Choose and Contract Claude Deployment Routes
Decide, per use-case class, which route carries Claude into production (Anthropic API, Claude Platform on AWS, Bedrock, Google Cloud, Microsoft Foundry, plus the Microsoft 365 Copilot surface); record who operates, bills and processes; size the diligence to the processor; check residency, zero data retention and HIPAA per feature; sign a contracting matrix row; and run Agent SDK or hosted agents with the controls the product does not supply. Provider-specific and dated.
- Version
- 0.1.0
- Owner
- DEIXEI
- Last reviewed
- Reading time
- 50 min
For Cto · Ai Architect · Platform Engineer · Security Engineer · Procurement · Dpo · Risk And Assurance · Internal Audit
Sources and details
- Section
- Build the platform
- Source works
- Book 3 · Research Series
- Source chapters
- 4, 7, 9, 10
- Source appendices
- 04:A, 04:C, 04:D, 04:H, 04:I, 04:J
- Refresh trigger
- vendor release
There is no single answer to “how do we run Claude?”. There are several commercial routes to the same model family, and each one assigns the operator, billing, data-processor and contracting roles differently. Each also offers a different residency posture, feature set and retention arrangement. Book 3 Ch 4 puts the decision in one line: choose a route per class of workload, not per institution. Most regulated firms will end up running two or three routes. This page turns that decision into a sequence you can follow, a matrix row both procurement and the second line can sign, and the run-time controls an embedded Claude agent needs whichever route carries it.
This page is provider-specific and dated. Every product statement carries its date. Three dates recur:
- Book 3 state (May to August 2026). Appendix A was drafted on 14 May 2026, and its Foundry entries were revised on 3 August 2026 after Foundry went GA on 29 June 2026. Chapters 4 and 9 are dated 3 August 2026.
- Research Paper 3 state (10 November 2026). Claude Code route facts checked against vendor documentation.
- Vendor documentation read for this page (15 November 2026). Anthropic’s pages on API data retention, data residency, Foundry, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Agent SDK hosting, session storage and permissions. Wherever these pages contradict the books, this page reports both and follows the vendor text, flagged as a correction.
Anything not checked against vendor documentation carries a [FACT-CHECK: …] marker. Documentation is not the contract: the signed order form, DPA and addenda govern. Bank A and Asset Manager X are composite institutions from Book 3, not real firms. Statements about what DORA, GDPR, the Swiss FADP or the EU AI Act require, and every regulatory date, are on the regulatory map and the regulatory calendar. This page gives design consequences only. It is not legal advice.
Decisions this page supports
| Decision | Decided by | Recorded in | Section |
|---|---|---|---|
| Which route serves each use-case class | Head of platform with lead architect; co-signed by procurement and second line | Contracting matrix row | Decision sequence |
| Who is the data processor, and therefore whose diligence | Procurement lead with legal and second line | Contracting matrix block 2; Register of Information row | Who operates, bills and processes |
| Whether the route meets the class’s residency rule | DPO with data office | Matrix block 3; transfer record | Residency |
| Which API features an agent may use without leaving ZDR or HIPAA cover | Wrangler with privacy office | Agent envelope (Means, Memory); ZDR configuration record | ZDR and HIPAA |
| Whether a hosted agent service is an outsourcing arrangement | Third-party risk owner with second line | Outsourcing register; matrix | Hosted agents |
| How a self-hosted SDK agent is isolated, bounded and evidenced | Platform engineering owner of the SDK runtime; SOC | Runtime profile; detection rules | Agent SDK |
| When a route decision is reopened | Route-portfolio owner | Route-decision register; quarterly minutes | Quarterly review |
The owners are drawn from Book 3 Ch 4 PR-2 (responsible roles) and Ch 9 PR-1. Align them with your own RACI.
Treat the route as a portfolio decision
Book 3 Ch 4 §5 names four factors. No route is best on all four.
| Factor | The question | What moves it |
|---|---|---|
| Data residency | Where is the prompt processed, and where is data stored at rest? | The route’s inference-location and storage-location controls (two separate settings on most routes) |
| Contractual posture | Who is the data processor, and is that processor a designated critical ICT third-party provider (CTPP)? | The route’s operator, not its biller |
| Region and model availability | Does this route serve the model version you need in the location you need, today? | Per-route model catalogues and regional endpoints; aliases resolve differently per route |
| Feature parity | Are the features the workload depends on (Files, Batch, Agent Skills, Compliance API, MCP connector, code execution) available on this route and this hosting option? | Per-route and per-hosting-option exclusion lists |
Three rules follow (Book 3 Ch 4 §3, §8):
- One matrix row per use-case class. “Engineering productivity”, “employee assistance” and “regulated workflow agents” are separate rows, and they can land on separate routes.
- The row expires. Re-derive block 2 (processor, biller, DPA) from current vendor text at every quarterly review, at every model release and at every vendor terms change. Do not copy it forward. In the Bank A composite, Foundry’s GA statement moved the processor from Microsoft to Anthropic. The route on two rows stayed the same, but their TPRM target, DPA reference and register entry all changed (Ch 4 §6).
- Pin model versions on every route. Use exact deployment, inference-profile or model IDs in configuration under change control. An alias is a laptop convenience, not a production control (Ch 4 §5; see model gateway: pin models).
Book 3’s reference architecture is Azure-first. It uses Foundry plus the direct Anthropic API behind an Azure API Management gateway, with Bedrock as the route for the residency-strict slice (Ch 4 §3, §8). AWS-first firms swap in Bedrock plus the direct API behind Amazon API Gateway. GCP-first firms swap in Google Cloud plus the direct API behind Apigee (Ch 4 §4.8). The method is the same in all three. That is the author’s position for an Azure estate, not a market finding.
Record who operates, bills and processes
The most expensive mistake in Book 3 Ch 4 §7 is to assume that the marketplace you buy through is the processor of your data. The table below separates the four roles. It combines Book 3 Appendix A §2 (Foundry column revised 3 August 2026) with the vendor documentation read on 15 November 2026.
| Route | Operator of inference | Biller | Data processor | Contracting entity (EU customers, per App A) |
|---|---|---|---|---|
Anthropic API (api.anthropic.com) | Anthropic | Anthropic | Anthropic PBC | Anthropic PBC, San Francisco; no EU establishment named |
| Claude Platform on AWS | Anthropic, on Anthropic-managed infrastructure | AWS Marketplace (Claude Consumption Units, billed in arrears on the AWS bill) | Anthropic for inference inputs and outputs; AWS processes billing and identity metadata (vendor docs, 15 Nov 2026) | AWS Marketplace terms layered over Anthropic Commercial Terms and DPA: two layers |
| Amazon Bedrock | AWS | AWS | AWS (App A: AWS EMEA SARL) | AWS EMEA SARL, Luxembourg; no separate Anthropic agreement |
| Claude on Google Cloud, Google managed (Google Cloud’s Agent Platform, formerly Vertex AI; Ch 4 §4.5, Paper 3 §4.3) | Google (App A: Google Ireland Ltd) | Google Ireland Ltd | ||
| Claude Models on Google Cloud, Anthropic managed (not in Book 3) | Anthropic | Google Cloud Marketplace | Anthropic; some supporting services, including Anthropic’s trust and safety systems, process data outside Google Cloud (vendor docs, 15 Nov 2026) | [FACT-CHECK: contracting layers for the Anthropic-managed Google Cloud offering] |
| Microsoft Foundry | Anthropic, on Azure infrastructure (“hosted on Azure”) or on Anthropic infrastructure (“hosted on Anthropic”) | Microsoft (Azure Marketplace; CCUs on the Azure bill) | Anthropic: Anthropic’s Foundry documentation says it “acts as an independent processor for Microsoft” (15 Nov 2026) | Microsoft (App A: Microsoft Ireland Operations Ltd under MOSPA). MOSPA covers the commercial wrapper only, not the Anthropic processing leg |
| Microsoft 365 Copilot (Anthropic models): arrives through an admin toggle, not procurement | Anthropic | Microsoft (M365 licensing) | Anthropic listed as a Microsoft sub-processor for select Copilot experiences (App A §3.2 sidebar, verified by the author 14 May 2026) | Microsoft 365 terms, not Foundry terms |
Read the table with these points in mind:
- On four of the seven rows, the biller is not the processor. On Claude Platform on AWS, the Anthropic-managed Google offering and Foundry, a cloud provider bills and authenticates while Anthropic processes. Book 3 Ch 4 §7 gives the test. If your TPRM register names only the cloud provider for one of these routes, or the DPA column says “MOSPA” on a Foundry row, the row is wrong. Re-derive the processor from the vendor’s current published text. Do not take it from the invoice, the login screen or last quarter’s diagram.
- Foundry hosting option is a per-deployment choice, and it changes the data path. For “hosted on Azure”, Anthropic’s documentation says prompts and completions stay within Azure, and only usage metadata and content flagged by Anthropic’s safety systems leave for Anthropic. “Hosted on Anthropic” runs on Anthropic infrastructure outside Azure (vendor docs, 15 Nov 2026; consistent with Research Paper 3 §4.1). In both cases Anthropic is the processor. Record the hosting option on the matrix row.
- Claude Platform on AWS and the Anthropic-managed Google offering each provision a new Anthropic organisation. It is separate from any existing Anthropic or Claude Enterprise organisation. Keys, workspaces, console settings and ZDR enablement do not carry over (vendor docs, 15 Nov 2026; Ch 4 §4.4). Add each organisation to the identity inventory and the access review.
- Draw the data-flow diagram for every route with the second line, the DPO and TPRM in one workshop. Date it and include it in the next supervisory pack (Ch 4 §7).
Source difference: who processes on Foundry
The books disagree, and this page follows the newer sources.
| Source | Says the Foundry processor is |
|---|---|
| Book 3 App A (revised 3 Aug 2026), Ch 4 §4.2 and §7 | Anthropic, per the GA statement of 29 June 2026, not qualified by hosting option |
| Book 3 App H §3.2 (per-route summary and second observation) | Treats Foundry as “cloud-processed”, with Anthropic as a sub-processor under MOSPA and a Microsoft CTPP lead-overseer relationship |
| Book 3 Glossary (“Azure AI Foundry”) and Ch 10 §5.2 | Microsoft is the data processor, or the cloud provider is processor (preview-era wording) |
| Research Paper 3 §4.1 (10 Nov 2026) | Anthropic is the seller and operator and acts as an independent data processor on both hosting options |
| Anthropic Foundry and data-retention documentation (15 Nov 2026) | Anthropic is the data processor; “independent processor for Microsoft” |
This page follows App A, Ch 4, Paper 3 and the vendor text: Anthropic is the processor on Foundry. App H’s Foundry row and the Glossary entry predate the GA revision. Do not use them to fill block 2 of the matrix. Apply the same check to exit sequencing: on Foundry, exit planning has to cover the Anthropic processing relationship as well as the Microsoft commercial one.
Size the diligence to the processor
Per Book 3 App A §1 and App I §3, Anthropic was not on the ESAs’ list of designated CTPPs, while AWS, Microsoft Azure and Google Cloud were. The list date, the relevant DORA articles and any later designation round are on the regulatory map and regulatory calendar. [FACT-CHECK: re-confirm Anthropic’s designation status against the current ESAs CTPP list before relying on this row.]
For the institution, the design consequence is that the diligence load follows the processor, not the biller:
| Route | Processor is a designated CTPP? | What third-party risk must do |
|---|---|---|
| Anthropic API | No | The full third-party diligence, register entry, contractual-clause check and exit strategy for Anthropic is the institution’s own work. No lead-overseer output exists to rely on (App I §3) |
| Claude Platform on AWS | No for the processing layer. AWS’s designation covers the Marketplace procurement layer only (App H §3.2) | As for the Anthropic API. Draw the AWS/Anthropic split explicitly in the Register of Information |
| Anthropic-managed Google Cloud | No (Anthropic) | As for the Anthropic API [FACT-CHECK: how this offering is papered] |
| Microsoft Foundry | No. The processor is Anthropic, even though Microsoft is designated (App A §3.2) | As for the Anthropic API, plus the Microsoft wrapper. Bank A’s July re-run added a second Anthropic row to the register and re-ran second-line sign-off for the two Foundry rows (Ch 4 §6, composite) |
| Bedrock | Yes (AWS) | Existing AWS TPRM workstream and register entry extend to the route; the designated-provider regime applies |
| Google-managed Google Cloud | Yes (Google) | Existing Google TPRM workstream extends |
| M365 Copilot (Anthropic) | Anthropic as a Microsoft sub-processor | A separate register row and owner from Foundry (App A §3.2 sidebar). See the Copilot surface |
Three supporting facts belong in the TPRM file:
- Mandatory contract clauses apply on any route that supports a critical or important function (CIF), whatever the CTPP status (Ch 4 §5). The CIF flag is decided at intake and kept separate from the risk tier (App I §3). Article references are on the regulatory map.
- Sub-processor notice differs by route. Anthropic publishes its sub-processor changes at
trust.anthropic.com/updateswith 15 days’ notice. The AWS list carries 30 days’ notice (App A §5). [FACT-CHECK: current notice periods in the executed DPAs.] Subscribe the TPRM mailbox to each list the matrix depends on. - WorkOS is a sub-processor on the Anthropic SSO path. Console and Claude Enterprise SAML SSO is brokered through WorkOS, so IdP metadata and certificate exchange pass through it rather than going directly between your IdP and Anthropic (App A §3.1; Glossary). Record WorkOS in the sub-processor register for the Anthropic API and Claude Platform on AWS rows. An Anthropic exit ends the WorkOS relationship without touching your own IdP (App H §3.2). [FACT-CHECK: whether WorkOS still mediates SSO, and for which plans.]
Exit cleavage points (contracting entity, notice, termination for cause, data return, sub-processor exit, suspension versus termination, CTPP implication, register-row implication) are covered in vendor concentration and exit.
Check residency route by route
Residency has two independent parts: where inference runs and where data is stored at rest (vendor docs, data residency, 15 Nov 2026; Paper 9 §6.2). Record both per deployment.
| Route | Book 3 state (App A §2, Ch 4 §4; May–Aug 2026) | Vendor documentation, 15 Nov 2026 | Consequence |
|---|---|---|---|
| Anthropic API | Workspace us only; inference us or global; no EU geo | Same. inference_geo accepts only "global" (default) and "us"; workspace geo is "us" only and cannot be changed after creation. Workspace settings allowed_inference_geos and default_inference_geo restrict and default the geo | No EU inference or storage. The EU-to-US transfer happens on every call and is carried by transfer instruments, not avoided |
| Claude Platform on AWS | In-region EU inference in Dublin, London, Frankfurt, Milan, Zurich, Paris and Stockholm | Correction. A workspace is bound to one AWS region, but that region “does not pin where model inference runs”. Inference geos are US or Global only. Workspace geo is not configurable. Workspaces created before 18 September 2026 may have requests processed and data stored outside AWS | Do not treat this route as an EU-resident route. The App A §4 step 3 outcome no longer holds (see the decision sequence) |
| Anthropic-managed Google Cloud | Not covered | Every workspace allows all inference geos and defaults to global; workspace geo us; set inference_geo: "us" per request for US-only | No EU inference or storage |
| Microsoft Foundry | US data zone only at GA; no EU data zone | Global Standard and US Data Zone Standard. The data zone is available only on the “hosted on Azure” option; inference_geo does not apply | No EU data zone. Paper 3 §4.1 notes data at rest stays in the chosen Azure geography while processing is global. Whether that satisfies your classification rules is your decision |
| Bedrock | In-region inference in eu-central-1, eu-west-1, eu-south-1, eu-north-1; AWS processor | Anthropic documents that on Bedrock the inference region is set by the endpoint or inference profile; Paper 3 §4.3 records eu.-prefixed cross-region profiles | The cleanest EU in-region option in the sources [FACT-CHECK: Claude model availability per EU region and the EU cross-region profile scope, from AWS documentation] |
| Google-managed Google Cloud | europe-west1 and the eu multi-region | Inference region set by endpoint URL (vendor docs) | EU in-region available [FACT-CHECK: per-model availability on europe-west1 and eu; a 404 usually means the model is served only on global or a multi-region (Ch 4 §4.5)] |
Regional premiums. As of 15 November 2026, US-only inference (inference_geo: "us") costs 1.1 times the standard rate for Claude 4.6 and later models. This applies on the Anthropic API, Claude Platform on AWS and the Anthropic-managed Google offering, and the same multiplier applies to Foundry’s US Data Zone deployments (vendor docs). Book 3 reports a 10% premium on Bedrock regional endpoints over global (from Claude 4.5) and on Google’s europe-west1 or eu locations (App A §2) [FACT-CHECK: current Bedrock and Google regional pricing]. Put the premium in the matrix row so finance sees the cost of residency rather than discovering it.
Treat the absence of an EU sovereign offering as a design constraint, not a dependency. Book 3 Ch 4 §4.7 compares the US government package (a named authorisation level, a written inference boundary, direct contracting) with the lack of any EU equivalent as of July 2026. The author’s advice is to design around the gap rather than put “awaiting EU sovereignty” on a roadmap as though it had a date. The author also notes that billing follows the assurance boundary. Where a route’s billing is most convenient, ask what assurance was traded away to get it. [FACT-CHECK: whether Anthropic has since published an EU sovereign or EU-resident offering.]
Record the transfer basis for every Anthropic-processed route
Zero data retention reduces copies. It does not legitimise a transfer (Ch 10 §5.2). Each Anthropic-processed row needs a transfer record, co-owned by the DPO and the procurement lead (App I §6.1):
- Check DPF certification. Verify Anthropic’s EU–US Data Privacy Framework status on the official DPF list, and re-verify at least annually. Track the status of legal challenges to the DPF (App I §6.1; Ch 10 §5.2).
- Fallback: SCCs plus a TIA. Where the DPF is unavailable or not relied on, use Standard Contractual Clauses Module 2 (controller to processor), which the Anthropic DPA carries with UK and Swiss adapters (App A §3.1), plus a documented Transfer Impact Assessment on file. Book 3’s composite firms run SCCs plus a TIA on every direct-API row (Ch 4 §6; Ch 9 §8).
- Swiss entities. Verify Anthropic’s Swiss–US DPF status and record it separately (App I §6.1).
- Record per row: mechanism relied on, verification date, TIA reference, owner and next check date.
The legal instruments, their dates and their litigation status are on the regulatory map. The data classes and lanes page decides which classes may take a cross-border lane at all.
Check zero data retention and HIPAA per feature, not per route
“ZDR applies” is a claim about one organisation, one route and one set of features. It is never a claim about a vendor (data classes: ZDR as recorded configuration).
ZDR by route (vendor docs, 15 Nov 2026)
| Route | ZDR |
|---|---|
| Anthropic API | Available on request, enabled per organisation by the account team. Enablement does not extend to other organisations under the same account. ZDR organisations cannot call the API from a browser (CORS requests are rejected) |
| Claude Code | Claude Code with Commercial-organisation API keys is covered by the API arrangement. Claude Code on Claude Enterprise has a separate ZDR offering for qualified accounts. If Claude Code metrics logging is on, productivity data is exempt from ZDR |
| Claude Platform on AWS | Available on request; otherwise the first-party retention policy applies |
| Anthropic-managed Google Cloud | Available on request. A ZDR arrangement on another Anthropic organisation does not carry over |
| Foundry, Bedrock, Google-managed Google Cloud | Anthropic’s ZDR does not carry. The cloud provider’s retention terms and settings govern (Ch 10 §5.2; App A §3). Paper 3 §4.1 found no ZDR option for Foundry on the Microsoft pages |
Features that leave ZDR
Book 3 (Ch 4 §4.1, Ch 10 §5.1, Glossary) lists the surfaces outside ZDR as Console, Workbench, Managed Agents, the Files API, the Batch API, Code Execution, Agent Skills, the MCP connector and Claude for Excel. Anthropic’s eligibility table on 15 November 2026 confirms that list and makes it more precise:
| Not ZDR-eligible (vendor table, 15 Nov 2026) | Endpoint or parameter | Retention note from the table |
|---|---|---|
| Files API | /v1/files | Retained until deleted or until a configured expiration |
| Batch processing | /v1/messages/batches | 29-day retention |
| MCP connector | /v1/messages with mcp_servers | Standard policy |
| MCP tunnels (research preview) | /v1/tunnels | Not eligible |
| Code execution tool, and programmatic tool calling built on it | /v1/messages with code_execution | Container data up to 30 days |
| Agent Skills | /v1/skills, /v1/messages with skills | Standard policy |
| Message threads | /v1/messages with thread | Not eligible |
| Claude Managed Agents | /v1/agents, /v1/sessions, /v1/environments | Not eligible except self-hosted Managed Agents |
| User profiles | /v1/user_profiles | Not eligible |
| Console (incl. playground), Claude for Excel, Teams/Enterprise product interfaces (apart from the separate Claude Code offering) | — | Not covered |
Four further rules from the vendor page:
- Using a non-eligible feature is not blocked under ZDR. The API accepts the call, and the feature’s own retention applies to that data. Only HIPAA readiness blocks non-eligible features, with a
400; the exception is client-side tools the table marks as not blocked, which are accepted but stay outside HIPAA cover. Under ZDR the control therefore has to sit with you, at the gateway and in the envelope. - Certain model families require 30-day retention. Anthropic calls them “Covered Models”, and they are not available under ZDR unless Anthropic expressly authorises it. Book 3 Ch 10 §5.1 makes the same point about one model. The retention posture limits model choice, so put it in the envelope.
- Flagged content and legal holds. Even under ZDR, Anthropic may retain data where the law requires it, and may keep inputs and outputs of flagged sessions for up to two years. Name both in the DPIA and the records schedule (Ch 10 §5.1).
- Book 3’s “computer use” note has aged. Ch 9 §4.2 calls computer use a beta capability that is not HIPAA-eligible. The 15 November 2026 table lists the computer-use tool as ZDR-eligible and HIPAA-eligible because it is a client-side tool. Book 3’s operational advice still stands: VM isolation, snapshot rollback, and admission only after an Agent Control Board review.
Controls that make ZDR checkable
| Control | Where | Owner |
|---|---|---|
The gateway refuses non-eligible features on ZDR-dependent routes. It strips or blocks mcp_servers, /v1/files, /v1/messages/batches, code_execution, skills and thread on requests whose envelope states a ZDR posture | Model gateway | Platform |
Envelope flag. persistent_storage: files_api_on is a deliberate non-ZDR choice. The registry flags the envelope for second-line review and records the posture (App D §2.3). Do the same for a Batch or MCP-connector dependency | Bounded agent | Wrangler; second line |
Pin the authenticated organisation. Coverage follows the organisation, not the device, so set forceLoginMethod and forceLoginOrgUUID on workstations (Ch 10 §5.1; Paper 3 §4.2) | Coding-agent workstation | Endpoint engineering |
| ZDR configuration record per deployment: route, organisation ID, enablement evidence, surfaces in and out, carve-outs, verification test, last verified date | Data classes and lanes | Privacy office |
The MCP connector: wire format only, broker decides
The Messages API mcp_servers parameter lets an embedded agent bind MCP tools directly. That has two consequences (Ch 7 §4.1; Ch 9 §4.3; App C §5 item 17):
- It takes the call out of ZDR. The vendor table (15 Nov 2026) confirms this. A ZDR-dependent workload either routes tool calls through the SDK’s own loop to registered servers, or records the exception in the envelope and the MCP register entry.
- It must not become a parallel tool surface. Embedded agents may name only servers from the connector broker’s signed manifest of approved servers. The runtime reads that manifest at instantiation, just as Claude Code reads
managed-mcp.json, and refuses any server that is not also on the envelope’s Means list. Themcp_serversparameter is the wire format. The MCP and connector broker is the authority.
MCP parity differs by route. Book 3 Ch 7 §4.1 (verified May 2026) reports: same-day parity on Claude Platform on AWS; Bedrock lagging beta features by about one release; Vertex with MCP tool search off by default; and Foundry carrying MCP through a separate Foundry MCP Server in public preview. Claude Platform on AWS now reaches public-internet MCP servers only, because MCP tunnels are not supported there (vendor docs, 15 Nov 2026). [FACT-CHECK: current MCP support per route.]
HIPAA by route
Book 3 states that HIPAA-ready cover is available only on the direct Anthropic API (App A §2; Ch 4 §4.1; Glossary). The vendor documentation of 15 November 2026 differs:
- Claude API: HIPAA readiness with a signed BAA and a HIPAA-enabled organisation, for eligible features only. Once enabled it is permanent, enforced per organisation, and the API returns
400for non-eligible features. Run HIPAA and general-purpose work in separate organisations. - Foundry: eligible on both hosting options, with a BAA with Anthropic plus the HIPAA configuration enabled on the Azure subscription.
- Claude Platform on AWS: not available. The AWS-specific page points HIPAA, FedRAMP High, IL4 and IL5 workloads to Bedrock.
- Bedrock and Google-managed Google Cloud: outside Anthropic’s programme. Refer to the provider’s compliance documentation.
- Claude Code: not covered by HIPAA readiness. This contradicts Ch 10 §5.2’s statement that a BAA extends to Claude Code where ZDR is active. Follow the vendor text and confirm in your BAA.
Walk the seven-step route decision
Book 3 App A §4 gives a walkable sequence: start at step 1 with the use case in hand, and the first fit is the recommendation. Below is the sequence with each output recorded in the contracting matrix, annotated where the 15 November 2026 vendor documentation changes the outcome. The annotations are this page’s correction. They are not in the book.
- Classify the workload. Record its AI Act class (prohibited / high-risk under Annex III / transparency / minimal), CIF yes or no, data class (the book’s Tier 0–4, or the C1–C4 classes on data classes and lanes), and personal and special-category data flags. If the use is prohibited, stop. Classification rules are on risk tiers and the regulatory map.
- Is in-EU inference mandatory? It is mandatory for high-risk workloads on EU residents, for CIFs, and wherever firm policy or the supervisor requires it (App A names BaFin, CSSF, FINMA, ACPR, DNB and CBI guidance as possible sources). Yes → step 3. No → step 5.
- Mandatory residency: which processor do you prefer?
- Book 3: Anthropic as processor → Claude Platform on AWS in a named EU region, accepting no HIPAA and the full non-CTPP diligence.
- Correction (15 Nov 2026): the vendor documentation offers only US or Global inference on Claude Platform on AWS and says the workspace region does not pin inference. No Anthropic-processed route currently offers EU in-region inference. If residency is mandatory, the answer is a cloud-processed route → step 4. Book 3 ties the Anthropic preference to Anthropic-native features (Compliance API, Activity Feed, ISO/IEC 42001 lineage, day-zero parity). If those features are essential, the workload has to be re-scoped, for example by tokenising so the residency rule no longer binds, or the residency requirement has to be formally relaxed by its owner. Neither is a route choice.
- Mandatory residency, cloud-processed route: which is your primary cloud?
- AWS → Bedrock in an EU region. Accept: no Compliance API, no Agent Skills, no Claude for Excel (App A §3.3); WebSearch unavailable on Bedrock (Ch 4 §4.3; Paper 3 §4.3).
- GCP → Google-managed Google Cloud on
europe-west1oreu. Accept the regional premium, no Compliance API, and verify each model on the chosen endpoint. - Azure → not Foundry for this workload. There is no EU data zone, and the processor is Anthropic, not Microsoft, so Foundry delivers neither residency nor a cloud-provider processor. Use Bedrock (Book 3’s choice), and defer Foundry for this class until an EU data zone is published. [FACT-CHECK: whether Foundry has since published an EU data zone.]
- Residency flexible: do you need MACC drawdown, with a residency-flexible workload (Book 3’s Tier 0–1, no customer personal data, no high-risk decisioning)?
- Yes → Foundry. Record the processor as Anthropic, record the hosting option, use Entra ID, document the EU data-zone gap as a watch item with a named owner and a quarterly date. Vendor docs (15 Nov 2026) describe Foundry billing as CCUs through the Azure Marketplace. Paper 3 §4.1 says spend counts against a Microsoft Azure Consumption Commitment and that Cloud Solution Provider subscriptions are not supported. [FACT-CHECK: MACC eligibility in your current SOW.]
- No → step 6.
- Does the workload process PHI that needs a BAA?
- Book 3: yes → direct Anthropic API only.
- Correction (15 Nov 2026): the Claude API or Foundry, each with an Anthropic BAA; or Bedrock under AWS’s own programme, per the Claude Platform on AWS page. Not Claude Platform on AWS. Not Claude Code. No → step 7.
- Does it depend on a feature available only on Anthropic-operated surfaces (Claude for Excel, Agent Skills, Compliance API, Activity Feed and similar)?
- Yes, residency flexible → Anthropic API (or Claude Platform on AWS where AWS Marketplace procurement is preferred: vendor docs list Agent Skills, code execution, Batch, the Files API, the Compliance API and Managed Agents there, but no HIPAA). Workspace
us; transfer record per the section above. - Yes, residency mandatory → Book 3 sends you back to step 3. After the correction there is no route that does both, so escalate the conflict to the use-case owner and the second line.
- No → choose between the Anthropic API (fastest feature parity) and Foundry (MACC, Entra ID) on procurement preference. On Foundry, check the exclusion lists: no Admin API, Advisor tool, Managed Agents, Compliance API, Models API, Message Batches API, server-side fallback or user-profile endpoints. “Hosted on Azure” additionally excludes code execution, Agent Skills, programmatic tool calling, the Files API and newer web search/fetch tool versions, and returns
400for them (vendor docs, 15 Nov 2026).
- Yes, residency flexible → Anthropic API (or Claude Platform on AWS where AWS Marketplace procurement is preferred: vendor docs list Agent Skills, code execution, Batch, the Files API, the Compliance API and Managed Agents there, but no HIPAA). Workspace
Output: one route per use-case class, written into the contracting matrix, signed by procurement and the second line. The same row feeds the Register of Information submission (App A §4).
The Azure-first shortcut (a second layer, not a replacement)
Book 3 Ch 4 PR-1 gives a shorter top-down tree for an Azure estate. Use it after the seven steps as a consistency check, not instead of them:
| Order | If… | Book 3 route | Note as of 15 Nov 2026 |
|---|---|---|---|
| 1 | Tier 0–1, internal, no customer data, no CIF adjacency | Foundry with Entra ID (DefaultAzureCredential, Azure AI User or Cognitive Services User; the vendor page now names Foundry User or Cognitive Services User) | Processor is Anthropic; EU data-zone gap is a watch item |
| 2 | Depends on Anthropic-native features | Anthropic API behind APIM; SCCs plus TIA; ZDR on the Messages API | Check each feature against the ZDR table above |
| 3 | High-risk or CIF and in-region EU inference non-negotiable | Bedrock eu-central-1 or eu-west-1 | Add a “revisit when an EU commitment is published” trigger |
| 4 | Anthropic as processor, EU residency, AWS Marketplace preferred | Claude Platform on AWS | Does not hold for residency as of 15 Nov 2026 (see step 3) |
| 5 | GCP primary, or the clearest documented EU region story needed | Google-managed Google Cloud on europe-west1 or eu | Verify per model |
Ch 9 PR-1 applies the same split to API agents. Tier 0–1 internal agents go to Foundry. Tier 2 and above needing Anthropic-native features go to the direct API behind APIM. Tier 3 and above with non-negotiable EU inference go to Bedrock. Book 3 reports this hybrid as where most peer banks have landed (Ch 9 §7). That is the author’s observation, not a survey.
Configure each route’s identity and client binding
Client and SDK configuration is where a route decision silently fails. The variable names below are from Book 3 App A §3 (May–Aug 2026) and Ch 4 PR-3, cross-checked where marked.
| Route | Client routing and auth (Claude Code / Agent SDK) | Grant needed | Traps |
|---|---|---|---|
| Anthropic API | ANTHROPIC_API_KEY (sent as X-Api-Key), ANTHROPIC_AUTH_TOKEN (bearer for a gateway), ANTHROPIC_BASE_URL (gateway endpoint) | Workspace-scoped key held in the secret store or gateway named values; separate “Claude Code” and “Developer” console key roles | A Console organisation is separate from a claude.ai Team or Enterprise organisation; settings and audit in one do not reach the other (Paper 3 §4.2 rule 5) |
| Bedrock | CLAUDE_CODE_USE_BEDROCK=1, AWS_REGION=<EU region>, AWS credential chain (role, Identity Center, environment) | bedrock:InvokeModel; Paper 3 §4.3 adds bedrock:InvokeModelWithResponseStream, bedrock:ListInferenceProfiles, bedrock:GetInferenceProfile | Model access is enabled per account per region before first use; prompt caching may not be available in every region; Claude Code uses the Invoke API rather than Converse (Ch 4 §4.3) |
| Google-managed Google Cloud | CLAUDE_CODE_USE_VERTEX=1, ANTHROPIC_VERTEX_PROJECT_ID, CLOUD_ML_REGION=europe-west1 (or eu), ADC or a service account | aiplatform.endpoints.predict (App A); Paper 3 §4.3 names roles/aiplatform.user | Model IDs are date-namespaced; not every model is on every regional endpoint |
| Foundry | CLAUDE_CODE_USE_FOUNDRY=1, ANTHROPIC_FOUNDRY_RESOURCE=<resource>; Entra bearer, resource key or the Azure default credential chain. SDK variables per vendor docs: ANTHROPIC_FOUNDRY_API_KEY, ANTHROPIC_FOUNDRY_RESOURCE or ANTHROPIC_FOUNDRY_BASE_URL (mutually exclusive) | Foundry User or Cognitive Services User (vendor docs, 15 Nov 2026) | Deployment names are administrator-chosen and immutable, and the model parameter takes the deployment name, so keep a deployment-name-to-model map. Book 3 reports no startup model check (failures occur at request time) and an opus alias resolving to an older default than on other routes (Aug 2026) [FACT-CHECK: current alias resolution]. Foundry omits Anthropic’s rate-limit headers, so watch limits through Azure Monitor. Keep both request-id and apim-request-id for support cases |
| Claude Platform on AWS | Book 3 Ch 4 PR-2’s matrix lists a _ANTHROPIC_AWS routing flag alongside CLAUDE_CODE_USE_BEDROCK (implying CLAUDE_CODE_USE_ANTHROPIC_AWS) [FACT-CHECK: Claude Code variable name]. SDK variables per vendor docs: ANTHROPIC_AWS_API_KEY (a key from the AWS console; standard Console keys do not work), ANTHROPIC_AWS_WORKSPACE_ID (wrkspc_ prefix), AWS_REGION, optional ANTHROPIC_AWS_BASE_URL, or SigV4 credentials | IAM service prefix aws-external-anthropic (e.g. aws-external-anthropic:CreateInference); enable outbound web identity federation once per account; sts:GetWebIdentityToken and sts:TagGetWebIdentityToken | A Bedrock-scoped role does not authorise this route (Ch 4 §4.4). Expect a permissions failure, and an access-review gap if nobody adds the new organisation to the inventory |
Auth precedence. Book 3 App A §3.1 gives the order Claude Code and the Agent SDK use: cloud-provider flags (Bedrock, Vertex, Foundry) → ANTHROPIC_AUTH_TOKEN → ANTHROPIC_API_KEY → apiKeyHelper script → CLAUDE_CODE_OAUTH_TOKEN → subscription OAuth. [FACT-CHECK: current precedence in the Claude Code authentication documentation.] Practical rules:
- Bind the route in the managed-settings
envblock, not in a shell profile, so a developer’s export cannot overrideANTHROPIC_BASE_URL(Ch 4 PR-3). On a gateway, Bedrock, Google Cloud or Foundry route, Claude Code skips the server-managed settings fetch, so device-delivered policy is the policy (Paper 3 §4.2 rule 6). Use the managed keyallowedProvidersto refuse sessions configured for any other provider (Paper 3 §4.2 rule 2; version-gated). Workstation detail is on coding-agent workstation. - Gateway in front of a cloud route. Point the client at the gateway with
ANTHROPIC_BEDROCK_BASE_URLorANTHROPIC_VERTEX_BASE_URL, and setCLAUDE_CODE_SKIP_BEDROCK_AUTH=1orCLAUDE_CODE_SKIP_VERTEX_AUTH=1so the gateway, not the laptop, holds the cloud credential (App A §3.3, §3.5). - No team-shared keys, no raw keys on endpoints. Use workload identity per agent (Ch 9 PR-1 step 2). See identity and secrets.
- Set every model-tier default explicitly (for example
ANTHROPIC_DEFAULT_HAIKU_MODEL) so background tasks do not fall back to the most expensive primary model (Ch 4 §6, Bank A composite).
Govern the Microsoft 365 Copilot surface as a separate route
Book 3 calls this “the sixth route, the one nobody procured” (Ch 4 §4.6). Per App A §3.2 (verified by the author against Microsoft Learn on 14 May 2026):
- Anthropic has been listed as a Microsoft sub-processor for select Copilot experiences since January 2026. The experiences named are Microsoft 365 Copilot (web, desktop, mobile), Researcher, Copilot Studio, Power Platform, Agent Mode in Excel, and the Word, Excel and PowerPoint agents.
- Those Anthropic models are excluded from the Microsoft EU Data Boundary, are not available in government or sovereign clouds, and are opt-in (off by default) for EU/EFTA and UK tenants.
- They are governed by Microsoft 365 terms, not Foundry terms.
[FACT-CHECK: current list of Copilot experiences using Anthropic models, their EU Data Boundary status, and whether they now run on Foundry-hosted capacity.]
Controls (Ch 4 §6–7):
- Separate register row and separate procurement and audit decision. Foundry Claude and Copilot Claude get two rows in the Register of Information.
- One shared owner who reads both the landing-zone configuration and the productivity-tenant configuration every quarter. In the Bank A composite, the DPO’s office owns the Copilot row.
- Bind AI-model enablement toggles to change control. That covers the M365 admin centre, Copilot Studio and every SaaS admin console with an AI sub-processor switch, all under the same change process as a Foundry deployment. Book 3’s test question for the architecture forum is: who can switch on an AI sub-processor here without raising a change? If the answer takes longer than ten seconds, you have found the gap.
- Withdraw the EU-residency assumption for any Copilot rollout that included Claude-served experiences.
Fill the contracting matrix
The contracting matrix is the decision’s evidence. Book 3 Ch 4 PR-2 defines one row per use-case class in four blocks, co-signed by procurement and the second line. The template below keeps the book’s fields. Fields marked † are this page’s additions, drawn from App A §5, App H §3 and the vendor corrections above.
Block 1: Scope
| Field | Values / notes |
|---|---|
| Use-case class | e.g. engineering productivity; employee assistance; regulated workflow agents |
| Risk tier / CIF flag / AI Act class | Tier per risk tiers; CIF yes/no; prohibited / high-risk / limited / minimal |
| Deployment route | One of the routes above. M365 Copilot needs its own row |
| Hosting option (Foundry only) | Hosted on Azure / hosted on Anthropic |
Block 2: Contracting and processing. This block expires. Re-derive it every quarter.
| Field | Values / notes |
|---|---|
| Data processor | Anthropic / Microsoft / AWS / Google, re-derived from vendor text this quarter, with the source and its date |
| Billing party | State it even when it equals the processor |
| CTPP-designated processor | Yes / No (Anthropic: No, per App A, subject to the FACT-CHECK above) |
| DPA reference | Anthropic DPA / AWS DPA / Google Cloud DPA. Never MOSPA alone on a Foundry row |
| DORA mandatory contractual clauses confirmed | Yes / No (article reference on the regulatory map) |
| Register of Information row ID | One row per processor relationship; Copilot separate |
| Transfer mechanism † | DPF (verified date) / SCCs Module 2 + TIA reference / Swiss–US DPF |
| Sub-processors in scope † | e.g. WorkOS on Anthropic SSO paths; sandbox providers for hosted agents |
Block 3: Technical configuration. The gateway and change register enforce these fields.
| Field | Values / notes |
|---|---|
| Region and inference residency | e.g. eu-central-1, in-region; Foundry, Global or US Data Zone; Anthropic API, inference_geo and workspace geo |
| Models authorised, pinned version | Exact deployment, inference-profile or model ID. No aliases |
| Client routing flag | CLAUDE_CODE_USE_BEDROCK / _VERTEX / _FOUNDRY / _ANTHROPIC_AWS / none |
| ZDR / HIPAA | Per route and organisation, with the ZDR configuration record reference † and the list of non-eligible features the class may use † |
| Expected consumption † | e.g. expected CCUs per workload on Claude Platform on AWS (App A §5); regional premium |
Block 4: Governance and review
| Field | Values / notes |
|---|---|
| Exit strategy reference | Vendor concentration and exit |
| Responsible roles | Named procurement lead, legal counsel, second-line reviewer, DPO, InfoSec architect |
| Second-line sign-off | Date and named reviewer |
| Review cadence | Quarterly minimum; re-review at every model release and every vendor terms change |
| Open watch items † | e.g. Foundry EU data zone; owner and next date |
Acceptance for a row: every block-2 cell cites a vendor source dated within the current quarter. The processor and biller are stated separately. The model ID is pinned. Every non-ZDR feature the class uses is named. Both signatures are present. The Register of Information row ID resolves.
Run the per-route procurement checklist
Book 3 App A §5 gives a per-route cheat-sheet for the procurement lead and contract lawyer. Verify every item against the order form on the contracting date.
| Route | Check at contracting |
|---|---|
| Anthropic API | Anthropic Commercial Terms (incorporating the DPA, Usage Policy and Privacy Policy), or an Enterprise MSA and PO. DPA with SCCs Module 2 plus UK and Swiss adapters; importer Anthropic PBC. ZDR addendum via the Enterprise account manager. BAA if needed, scoped to eligible features. Sub-processor feed subscribed. SLA: Book 3 cites a 99.99% Enterprise figure from third-party reporting only [FACT-CHECK: SLA in the executed MSA] |
| Foundry | MOSPA and Product Terms for the wrapper. Separately confirm DPA cover for the Anthropic processing leg and put that relationship on the register [FACT-CHECK: how the Anthropic leg is papered]. Hosting option chosen per deployment. Subscription tagged for cost allocation. RBAC role assigned. A written question to Microsoft on the EU data zone, with the answer filed. Private network: the vendor page now documents adding the resource to an Azure Virtual Network (Book 3 found no such guidance in August 2026) |
| Bedrock | AWS Customer Agreement or EA, AWS DPA and Service Terms; no separate Anthropic agreement. Customer-managed KMS key recommended for Book 3’s Tier 1 and above. PrivateLink endpoint per region, with the endpoint policy and IAM scoped to Bedrock invoke actions. Bedrock SLA version in force. AWS sub-processor list. Configure Bedrock retention and invocation-logging settings, and opt out of improvement programmes (Ch 10 §5.2) |
| Claude Platform on AWS | Two contractual layers (Marketplace order plus Anthropic Commercial Terms and DPA): verify both, and lock the DPA version referenced in the order form. ZDR scope in the addendum. No HIPAA. Confirm in writing that data is processed by Anthropic outside the AWS security boundary and draw it on the data-processor diagram. CMEK is available with AWS KMS keys per vendor docs (15 Nov 2026; Book 3 reported BYOK as “announced”) |
| Google-managed Google Cloud | Google Cloud Master Agreement, DPA and service-specific terms covering partner models. CMEK keyring in the endpoint’s region with a rotation policy. VPC Service Controls perimeter including Vertex AI. Private Service Connect. Access Transparency logs. Cloud Audit Logs |
| Anthropic-managed Google Cloud | Not in Book 3. Treat it as an Anthropic-processed route: Anthropic DPA, transfer record, new organisation in the identity inventory, ZDR enabled separately [FACT-CHECK: contract structure] |
Every contracting decision is logged in the Register of Information. The row template is in Book 3 App I §2, and the obligation is on the regulatory map.
Treat hosted managed agents as outsourcing
Book 3 Ch 9 §4.4 states its position plainly. Claude Managed Agents is an Anthropic-managed harness that provides state, memory, permissions, scheduling, sandboxing and tool execution. It is an ICT third-party service supporting whatever function the agent supports, not a deployment option you operate. As of 3 August 2026 the book reports it in public beta at token rates plus $0.08 per session-hour [FACT-CHECK: current status and pricing].
What follows for a regulated firm (Ch 9 §4.4, §9):
| Point | Consequence |
|---|---|
| If the supported function is critical or important, the third-party obligations attach (register entry, exit strategy, audit access) | Register it as an outsourcing arrangement before the first production session |
| Self-hosted sandboxes narrow the exposure but do not remove it. Tool execution moves to your infrastructure, but Book 3 notes the orchestration loop stays on Anthropic’s. Vendor docs (15 Nov 2026) add that attached memory stores remain stored by Anthropic and are copied to your sandbox for the session | Prompts, tool arguments and results still transit Anthropic: keep the transfer record and DPA cover |
| Third-party sandbox providers are fourth parties. Book 3 names Cloudflare, Daytona, Modal and Vercel [FACT-CHECK: current list] | Register the sub-outsourcing chain |
| No CIF on a beta service. The author’s position: Tier 0–1 workloads are a reasonable way to learn the surface; nothing supporting a CIF goes there until GA and a contract with SLAs, exit terms and audit access | Gate admission by tier in the envelope registry |
| ZDR: Managed Agents is not ZDR-eligible except self-hosted Managed Agents (vendor table, 15 Nov 2026) | Record the retention posture in the envelope |
Residency: inference_geo can be pinned per agent and overridden per session; on Claude Platform on AWS, sessions run with workspace geo us (vendor docs, 15 Nov 2026) | Same EU constraint as the parent route |
| Vaults keep credentials at the network boundary, so the agent never holds the real secret. But an injected agent can still use the credential against an allow-listed domain, and revoking a credential does not end a running session | Least-privilege scoping stays yours. Treat revocation as a session-termination event |
| Dreaming (scheduled memory curation) rewrites tomorrow’s memory | Manual review mandatory for a regulated function |
| Scheduled deployments are standing changes | The schedule needs approval, not only the agent |
Run Agent SDK agents as processes, not services
Choose the SDK or the raw API deliberately
Book 3 Ch 9 §5.5 treats this as a craft decision, governed by the same envelope either way:
- Agent SDK: multi-step reasoning with tool use, such as research synthesis, incident summarisation, regulatory-change diffing and evidence preparation. These need the hook model and subagents.
- Raw Messages API: a single-turn answer grounded in Citations, batched analysis through the Batch API (accepting that Batch is not ZDR), and tight latency budgets the SDK loop overhead cannot absorb.
- Subagents give you context isolation, tool scoping and parallelism. Spawn each with a tighter tool allow-list, a separate context and an explicit return contract. Subagents inherit the parent’s identity unless re-scoped at spawn (Ch 9 §5.1). Permission modes also inherit. Vendor docs (15 Nov 2026) say a subagent inherits
plan,acceptEdits,autoorbypassPermissionsfrom the main session, and runs its own mode only under adefaultordontAskparent. Bound each subagent withmaxTurnsin itsAgentDefinition. The vendor page states there is no per-subagent wall-clock deadline.
Know what you are deploying: the subprocess model
Book 3 Ch 9 §5.2 matches the vendor hosting page (15 Nov 2026). Calling query() spawns a separate claude CLI subprocess that talks over stdio and owns a shell, a working directory and session files on disk. One session is one subprocess.
| Property | Value / behaviour | Control |
|---|---|---|
| Starting capacity per agent | 1 GiB RAM, 5 GiB disk, 1 CPU. The vendor page calls this a floor, not the ceiling | Size hosts as (host RAM − overhead) ÷ measured peak per-session RSS |
| State on local disk | Transcripts (~/.claude/projects/ or under CLAUDE_CONFIG_DIR), CLAUDE.md memory files, working-directory artefacts; none survive a restart | Session store for transcripts; separate storage for memory and artefacts |
| Session timeout | None at the top level | maxTurns / max_turns plus an orchestrator-level pod deadline; recycle subprocesses to cap memory growth |
| Kill semantics | Killing the pod kills the shell and every child it spawned | That is your suspension mechanism. Make sure an autoscaler cannot trigger it by accident |
| Routing | Long-running sessions pin to one container | Consistent hashing on sessionId behind the load balancer (Ch 9 PR-1 step 7) |
| Version | The SDK bundles a CLI binary pinned to the SDK package version | Updating the SDK is how you update the CLI, so put SDK bumps under change control |
The SDK does not supply the envelope, the policy decision point, the evidence pipeline or the isolation posture (Ch 9 §5.6). Those come from the bounded-agent pattern, the model gateway, audit evidence and your platform team. The runtime must refuse to instantiate without a valid envelope ID, and the gateway must refuse traffic without one (Ch 9 §9).
Treat session-store mirroring as best-effort
The SessionStore adapter mirrors transcripts into storage you govern, and the vendor lists compliance and audit among its purposes. Book 3 Ch 9 §5.2 warns that it is a mirror, not a guarantee. The vendor page (15 Nov 2026) gives the detail:
- Retries: a rejected
append()is retried up to two more times (three attempts in total) with short backoff. A timed-out call is not retried. After that the SDK logs the error, emits{ type: "system", subtype: "mirror_error" }, drops the batch and continues. The agent does not stop. - Which copy is durable depends on how the run started. For a fresh session, the local transcript outlives the run and the store holds a copy. For a run resumed from the store, the local copy is deleted at run end, so the store holds the only durable copy, and a dropped batch on such a run has no surviving copy anywhere. Book 3’s statement that “the local file stays authoritative” is true only for the first case.
- Duplicates: retries can re-deliver entries. Deduplicate by
entry.uuidinappend(), and keep entries with nouuid. - Audit retrieval:
getSessionMessagesreturns the post-compaction chain the agent would see on resume. For the full record, callstore.load(key)(Book 3 Ch 9 PR-1 step 7 gives the same advice). - Retention: the SDK never deletes from your store. Lifecycle rules are the adapter’s job. Local transcripts are swept separately by
cleanupPeriodDays. - Conflicts:
persistSession: false(TypeScript) and file checkpointing both throw at startup when combined with a store. - Scope: only transcripts are mirrored, not
CLAUDE.mdor working-directory artefacts.
Controls. (1) A SOC detection rule on every mirror_error, with an owner. (2) For fresh sessions, reconcile the local JSONL against the store before container teardown. (3) For store-resumed sessions, treat a mirror_error as an evidence gap to record and report. (4) Do not assert transcript completeness to a supervisor without (1) to (3) in place (Ch 9 §9).
Put mandatory checks where the permission pipeline cannot skip them
Book 3 Ch 9 §5.3 and the vendor permissions page (15 Nov 2026) give the same six-step order. Assign an owner to each step:
| Step | Evaluated | Behaviour that matters | Owner (Book 3) |
|---|---|---|---|
| 1. Hooks | First | A hook deny holds in every mode, including bypassPermissions. A hook allow does not skip steps 2–3 | Platform security |
| 2. Deny rules | disallowedTools plus loaded settings | Deny even in bypassPermissions and even after a hook allow | Central policy |
| 3. Ask rules | settings ask | Falls through to canUseTool, even in bypassPermissions; denied in dontAsk | Central policy |
| 4. Permission mode | Active mode | bypassPermissions approves what reaches it; default and dontAsk move on to allow rules | Platform (mode is set in code; review it) |
| 5. Allow rules | allowedTools plus settings | Approved here never reaches canUseTool | Application team, within the envelope’s Means |
6. canUseTool callback | Last | Skipped and denied in dontAsk | Application team (user experience, not control) |
Book 3 calls the two documented warnings pre-written post-mortems:
- A compliance check in
canUseToolis silently bypassed for every call approved earlier by an allow rule,acceptEditsorbypassPermissions. Mandatory checks belong in aPreToolUsehook. The TypeScript SDK now emits aCLAUDE_SDK_CAN_USE_TOOL_SHADOWEDprocess warning when the configuration shadows the callback (vendor docs, 15 Nov 2026). Make that warning fail the build. allowedToolsdoes not constrainbypassPermissions. Unlisted tools are still approved. For unattended agents, use the documented locked-down pattern ofallowedToolspluspermissionMode: "dontAsk": listed tools run, every call that would prompt is denied, and there is no prompt path. In an unattended agent, behaviour must not depend on an absent human.
CI and unattended runs (Ch 9 §5.3, PR-1 step 8; version-gated to Claude Code v2.1.x, August 2026): claude -p --bare --allowedTools "<explicit set>" --permission-mode dontAsk --output-format json. Keep each run’s session_id and total_cost_usd as evidence, and fail the job whenever the system/init event carries a non-empty mcp_server_errors or plugin_errors array. [FACT-CHECK: flags and event fields against the current headless documentation.]
Harden the runtime: a bank-grade profile
Book 3 Ch 9 §5.4 collapses Anthropic’s secure-deployment menu into one profile for Tier 2 and above. The author derived it from the vendor page but it is not quoted from it. [FACT-CHECK: isolation figures and the proxy caveat against the current secure-deployment page.]
| # | Control | Setting | Acceptance check |
|---|---|---|---|
| 1 | Isolation by zone | Development zone: hardened Docker. Untrusted external content or multi-tenant: gVisor (--runtime=runsc) or a Firecracker microVM (Book 3 cites under 125 ms boot, under 5 MiB overhead) | Runtime class recorded per envelope; untrusted-content agents never on plain containers |
| 2 | Egress through a proxy, always | --network none; the only way out is a Unix socket mounted from the host, behind which a proxy enforces the destination allow-list, adds credentials and records each request | A direct outbound connection from the container fails; the proxy log shows every call |
| 3 | Credentials outside the boundary | An MCP or custom tool relays to an authenticated service, or a TLS-terminating proxy injects the header (Book 3 names Envoy’s credential_injector). Book 3 prefers the custom-tool route for banks: no certificate estate, and each call is logged against the agent’s identity. Do not rely on sandbox-runtime’s built-in proxy: it allow-lists by client-supplied hostname without TLS inspection, so domain fronting bypasses it | No secret readable inside the container; egress enforced at the network |
| 4 | Container flags | --cap-drop ALL, --security-opt no-new-privileges, --read-only with a noexec,nosuid tmpfs, memory/CPU/PID limits, non-root user, read-only code volume | Image scan and admission policy enforce the flags |
| 5 | No credential files in any mount, read-only included | .env, ~/.git-credentials, ~/.aws/credentials, ~/.azure/, ~/.kube/config, ~/.docker/config.json, .npmrc, *.pem, *.key | Mount manifest test fails on any match |
| 6 | Multi-tenant isolation | See below | Cross-tenant canary test |
In the Asset Manager X composite (Ch 9 §8), the research agent ingests external research, so it runs on gVisor with --network none and a Unix socket to the egress proxy that holds the APIM bearer token. The direct API behind APIM is itself the proxy-based credential-injection pattern (Ch 9 §7).
Isolate tenants with four controls that travel together
One tenant’s CLAUDE.md or auto-memory appearing in another tenant’s prompt is a data-segregation breach, not a quality bug (Ch 9 §5.4 item 6). The vendor hosting page (15 Nov 2026) confirms the set. Apply all four on every query():
settingSources: [](Pythonsetting_sources=[]), so no user, project or local settings load.CLAUDE_CODE_DISABLE_AUTO_MEMORY=1inenv. Auto memory loads into the system prompt regardless ofsettingSources.CLAUDE_CONFIG_DIRpointed at a per-tenant directory, so tenants do not share the global config. Build the path so no other tenant can read it.- A per-tenant
cwd, passed explicitly on every call.
Then add per-tenant egress rules at the proxy (distinct outbound identity, credentials or domain allow-lists), so a compromised tenant cannot exfiltrate through another tenant’s policy. One implementation trap from the vendor page: in TypeScript, env replaces the subprocess environment, so spread process.env deliberately. In Python, env merges. Note too that setting settingSources: [] also drops project .claude/settings.json permission rules, so supply those rules in code.
Acceptance test: plant a unique canary string in tenant A’s CLAUDE.md, auto-memory and working directory. Run tenant B’s session and assert that the canary is absent from the prompt, the transcript and all outbound traffic.
Re-run the portfolio every quarter
Book 3 App A §6 reviews route choices against six axes every quarter. The output is a row update in the route-decision register and an addendum to the next Register of Information submission. The review is a deliverable for the management body and a working artefact for the second line and internal audit.
| Axis | Check | Source of truth |
|---|---|---|
| 1. Model lifecycle | New models per route, deprecations and exposed workloads, and the next two quarters’ retirement calendar. Deprecation schedules can differ by route. Vendor docs (15 Nov 2026) state Foundry follows the Claude API lifecycle; Bedrock and Google follow their own policies | Anthropic model-deprecations page; each cloud’s policy |
| 2. Regulatory | Changes to the CTPP list, AI Act timing, supervisory guidance, incident-reporting templates | Regulatory map, calendar |
| 3. Anthropic | Trust Center updates, certifications, sub-processor changes, the policy versions in force | trust.anthropic.com |
| 4. Cloud provider | Foundry EU data-zone status; re-derive the Foundry processor statement; new Bedrock and Google regions and model IDs; Claude Platform on AWS feature and region changes (re-check the inference-geo position); provider sub-processor lists | Vendor documentation, dated |
| 5. Workload portfolio | New use cases and their routes; classifications that moved (Book 3’s typical case is a Tier 1 internal assistant becoming a Tier 2 customer-facing one); CIF flags added or removed; changed data-flow diagrams | Use-case registry |
| 6. Cost and concentration | Spend per route; concentration per provider as a share of the AI estate against limits; exit-rehearsal readiness per route | Vendor concentration and exit; running the factory |
Also reopen a row outside the quarterly cycle at every model release and every vendor terms change (Ch 4 PR-2). Bank A’s July re-run took eleven days, and Book 3 attributes that to a vendor fact changing, not a wrong choice. Its model-risk lead put it as choosing on facts that later expired (Ch 4 §6, composite).
Failure modes and recovery
Book 3 Ch 4 §7 and Ch 9 §9 name these patterns. The recognition signals and recoveries are condensed from them.
| Failure | Recognise by | Recover |
|---|---|---|
| One route for everything | ”We’ll just put everything on Foundry”, with no mapping to use-case classes | Push each class through the matrix against the four factors |
| Marketplace wrapper mistaken for processor | TPRM register names only the cloud provider on an Anthropic-processed route; “MOSPA” in a Foundry DPA cell | Re-derive the processor from current vendor text and add the Anthropic register row |
| An admin toggle opens a route | Copilot or SaaS AI features enabled with no change record | Put toggles under change control and name one owner across tenants |
| No data-flow diagram | Second line, DPO and TPRM describe the flow differently | One workshop; dated diagrams in the supervisory pack |
| Pilots leaked across routes | The AI inventory shows the same model on three routes with three evidence sets | Draw a freeze line on new pilots and consolidate within the next two release windows. Book 3 accepts the cost in goodwill; the alternative costs a year |
| Residency assumed from a region name (this page’s addition) | Claude Platform on AWS or Foundry recorded as “EU” because the workspace or resource sits in an EU region | Record inference geo and storage geo separately; the region does not pin inference on these routes |
| ZDR assumed for a whole agent | Envelope says ZDR while the agent uses Files, Batch, the MCP connector, Skills or code execution | Gateway refusal on non-eligible features; envelope flag for second-line review |
| No envelope, no production | A successful Claude call with no envelope reference | Runtime refuses without an envelope ID; gateway refuses forwarding; SOC rule on envelope-less calls |
| Silent evidence loss | Transcript gaps with no failed request anywhere | Alert on mirror_error; reconcile before teardown |
Compliance check in canUseTool | Allow-listed tools bypass the check | Move it to a PreToolUse hook; fail the build on the shadowing warning |
| Credential revocation that does not stop the agent | A revoked credential still produces successful calls | Treat revocation as session termination |
Ownership
| Artefact or control | Accountable | Consulted |
|---|---|---|
| Route-decision register and quarterly review | Head of platform (route-portfolio owner) | Lead architect, procurement, second line |
| Contracting matrix row | Procurement lead and second-line reviewer (co-sign) | Legal, DPO, InfoSec architect |
| Register of Information rows, including Copilot | Procurement lead | Second line, DPO |
| Transfer record (DPF / SCCs + TIA / Swiss DPF) | DPO with procurement lead | Legal |
| ZDR configuration record; non-eligible-feature exceptions | Privacy office; wrangler per envelope | Platform |
Client route binding (managed env, allowedProviders) | Endpoint engineering | Platform security |
| SDK runtime profile, tenant isolation, session store | Platform engineering owner of the SDK runtime | Security engineering |
Detection rules (mirror_error, envelope-less calls, token spikes, recursion depth) | SOC engineering lead | Wrangler |
| Managed Agents admission | Third-party risk owner and Agent Control Board | Second line |
Acceptance checks
- Every production use-case class has one signed matrix row. Block 2 cites vendor text dated within the current quarter, and processor and biller are stated separately.
- Every Foundry row names Anthropic as processor, records the hosting option, and has its own Anthropic register entry and DPA reference.
- Every Anthropic-processed row has a transfer record with a verification date, and a TIA reference where SCCs are relied on.
- No row claims EU in-region inference for the Anthropic API, Claude Platform on AWS, the Anthropic-managed Google offering or Foundry unless current vendor text supports it.
- Microsoft 365 Copilot’s Anthropic experiences have a separate register row, a named owner, and toggles under change control.
- Each organisation provisioned through a marketplace (Claude Platform on AWS, Anthropic-managed Google) is in the identity inventory and the access review.
- For every ZDR-dependent envelope, a synthetic request using
mcp_servers,/v1/filesor/v1/messages/batchesis refused at the gateway. - Model IDs are pinned on every route; a configuration scan finds no aliases in production settings.
- Managed Agents are absent from CIF-supporting envelopes; any Tier 0–1 use is in the outsourcing register with its fourth parties.
- SDK runtimes pass the hardening checks above, a
mirror_errortest fires the SOC rule, and the cross-tenant canary test passes. - No mandatory check lives only in
canUseTool, and unattended agents run withdontAskand an explicitallowedTools. - The quarterly review minutes cover all six axes, and the Register of Information addendum is filed.
Source status and differences
| Topic | Difference | Page follows |
|---|---|---|
| Foundry processor | App A (Aug 2026), Ch 4, Paper 3 and vendor docs say Anthropic; App H §3.2, the Glossary and Ch 10 §5.2 say Microsoft or the cloud provider | Anthropic |
| Claude Platform on AWS residency | App A and Ch 4 list EU regions with in-region inference; vendor docs (15 Nov 2026) give US or Global inference only, with the region not pinning inference | Vendor docs. App A §4 step 3 and Ch 4 PR-1 item 4 are corrected above |
| HIPAA | Book 3: direct API only, BAA extends to Claude Code with ZDR; vendor docs (15 Nov 2026): Claude API and Foundry eligible, Claude Code not covered | Vendor docs |
| Session-store authority | Ch 9 §5.2: local file stays authoritative; vendor docs: true for fresh sessions only, store-resumed runs delete the local copy | Vendor docs |
| Computer use | Ch 9 §4.2: not HIPAA-eligible; vendor table: ZDR- and HIPAA-eligible (client-side) | Vendor docs for eligibility; Book 3 for admission controls |
| Foundry feature and networking gaps | Book 3 (Aug 2026): no published feature-limitation list or VNet guidance; vendor docs (15 Nov 2026) publish both | Vendor docs |
| Anthropic-managed Google Cloud | Absent from Book 3 | Included from vendor docs; contracting marked for fact-check |
| Bedrock, Google-managed Google Cloud, M365 Copilot, Managed Agents status | Not re-verified for this page | Book 3 / Paper 3, marked [FACT-CHECK] |
Regulatory status, articles and dates are deliberately not restated here. See the regulatory map and regulatory calendar.
Source material
- Book 3 (Deploying Claude Safely), Chapter 4, The Deployment Decision (§3 main idea; §4.1–4.6 the routes and the M365 Copilot surface; §4.7 sovereignty gap; §4.8 AWS and GCP equivalents; §5 four decision factors; §6 Bank A worked example and July re-run; §7 anti-patterns; PR-1 Azure-first decision tree; PR-2 contracting matrix template; PR-3 cookbook)
- Book 3, Chapter 7, MCP and the Connector Broker (§4.1 the Messages API MCP connector and ZDR; per-route MCP parity)
- Book 3, Chapter 9, The Anthropic API and Agent SDK (§4.1–4.4 API surface, MCP from the API side, Managed Agents as outsourcing; §5.1–5.6 Agent SDK subprocess model, permission pipeline, runtime hardening, SDK versus raw API; §7 cloud equivalents; §8 Asset Manager X; §9 anti-patterns; PR-1 API agent decision tree)
- Book 3, Chapter 10, Data Classification, Retention, and DLP (§5.1–5.2 ZDR scope on the direct platform and the cloud routes)
- Book 3, Appendix A, Anthropic Deployment Options Matrix (§1 overview; §2 master matrix; §3.1–3.5 per-route deep dives and the M365 Copilot sidebar; §4 seven-step decision sequence; §5 procurement and contracting cheat-sheet; §6 quarterly review checklist)
- Book 3, Appendix C, MCP Server Registration and Approval Checklist (§5 item 17, ZDR posture of the mcp_servers parameter)
- Book 3, Appendix D, Bounded Agency Envelope Template (§2.3 persistent_storage files_api_on validation rule)
- Book 3, Appendix H, Vendor Concentration and Exit Playbook (§3.1 eight cleavage points; §3.2 per-route summary)
- Book 3, Appendix I, Regulatory Mapping (§3 DORA and Anthropic's CTPP status; §6.1 GDPR/FADP transfer mechanisms)
- Book 3, Appendix J, Glossary (Azure AI Foundry, Claude Platform on AWS, WorkOS, Sub-processor, Zero data retention, HIPAA-ready, Batch API, Files API)
- Research Paper 3, Securing Claude Code in the Enterprise (§4.1 approved routes and the Foundry hosting, residency and provider facts; §4.2 rules; §4.3 Amazon and Google equivalents)
- Research Paper 9, Data Classes, Processing Lanes and Controller/Processor Roles for AI (§6.2 where inference runs; §8.2 vendor documentation)
Related guidance
-
Design and Operate the Model GatewayBuild the model gateway as the governed edge for every model call: per-client identity, per-request highest-class routing, tokenisation, pinned models, a conformant wire contract, a correlated evidence record, a tested concentration answer, and day-two cost, latency and reliability operations across two estates and three meters. -
Data Classes, Processing Lanes and TokenisationClassify every model request by its data, route it only to a processing lane the class permits, keep derived artefacts inside their source class, and treat retention, roles and tokenisation as recorded, testable decisions. -
Design the Cloud Production Line and Hosting BoundariesClassify, segment, constrain and observe every workspace at the plane where actions execute, choose hosting routes per request, and keep vendor maturity, contracts and exit equivalence dated and explicit. -
Identity, Secrets and Configuration for Agent WorkloadsGive every actor one owned identity, size its authority for the worst case, separate runtime from deployment, federate instead of holding keys, refuse impersonation and silent fallback, govern configuration by class, and audit the estate until flagged identities decline. -
Govern MCP Connectors Through a BrokerAdmit, tier, pin and enforce every MCP server through one registry and two enforcement points, keep tokens audience-bound, govern A2A delegation through its own registry and gateway, and expose client-facing agent channels only through the same policy layer. -
Coding-Agent Workstation Baseline (Claude Code)Put Claude Code on developer workstations and build agents as privileged automation: one device-delivered managed policy per surface, lock flags that stop rules merging upward, command hooks for enforcement and HTTP hooks only for evidence, the OS sandbox where it runs, telemetry by policy, and the workstation recorded as a processing location. Dated to November 2026. -
Prepare a Governed Landing Zone for AI DeliveryDesign the management-group hierarchy, enforce guardrails in deny mode, give humans and agents the same paved workspace, integrate the eight agent-platform components, close the perimeter, and prove the zone is done with seven observable exit criteria and a second-line sign-off. -
Route Regulatory Questions to Obligations and EvidenceA dated, article-level reference that routes AI design questions to the EU AI Act, DORA and its RTS, GDPR and sectoral rules, and turns each obligation into a control, an owner and evidence: change records under RTS 2024/1774 Art 17, the 4 h / 72 h / one-month incident clocks, the CIF hinge and its five artefacts, the three-count Register of Information reconciliation, Art 14(4) oversight controls, the Art 26(5) discharge, the FRIA gate, Art 49 and Art 50, SCHUFA-grade review and supplier controls. -
Vendor Concentration and ExitMeasure concentration on three axes against 70% and 85% discussion triggers, keep a tested successor for every material model usage, record eight contract-cleavage points per route, prove evidence portability with a 15-item checklist, and rehearse exit on one material use case every year with a filed close-out report. -
The Audit Spine and the Evidence RecordHow to capture, join, protect, retain and prove per-action evidence: the AES v0.1.0 record contract, three capture points and one correlation ID, tiered immutable storage, retention and legal hold, replay rehearsal, the route-approval evidence gate, and the Claude-specific ways evidence goes missing. -
Model Governance: Usage Registry and Upgrade LifecycleGovern models at the grain regulators ask about: a catalogue for what exists, an 18-field usage registry for what is used, risk classified per usage, a registry the policy decision point reads by usage_id, pinned model IDs, and a four-phase upgrade lifecycle from Day −14 to Day +90 with numeric rollback thresholds. -
Bounded-Agent PatternA class 3 agent lets a model choose the next step, but only inside a registered, validated and enforced envelope of Mandate, Means, Memory and Monitoring. -
Operate the Agent LifecycleInventory, register, sign, zone, review, suspend, recover and retire agents through a six-stage lifecycle with named human ownership, zone-specific authority and evidence at every transition. -
Respond to Agent IncidentsSort an agent failure into one of four incident shapes, suspend it in three verified layers, keep humans on every classification and disclosure decision, route one playbook to every authority that needs a cover sheet, and close each incident with a named control change. -
Classify Use Cases into Risk TiersThe T0–T4 use-case taxonomy as a working decision reference: the prohibited-use gate, four intake questions, data and action boundaries per tier, the twelve-area control matrix, retention and payload rules, runtime enforcement, approval gates, regulatory routing and re-tiering.