week

  • List every AI call your systems make and the region each one lands in. Include logging and backups.
  • For anything that must stay in one country, replace the geo inference profile with the single-region model ID.
  • Run the CloudTrail query and keep the result with your compliance register.

Want the same thing done for you?

Join the launch list and you get your personal start link the day it opens.