Data protection and AI processing

How BundleWright uses Amazon Bedrock, and what that means for GDPR and HIPAA.

UK GDPR and GDPR

  • AWS states that customers can use Amazon Bedrock in compliance with the GDPR. AWS acts as our processor under the AWS Data Processing Addendum, which forms part of the AWS Service Terms and includes the UK Addendum.
  • Case data stays in the United Kingdom. The model is called in AWS London (eu-west-2), on demand, with cross-region inference blocked by policy, so a request cannot be routed to another country.
  • Your documents and case context reach the model only as anonymised text. Names, identifiers, dates of birth and organisations are replaced with placeholders inside BundleWright first, and there is no setting that sends the original text of your documents.
  • Two short pieces of text you type are sent as written, because they are used before anonymisation can run. If you write your own chronology instructions, they are checked by an AI model on Amazon Bedrock before your case starts; we first remove what we can recognise by its format (email addresses, phone numbers, NHS numbers, labelled National Insurance, hospital and case numbers, labelled dates of birth, postcodes, street addresses, organisation names and long numbers), but names of people are not detected, so describe people by role. The text you give an “Other” document role (up to three words) is also sent as typed. The standard and concise instructions are never checked this way.
  • The AI service is zero data retention AI: AWS does not store your prompts or the model’s replies, does not share them with model providers, and does not use them to train any model (AWS Service Terms, section 50.3). We have also left Bedrock model invocation logging switched off, so no copy is kept in our AWS account either.
  • Deleting a case destroys its encryption key, which makes every stored copy and backup unreadable. Retention is chosen per case.
  • The Terms and Conditions you accept at registration name Amazon Web Services as our sub-processor (Schedule 1), and we give 30 days’ notice of any change.

Zero data retention AI

  • The AI service we use, Amazon Bedrock, is zero data retention AI: it does not keep your documents, prompts or outputs once processing is done (AWS Service Terms, section 50.3). It does not share them with model providers or use them to train any model.
  • We leave Bedrock model invocation logging switched off, so no copy is kept in our AWS account either. Zero retention here describes the AI service. It is separate from the retention choice you make for each case, which controls how long we keep your case before deleting it.
  • The precise facts behind it: the model runs in AWS London (eu-west-2) with no cross-region inference, your case data is encrypted with a separate key per case, and the two pieces of typed text described above are the only text sent as typed.

HIPAA

  • BundleWright is built to meet both GDPR and HIPAA. Amazon Bedrock is a HIPAA eligible AWS service.
  • HIPAA applies to US covered entities and their business associates. We offer a Business Associate Agreement (BAA) on request, and you should ask for it before you upload any US protected health information.
  • Compliance also depends on how each customer uses the service. You remain responsible as the controller and, if you are a US covered entity, as the covered entity.

Independent assurance AWS holds for Bedrock

  • SOC 1, SOC 2 and SOC 3 (Type 2) reports, available from AWS Artifact.
  • ISO 27001, 27017, 27018, 27701, 22301, 9001 and 20000, and CSA STAR Level 2.
  • Encryption in transit and at rest, and access control through AWS IAM. Our case data is further encrypted with a separate key per case (AWS KMS).
  • These are AWS’s certifications of its own service. BundleWright is not itself certified to any of these standards.

What to remember

  • Anonymisation is automated and measured, not perfect. Some identifiers may reach the model, which is why the model provider’s no-storage, no-training terms matter.
  • After the first draft, the same model reviews the chronology twice, always on anonymised text: once to shorten entries and flag any whose summary the source quote does not support, and once, acting as a case handler, to combine entries that clearly describe the same event. A combined entry keeps every page citation of the entries it replaces. Where the model is unsure, entries are left separate and marked “Needs human review” with a reason. Rewritten and combined text is not kept in its earlier form.
  • The chronology is an AI draft. Every entry shows its source page, and you must check it before relying on it.
  • Your own obligations as controller, including your lawful basis and what you tell the people in the records, remain yours. Our DPIA is available to review with your data protection adviser.

Check AWS’s statements at the source:

Legal documents · Back to BundleWright