Rubric
Contents — domains, guide and mocks

Regulatory compliance by design

CCAR-P 5.413 min read · checked 21 September 2026

Task statementEnsure compliance with regulations (e.g., GDPR, HIPAA, FedRAMP)

Compliance is a stack, not a checkbox

  1. Legal basis and agreementsDPA, BAA, government terms
  2. Deployment pathAnthropic API, cloud platform, gov offering
  3. Enabled featuresonly those eligible for the regime
  4. Data flowswhat is sent, stored, logged, and where
  5. Controls and oversightaccess, retention, review, audit trail
Each layer can break compliance on its own. A signed agreement does not help if your application logs patient data to an unapproved tool.

GDPR: principles that become design decisions

The EU General Data Protection Regulation applies to processing personal data. Article 5 sets the principles: lawfulness, fairness and transparency; purpose limitation; data minimisation; accuracy; storage limitation; integrity and confidentiality — and Article 5(2) makes the controller accountable for being able to demonstrate compliance. For an architect, each principle maps to a choice. Minimisation means sending the model only the fields the task needs. Storage limitation means deciding how long prompts, outputs and logs live. Accuracy matters because a model can state wrong facts about a person.

GDPR articleWhat it requires (paraphrased)Architecture consequence
Art. 5Six processing principles plus accountabilityMinimise fields sent; set retention; keep evidence of decisions
Art. 22Limits on solely automated decisions with legal or similarly significant effectsKeep meaningful human involvement; allow people to contest (5.3)
Art. 25Data protection by design and by defaultPrivacy controls built in from the first design, not bolted on
Art. 28Processors need sufficient guarantees and a binding contractA data processing agreement with the AI vendor; sub-processors authorised
Art. 35Impact assessment for likely high-risk processing, especially new technologiesRun a DPIA before launching high-risk AI processing

Article 22 deserves care because AI decisions touch it directly. It gives people the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. There are exceptions — where necessary for a contract, authorised by law, or based on explicit consent — but where they apply, the controller must put safeguards in place, including at least the right to obtain human intervention, to express a point of view and to contest the decision. Article 13 also requires meaningful information about the logic involved when such decision-making is used. In practice, a design with a genuine human decision-maker (5.3) is usually the simpler path.

HIPAA: business associates and eligible features

In the US, HIPAA governs protected health information (PHI) held by covered entities such as health plans and most providers. HHS defines a business associate as a person or organisation that performs functions on a covered entity’s behalf involving PHI — creating, receiving, maintaining or transmitting it — and notes that business associates are directly liable for complying with certain HIPAA provisions. A business associate agreement (BAA) must describe the permitted uses and disclosures of PHI and include safeguards and breach-reporting obligations.

Anthropic’s documentation describes a HIPAA readiness setting for API organisations: an admin enables it and executes a BAA, after which the API rejects features that are not HIPAA-eligible rather than letting them run. The eligible set includes the Messages API and several related features; the docs list others — including the Files API, the Message Batches API and code execution — as not eligible. The same page warns not to put PHI in JSON schema definitions used for structured outputs, because schemas are cached; patient details belong in message content. Anthropic’s privacy center adds that BAA customers face configuration requirements and limits on available features and integrations.

A hospital pipeline checked against HIPAA readiness

  • Passes: BAA executed with the vendor
  • Passes: Notes sent inline via the Messages API
  • Fails: Nightly bulk run via the Batches APInot HIPAA-eligible
  • Fails: Scans uploaded via the Files APIsend PDFs inline instead
  • Fails: Patient names in a JSON schema enumPHI belongs in messages
  • Check: App logs with PHI in a general toolyour own obligation
The design is sound until it reaches for convenient features. Each non-eligible feature needs an eligible alternative.

Retention: what zero data retention does and does not cover

Anthropic’s docs describe zero data retention (ZDR) as an arrangement where customer prompts and responses are not stored at rest after the API response is returned, apart from what is technically necessary or needed to comply with law or combat misuse. It is enabled per organisation through Anthropic’s sales team. Some features are not ZDR-eligible because they are stateful by nature — the docs list the Files API, batch processing and code execution among them — and using one of those is a choice to step outside ZDR for that data. The privacy center adds that ZDR applies to commercial API use and not to consumer plans.

FedRAMP: authorisation belongs to the cloud service

FedRAMP is the US government programme for authorising cloud services used by federal agencies. Anthropic’s Public Sector FAQ makes a point worth remembering: FedRAMP and DoD Impact Levels certify cloud services, and AI models are software components, not cloud services. So the architectural question is which hosting path carries the authorisation. The FAQ names Claude for Government as Anthropic’s FedRAMP High authorised product, lists Claude in Amazon Bedrock in AWS GovCloud as approved for FedRAMP High and DoD IL4/5 workloads, and Google Vertex AI as authorised at FedRAMP High. It also states that Claude Enterprise purchased through AWS Marketplace is not FedRAMP authorised.

Choosing a deployment path by regime

Which regime governs this workload?
  • US federal, FedRAMP High
    Authorised gov pathClaude for Government, Bedrock GovCloud, Vertex AI
  • US health data (PHI)
    BAA + eligible featuresHIPAA readiness; no stateful extras
  • EU personal data
    DPA, minimise, DPIAretention and residency checked
  • No regulated data
    Standard commercialstill set retention and logging
The regime narrows the path first; features and data flows are then checked within it.

Traps the wrong answers are built from

Tempting but wrongDo this instead
Saying the model itself is HIPAA or FedRAMP compliant.Compliance comes from the agreement, the authorised deployment path, the eligible features and your own data handling.
Signing a BAA and then using any convenient API feature.Use only HIPAA-eligible features; replace ineligible ones such as batch or file uploads with eligible alternatives.
Assuming zero data retention covers every feature and every plan.Check which features are ZDR-eligible; stateful features and consumer plans sit outside it.
Letting a model make significant decisions about EU residents with no human involved.Keep a meaningful human decision-maker or put Article 22 safeguards in place, and assess with a DPIA.
Forgetting your own application’s logs, caches and analytics tools.Map every data flow, including your own logging, and apply the same retention and access rules.

You should now be able to

  • Translate GDPR Articles 5, 22, 25, 28 and 35 into concrete design choices for a Claude system.
  • Explain HIPAA business associate obligations and what a BAA-covered Claude configuration requires.
  • Describe what zero data retention covers, and recognise features that fall outside it.
  • Choose a FedRAMP-appropriate deployment path, knowing that authorisation applies to cloud services rather than models.
  • Identify compliance gaps in an architecture across agreements, hosting, features, data flows and oversight.

Practice questions

Original questions written for this lesson, in the exam’s style. Answer first, then open the reasoning — every option is explained, including why the wrong ones are tempting.

  1. Question 1

    A hospital network has signed a BAA and enabled HIPAA readiness on its API organisation. Engineers want to process a backlog of 50,000 clinical notes overnight using the Message Batches API to save cost.

    What should the architect advise?

    1. AProceed; the BAA covers every feature of the API organisation.
    2. BProceed, but strip patient names from the notes before batching.
    3. CUse eligible Messages API calls run by the network’s own job queue.
    4. DMove the workload to a consumer Claude plan for the bulk run.
    Show answer and reasoning
    1. AIncorrect. The BAA does not make every feature eligible; the docs list batch processing as not HIPAA-eligible.
    2. BIncorrect. Removing names alone rarely de-identifies clinical notes, and it sidesteps rather than answers the eligibility question.
    3. CCorrect. This keeps PHI within HIPAA-eligible features while still processing the backlog overnight.
    4. DIncorrect. Consumer plans are outside these commercial arrangements entirely and would be the worst choice for PHI.

  2. Question 2

    A US federal agency built a successful pilot on a Claude Enterprise subscription purchased through AWS Marketplace. Production requires FedRAMP High authorisation.

    Which statement is most accurate?

    1. AClaude models are FedRAMP High authorised, so any subscription qualifies.
    2. BThe Marketplace offering is not FedRAMP authorised; move to an authorised path.
    3. CFedRAMP applies only to storage, so the model call can stay where it is.
    4. DAdding zero data retention to the subscription satisfies FedRAMP High.
    Show answer and reasoning
    1. AIncorrect. Anthropic’s FAQ notes that FedRAMP certifies cloud services, not models.
    2. BCorrect. The FAQ says so explicitly and names Claude for Government, Bedrock GovCloud and Vertex AI paths.
    3. CIncorrect. The model service is part of the processing path and must sit within an authorised boundary.
    4. DIncorrect. Retention settings are useful but are not a substitute for an authorised cloud service.

  3. Question 3

    A French bank plans to use Claude to decide automatically whether to grant small personal loans, with no human involved. Legal asks what GDPR requires.

    Which two design responses best address GDPR? (Select 2.)

    1. APut a credit officer in the decision, with the right to contest built in.
    2. BCarry out a DPIA before launch, documenting risks and mitigations.
    3. CAdd a notice that decisions are final and not subject to review.
    4. DSend the full customer file to the model so it has maximum context.
    5. ERely on the vendor’s certifications instead of a processing agreement.
    Show answer and reasoning
    1. ACorrect. Article 22 restricts solely automated significant decisions; meaningful human involvement and contestability address it.
    2. BCorrect. Article 35 calls for an impact assessment for likely high-risk processing, especially with new technologies.
    3. CIncorrect. This contradicts the Article 22 safeguards, which include the right to contest.
    4. DIncorrect. Data minimisation (Article 5) points the other way: send only what the decision needs.
    5. EIncorrect. Article 28 requires a binding contract with the processor; certifications do not replace it.

Practise every question in Domain 5 →

Sources

Drafted with AI assistance and checked against the sources above; expert review is in progress. Spotted an error? Tell us and it gets fixed, dated and listed on how this is written.