Vendor Concentration and Exit
Measure 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.
- Version
- 0.1.0
- Owner
- DEIXEI
- Last reviewed
- Reading time
- 24 min
For Cto · Ai Architect · Platform Engineer · Procurement · Risk And Assurance · Internal Audit · Legal And Compliance
Sources and details
- Section
- Govern and prove
- Source works
- Book 3 · Book 1 · Research Series
- Source chapters
- 6, 10, 20
- Source appendices
- 04:A, 04:H
- Tool specifications
- audit_evidence_schema@0.1.0
- Refresh trigger
- quarterly
Concentration is the price of the convergence that makes enterprise agent platforms usable. Book 1 states the trade-off in one line: the firm designs to the converged platform primitives and contracts and exits against the concentration they create (Ch 10 §3). This page is the decision guide for the second half: how to measure concentration, when it must be escalated, what the contract must say for exit to be possible, what evidence must survive a route change, and how to prove, once a year, that the exit plan works.
Reading the labels. Source claim means a book states it. Author model means a framework the books propose (the three dimensions, the 70%/85% triggers, the eight cleavage points, the seven-step rehearsal, the 15-item checklist); these are not regulatory figures. Recommendation is this page’s own synthesis. Bank A is a composite firm written for the books. Vendor facts are dated “as of 14 May 2026, per Book 3 App H” unless stated and carry [FACT-CHECK] where they are current product or contract claims not yet re-verified against vendor documentation. The legal basis (DORA third-party risk, exit-strategy and suspension provisions, the Register of Information, the designation of critical ICT third-party providers, management-body duties) is stated, with article references and dates, only on the regulatory map and the regulatory calendar; this page links there rather than restating it.
Decisions this page supports
| Decision | Who decides | Section |
|---|---|---|
| Is our concentration on one provider, model or region high enough to escalate? | AI Steering Committee; management body above 85% | Measure on three axes |
| Which measure do we use: 70/85 per processor, or 80% criticality-weighted? | AI Platform Owner, second-line risk | Two threshold models |
| Does every material model usage have a tested successor? | AI Platform Owner with model risk | Model concentration |
| Can we suspend or leave this route cleanly, under the contract we actually signed? | Procurement lead, Legal and Compliance | Contract cleavage |
| How portable is this workload if we must change model provider, not just route? | Solution architect, wrangler | Envelope-level neutrality |
| Which use case do we rehearse this year, at what scope? | AI Steering Committee | Exit rehearsal |
| Did the evidence and the controls move with the workload? | Audit-evidence owner, second line | Evidence portability |
| Is our multi-cloud strategy real? | CIO, CRO | Multi-cloud theatre |
Measure concentration on three axes
Book 3 App H §2 separates three dimensions, each with its own measure, threshold, escalation and remediation lever. Assess all three every quarter (author model).
Vendor concentration: share by data processor
Measure the share of the AI estate by data processor, not by contract or invoice. The processor depends on the route:
| Route | Data processor at the processing layer (as of 14 May 2026, per Book 3 App H §2.1) |
|---|---|
| Direct Anthropic API | Anthropic |
| Claude Platform on AWS | Anthropic (AWS is the procurement layer) |
| Microsoft Foundry (Azure) | Microsoft, per App H; see the source conflict below |
| AWS Bedrock | AWS |
| Google Vertex AI | |
| Microsoft 365 Copilot with Anthropic as sub-processor | Microsoft at the surface; Anthropic at the sub-processor layer |
[FACT-CHECK: re-derive the data-processor statement for each route from current vendor terms before using it in a Register of Information entry; Book 3 App A §6 itself instructs this.]
Source conflict. Book 3 App H §2.1 records Foundry as Microsoft-processed. Book 3 App A §6, which carries later-dated facts (it records Foundry GA on 29 June 2026), tells the reader to re-derive the Foundry processor statement from current vendor text and notes it as “Anthropic per the GA announcement”. This page does not resolve the question. It follows App A: treat the Foundry processor as an open, dated fact; record the surface processor and any sub-processor as two layers in the concentration view, so the share can be computed either way; and confirm against the executed contract.
Three measurement bases. Compute the share on each of: token volume, inference cost, and number of registered envelopes per route. A breach on any one counts.
Thresholds (author model, Book 3 App H §2.1):
| Share on a single data processor, institution level, any basis | Consequence |
|---|---|
| Above 70% | Elevated-concentration discussion at the AI Steering Committee |
| Above 85% | Escalated to the management body |
The thresholds are discussion triggers, not caps. A firm running most of its estate on one route because that route is its strategic cloud destination is not in breach. It is in elevated-concentration territory and owes the management body two things: a documented rationale and a rehearsed exit path.
Model concentration: share by model class and version
Measure the share by model class (for Claude: Opus, Sonnet, Haiku) and by version within each class. The risk is concrete: a single-model dependency on a version approaching retirement. Deprecation schedules differ by route (the Anthropic API schedule and each cloud provider’s own policy), so track them per route in the model usage registry. The specific retirement dates quoted in Book 3 were as of May 2026 and have since passed or moved; read the current schedule at the vendor’s deprecation page at every quarterly review. [FACT-CHECK: current Claude model retirement dates per route, from Anthropic's model-deprecations page and each cloud provider's model lifecycle page.] The controls are on the next section.
Substrate concentration: share by region of inference
Measure the share of inference traffic by region where inference is served, not the region where the project or resource is created. Book 3 App H §2.3 treats substrate as a separate risk axis from vendor. Its route notes, as of 14 May 2026:
| Route and configuration | In-region for an EU firm? (per Book 3 App H §2.3) |
|---|---|
Bedrock in eu-central-1, eu-west-1, eu-south-1, eu-north-1 | Yes |
Vertex AI on europe-west1 or the eu multi-region | Yes |
| Foundry Global Standard | No: inference may be served outside the resource region; in-region only once an EU data zone exists |
Direct API workspace in us | No |
[FACT-CHECK: whether Microsoft has published an EU data zone for Claude on Foundry since June 2026, and current Bedrock and Vertex EU region availability for each Claude model ID.] Book 3 is itself inconsistent on Foundry (App H says an EU DataZone is “coming 2026”; App A §6 says only a US data zone existed at GA); treat it as an open watch item. The route mechanics are on deployment routes.
The concentration view and its cadence
Two sources set different cadences. Use both, as layers:
| Layer | Cadence | Content | Source |
|---|---|---|---|
| Concentration view (data) | Monthly refresh | Exposures aggregated by designated critical provider across all critical or important functions and all agents, so the firm’s total exposure to each provider is visible, not just each contract | Book 1 Ch 6 §8 (anti-pattern four) |
| Concentration assessment (decision) | Quarterly, to the AI Steering Committee | One sheet: one row per dimension, current measurement against threshold, movement since last quarter, open remediation actions | Book 3 App H §2.4 |
| Management-body extract | Annual | The three dimensions, rehearsal outcome, cleavage status per route, open triggers. In Book 3 Ch 20 §5 this sits in the annual board paper (the CFO or CIO presents, the CRO countersigns), with the exit-rehearsal close-out attached as an appendix | Book 3 App H §2.4, §7; Ch 20 §5 |
| Internal Audit review | Annual | The assessment and the playbook artefacts, typically alongside the Register of Information audit | Book 3 App H §7 |
The quarterly review takes three inputs (the concentration assessment, the cleavage status per route, the rehearsal calendar status) and produces one sheet. Rule: a route choice that worsens concentration without an offsetting trade-off gets a documented justification, not a quiet sign-off (App H §2.4). Book 1 Ch 6 §7 adds that each exit strategy for a dependency on a designated provider must be current, have a costed alternative path and be approved at management-body level.
Failure mode: a register complete per contract but blind in total (Book 1 Ch 6 §8). Every deployment on one hyperscaler is recorded individually; nobody has summed the rows. When a supervisory recommendation concerning that provider arrives, the firm cannot answer the impact question quickly. Recovery: the monthly concentration view above. Check: can the firm state, today, the share of its critical or important functions that depend on each designated provider?
Two threshold models: read them as layers
Book 1 and Book 3 give different numbers. They measure different things and do not conflict once separated:
| Model | Measures | Threshold | Purpose | Source |
|---|---|---|---|---|
| Concentration trigger | Share of AI estate per data processor, on tokens, cost or envelope count | 70% (committee), 85% (management body) | When to escalate the concentration conversation | Book 3 App H §2.1 |
| Multi-cloud theatre test | Share of agent workload, weighted by criticality, on the dominant hyperscaler | Above 80% means a multi-cloud declaration is theatre | Whether a declared multi-cloud posture is real | Book 1 Ch 10 §8 |
| Credible second platform | Criticality-weighted volume on a second hyperscaler | Roughly 20–30% | What “real” multi-cloud looks like | Book 1 Ch 10 §8 |
| Stress shift | Documented runbook | Move at least 10% within four weeks | Whether the second platform can carry load under stress | Book 1 Ch 10 §8 |
This page follows both. Run the 70/85 triggers every quarter for escalation. Run the 80% theatre test whenever the firm claims a multi-cloud strategy in a board paper, a supervisory answer or an exit plan. All figures are author proposals.
Model concentration: keep a tested successor
Control (author model, Book 3 App H §2.2): every Tier 2 and above envelope (tiers on risk tiers) has a tested successor model identified at all times. “Tested” means the successor has been run through the usage’s evaluation suite, under the same regime as a model upgrade (model governance), and the result is recorded against the usage row.
Open-risk rule: a workload pinned to a model within 90 days of retirement that has not been re-evaluated against its successor is an open concentration risk and is surfaced in the quarterly review.
| Check | Owner | Evidence |
|---|---|---|
| Successor named for every Tier 2+ usage | AI Platform Owner | Model usage registry row |
| Successor evaluated on the usage’s suite | Model risk, wrangler | Evaluation run reference and result |
| No pinned usage inside 90 days of retirement without re-evaluation | AI Platform Owner | Quarterly report listing exceptions |
| Retirement dates tracked per route | Platform | Registry, per-route deprecation source and date read |
Book 3 App A §6 places this inside a six-axis quarterly route review (model lifecycle, regulatory, Anthropic trust posture, cloud provider, workload portfolio, cost and concentration). The output is a row update in the route-decision register and an addendum to the next Register of Information submission.
Record eight contract-cleavage points per route
The contract decides how cleanly the firm can suspend or leave a route. Book 3 App H §3 names eight points to assess at contracting and at every quarterly review. The data populates the contracting matrix and the Register of Information row.
| # | Cleavage point | What to record | Why it matters for exit |
|---|---|---|---|
| 1 | Contracting entity | Named legal entity signed with | Sets jurisdiction, data-protection establishment and the supervisory relationship |
| 2 | Termination notice period | Notice to terminate without cause | Sets the minimum transition time |
| 3 | Termination for cause | Material breach, regulatory order, service-failure thresholds, supervisor instruction | Whether a supervisory instruction or failure lets the firm leave early |
| 4 | Data return and deletion | What is returned at termination, in which format, by when | Whether evidence and data leave with the firm |
| 5 | Sub-processor exit | What happens to sub-processors (for Anthropic-processed routes, the SSO mediator; vendor MCP servers) on route change | Hidden dependencies that outlive the main contract |
| 6 | Suspension versus termination | Right to suspend without terminating, and the mechanism | The operational lever short of exit |
| 7 | Designated-provider status | Whether the processor is a designated critical ICT third-party provider | Changes how exit is sequenced with the supervisor |
| 8 | Register of Information implication | Which register fields an exit would update | The exit is not done until the register says so |
The executed agreement, not the public terms, is authoritative for notice and data return. Verify per route at every contracting cycle.
Per-route snapshot (as of 14 May 2026, per Book 3 App H §3.2; every cell needs contract verification):
| Route | Contracting entity | Notice | Data return | Supervisory engagement |
|---|---|---|---|---|
| Direct Anthropic API | Anthropic PBC; no EU establishment; SCCs with UK/Swiss adapters | [FACT-CHECK] commercial terms 30 days; enterprise agreement negotiated | Per Anthropic DPA on customer election; Compliance API retention six years | Anthropic not designated; firm deals with its own competent authority |
| Microsoft Foundry | Microsoft Ireland Operations Ltd | [FACT-CHECK] per Microsoft product terms and enterprise agreement | 90-day post-termination retrieval per the Online Services DPA [FACT-CHECK] | Microsoft designated; joint ESA lead overseer |
| AWS Bedrock | AWS EMEA SARL (Luxembourg) | [FACT-CHECK] per AWS Customer Agreement and enterprise agreement | Per AWS DPA; customer-managed keys retained in customer account | AWS designated |
| Claude Platform on AWS | AWS Marketplace plus Anthropic commercial terms (two layers) | [FACT-CHECK] contracting surface was settling at GA | Per Anthropic DPA | AWS designation covers the procurement layer only |
| Google Vertex AI | Google Ireland Ltd | [FACT-CHECK] per Google Cloud Master Agreement | Per Google Cloud DPA; CMEK retained in customer project | Google designated |
The designation list and its updates are on the regulatory calendar; the book’s snapshot predates any later round. [FACT-CHECK: whether Anthropic has been designated in a later round.]
Three readings for the quarterly review (source claims, App H §3.2):
- Two routes sit outside the designated-provider regime at the processing layer (Direct API, Claude Platform on AWS). The firm carries the full third-party diligence load. On Claude Platform on AWS, draw the split explicitly in the Register of Information: AWS’s designation covers procurement, not processing. Book 3 quotes AWS documentation stating customer data is processed outside the AWS security boundary.
[FACT-CHECK: current AWS documentation for Claude Platform on AWS.] - Sub-processor exit differs by route. On Anthropic-processed routes, leaving Anthropic ends the SSO-mediator relationship without touching the firm’s own identity provider. On cloud-processed routes the cloud provider’s sub-processor list governs, and Anthropic may itself appear as a sub-processor. Sequence the exit accordingly.
- Suspension is the first lever. Foundry resource pause, Bedrock model-access revocation, Vertex endpoint pause and Anthropic workspace suspension preserve the contract; termination resets it. The rehearsal below starts with suspension.
Write the contract spine at the interface level. Name the capabilities the firm needs (per-agent identity, registry, runtime policy, identity-aware gateway, evaluations), not the vendor’s products, and map each vendor’s implementation to the list. For runtime policy, list the capabilities explicitly (content filtering across the OWASP categories, denied topics, PII redaction, output schema validation, formal-logic verification where the use case needs it). Policy engines from different vendors are not interchangeable; swapping one for another without re-running the map quietly loses a control (Book 1 Ch 10 §8, anti-pattern 3). The five-week procedure and the primitive equivalence map are on cloud and hosting; the per-primitive implementations are on the enterprise control plane.
Keep vendor neutrality at the envelope, not the feature
Route migration keeps the model and changes the path. Model-provider change replaces the model. Book 3 App H §6 is candid that the second is harder.
- Envelopes are vendor-neutral by design. Mandate, Means, Memory and Monitoring (bounded agent) say nothing about which model implements the agent. An agent built against the envelope abstraction can change model at manageable engineering cost; one with provider-specific prompt patterns hard-coded in its code path cannot.
- Some native features have no clean equivalent elsewhere. As of May 2026, per Book 3 App H §6: Claude for Excel, the Compliance API with its activity feed and six-year retention, specific Citations-plus-Files combinations and the Agent Skills catalogue.
[FACT-CHECK: current availability and retention terms of these Anthropic features.]A workload that depends on them is less portable than one on the Messages API alone. - Prompt portability is a real cost. Prompts tuned for one model do not transfer without re-evaluation. A workload moved to another provider goes through the same evaluation regime as a model upgrade.
Stance (author model, recorded in the AI risk-appetite statement): maintain envelope-level vendor neutrality as a control; accept feature-level dependency only where it is consciously incurred; record the portability cost per use case in the operating-model artefact at design time; re-evaluate it at every quarterly review and every model-upgrade change event.
The “no vendor names” test (Book 1 Ch 10 §8): for any agent in production, describe what it does and how it is controlled without the vendor’s product names. If that is not possible, the exit plan is performative. Retrofitting an interface layer once vendor specifics have leaked costs, by the author’s estimate, three to twelve person-months per primitive.
Configured platforms enter exit planning too. Research Paper 7 (§6.3) argues that a workflow configured into a proprietary low-code or integration platform that supports a critical or important function brings that platform into exit planning: what would replace it, how long the transition takes, and whether the configured logic can be extracted in usable form. The same capability built as code on commodity compute exits more simply. The paper treats this as a weight on the build side of a sourcing decision, not a verdict; legal and risk decide how heavy it is.
Rehearse exit every year
Rule (author model, Book 3 App H §4, Ch 20): rehearse exit on at least one material use case per year. The rehearsal is the evidence that the exit strategy operates. A control that has never been run is not yet a control.
Annual calendar
| Quarter | Activity |
|---|---|
| Q1 | Select the use case and scope |
| Q2 or Q3 | Execute the rehearsal |
| Q4 | Close-out report and the annual Register of Information update |
Over a three-year cycle, rotate across routes; Book 3 names at least Foundry, Direct API and Bedrock. In the 18-month roadmap (Book 3 Ch 20 §4.6), the first rehearsal is a Phase 5 exit criterion (months 12–18): exit and suspension rehearsed for at least one material use case, with the close-out report filed against the Register of Information row.
The seven steps
| # | Step | What happens | Owner | Output |
|---|---|---|---|---|
| 1 | Select | Pick one material use case. Material means any of: Tier 2 or above; supports a critical or important function; more than 100 active users; more than €100,000 annual run-rate cost | AI Steering Committee | Named use case, rationale |
| 2 | Scope | Choose one of three: provider suspension (fail over to a contracted alternative; hours), route migration (move to another route; days), full vendor change (exit the provider; weeks to months, table-top only) | AI Steering Committee | Scope statement |
| 3 | Brief | No surprise rehearsals. Brief the wrangler (holds the envelope), the business owner (holds workflow continuity), the second line (Legal and Compliance plus Operational Risk; holds the audit record). Brief the DPO if personal data is in scope. Brief Internal Audit for awareness, not sign-off | AI Platform Owner | Briefing record |
| 4 | Execute in non-production | Disable the route at the gateway (the route’s kill switch on the model gateway; suspension mechanics on incident response). Migrate to the alternative route in the contracting matrix | Platform, wrangler | Timestamped log |
| 5 | Evidence portability check | Sample correlation IDs from before the migration and reconstruct their evidence records from the alternative route’s telemetry. Run the 15-item checklist | Audit-evidence owner | Pass, or gaps with named owner and date |
| 6 | Operational continuity check | Quantify cycle-time impact, run-rate cost differential and quality impact against the use case’s evaluation set | Business owner signs | Continuity assessment |
| 7 | Close-out | One page minimum: use case, scope, timings, gaps, remediation owners, continuity quantification. File with the Register of Information row and the route’s contracting-matrix entry; the annual board briefing cites it | Second line | Close-out report |
Owners for steps 1 and 6 are the source’s (AI Steering Committee; business owner signs). The other owners and the outputs for steps 3, 4 and 7 are this page’s recommendation; the source names who is briefed and that the second line holds the audit record.
Timings to capture, to the second (step 4): wrangler decision; gateway kill; first request to the alternative route; full traffic redirection; first evidence record on the alternative route.
Production or non-production? Book 3 runs the rehearsal in non-production and flags supervisory expectation on production rehearsal as open. [FACT-CHECK: supervisory expectation on production versus non-production exit rehearsal.] A firm that chooses production documents the choice.
Acceptance check for a rehearsal: the close-out report exists, is filed against the register row, names every gap with an owner and date, carries a business-owner signature on continuity, and the next quarter’s remediation plan contains the gaps. Gaps from step 5 go to the next quarter’s plan with a named owner and date; they do not wait for the next rehearsal.
Composite example: Bank A’s first rehearsal
Composite, Book 3 App H §4.1; planned for May 2026. Bank A migrates one Tier 2 employee-assistance workflow (a meeting-summary assistant) from Foundry to Bedrock in eu-central-1, to show that Bedrock is viable as the alternative for that class even though Foundry is the strategic destination. Expected gaps the book names:
- Tokenisation cache discontinuity. The gateway’s tokenisation cache is keyed on correlation ID; on migration the mapping must be reloaded against the destination route’s evidence-record schema. Mitigation: dual-write the cache during the transition window, exercised under load in the rehearsal.
- Zero-data-retention posture changes. Per Book 3, Anthropic’s ZDR arrangement on the Messages API does not carry to the cloud-marketplace routes, where the cloud provider’s data-handling terms apply instead.
[FACT-CHECK: current ZDR availability per route.]The close-out must name which workloads tolerate the change without re-papering and which trigger a renewed third-party contract review.
Lessons feed the contracting matrix and the next year’s scope (planned in the composite as Direct API to Foundry).
Composite inconsistency. Book 3 Ch 20 tells Bank A’s story differently: its first Article 28(8) exit rehearsal is a route migration of a regulatory-summary workflow from Foundry to the Direct API behind the gateway, in October, with a separate suspension rehearsal (gateway kill 38 seconds after the wrangler’s decision) closing Phase 2. Both are illustrations, not data; the playbook steps are the same either way.
Check evidence portability before and after
Run this per use case before and after a migration (Book 3 App H §5). The delta is the evidence-portability outcome.
| # | Item | Where to check |
|---|---|---|
| 1 | Audit-evidence records exportable in machine-readable form (JSON or Parquet) | SIEM export plus object-storage copy (audit evidence) |
| 2 | Bounded-agency envelope record exportable | Agent registry, JSON export |
| 3 | MCP register entries for the use case exportable | Connector broker registry |
| 4 | Wrangler-attestation history exportable | Wrangler portal or registry; six-year retention in the source |
| 5 | Incident records for the use case exportable | Incident management system |
| 6 | Replay capability preserved | Replay procedure re-tested |
| 7 | Model-version pinning preserved or re-pinned | Model usage registry, per route |
| 8 | Correlation IDs preserved across the boundary | X-Correlation-ID propagation tested on both routes |
| 9 | Data-classification labels preserved | Labels re-applied at the new gateway hand-off |
| 10 | ZDR / HIPAA posture documented for the destination | Per-route assessment [FACT-CHECK: per-route HIPAA eligibility; Book 3 lists Direct API only, as of May 2026] |
| 11 | Sub-processor list updated for the destination | Vendor trust centre and provider list |
| 12 | Contract clauses updated and re-signed where needed | DPA, MSA, enterprise agreement, marketplace order |
| 13 | Register of Information row updated | Data processor, location, critical-function flag, exit reference |
| 14 | Operating-model documentation updated | RACI (accountability); runbooks |
| 15 | Wrangler training updated for the destination | Training and competence records |
Items 1–9 are technical portability (what the evidence pipeline must preserve). Items 10–15 are operational portability (what the operating model must update). The most common gap the source names: items 1–9 ticked and 10–15 untouched. The workload moved; the controls did not. Flag it in the close-out report.
Test whether multi-cloud reduces concentration
A multi-cloud declaration where production runs on one hyperscaler, a second hosts an idle development environment and a third hosts one proof of concept is theatre (Book 1 Ch 10 §8, anti-pattern 2): the firm pays for three platforms, operates one, and concentration is unchanged. Apply the 80% criticality-weighted test. Then choose one honest position: commit material workload to a second platform and rehearse the four-week shift, or declare single-hyperscaler dependency and design the exit plan, contract spine and resilience around it. The engineering detail is on cloud and hosting. The model gateway is itself a chosen concentration and needs its own entry.
System-level concentration. Book 1 Ch 10 §4.5 notes that sector-wide concentration on a few providers is a supervisory theme (FSB and BIS publications are cited there and on the regulatory map). Recommendation from the source: name an owner, typically in regulatory affairs with CIO and CRO input, to follow that dialogue and answer consultations with the firm’s own dependency data.
Triggers for an out-of-cycle review
Re-run the assessment and review the playbook, outside the quarterly cycle, on any of (Book 3 App H §7):
- a new route entering the portfolio (for example an EU data zone becoming available on a route already used);
- a change in designated-provider status (a new designation round);
- a change in a vendor’s model deprecation schedule;
- an incident that exposed an exit-mechanism gap (the incident close-out triggers the review);
- a contract renewal or amendment that materially changes any of the eight cleavage points;
- and, from this page: a concentration threshold crossed between quarters, or a Tier 2+ usage entering its 90-day pre-retirement window without a tested successor.
Ownership
| Activity | Accountable / runs | Consulted | Source |
|---|---|---|---|
| Quarterly concentration assessment | AI Platform Owner runs, with the Procurement Lead; AI Steering Committee reviews | Executive sponsor, model risk, Legal and Compliance | Book 3 App H §7; accountability row 27 (procurement Accountable, no Responsible named) |
| Exit rehearsal | Procurement Accountable; AI Platform Owner, AI Platform Product Owner and SOC Responsible | CISO, Legal and Compliance | Accountability row 28 |
| Continuity sign-off | Business owner | Wrangler | Book 3 App H §4 step 6 |
| Close-out filing and register update | Second line | Procurement | Book 3 App H §4 step 7 |
| Annual review | Internal Audit (third line) | — | Book 3 App H §7 |
The accountability matrix leaves row 27 without a Responsible party. Recommendation: name the AI Platform Owner, consistent with App H §7.
Failure modes and recovery
| Failure | Signal | Recovery |
|---|---|---|
| No exit rehearsal (Book 3 Ch 20 §7) | Contract clauses, alternatives and suspension mechanisms exist but have never run; the exit question is answered in the future tense | Annual rehearsal: one material use case, one named owner, one filed close-out |
| Performative exit (Book 1 Ch 10 §8) | Agents, policy and evaluations are expressed only in vendor-specific forms; the “no vendor names” test fails | Retrofit the interface-level spine primitive by primitive, vendor implementation isolated behind it |
| Multi-cloud theatre | Above 80% of criticality-weighted workload on one hyperscaler | Commit a real second platform or declare dependency honestly |
| Register blind to total exposure (Book 1 Ch 6 §8) | Rows exist per contract; nobody has summed them by provider | Monthly concentration view by provider across functions and agents |
| Workload moved, controls left behind | Checklist items 1–9 pass, 10–15 do not | Close-out flags it; remediation owner per item |
| Quiet sign-off of worsening concentration | Route approved that raises a share past a trigger, without justification | Documented justification required at the procurement gate; see change governance |
| Successor never tested | Pinned Tier 2+ usage within 90 days of retirement, no re-evaluation | Open risk in quarterly review; evaluation run before the window closes |
Acceptance checks
The programme is operating when a reviewer can confirm:
- a quarterly sheet exists for each of the last four quarters, with all three dimensions measured on the named bases and every threshold crossing escalated or justified;
- the concentration view aggregates exposure by provider across all critical or important functions and agents and is at most one month old;
- every Tier 2+ usage has a named, evaluated successor, and no pinned usage is inside 90 days of retirement unevaluated;
- every route has its eight cleavage points recorded from the executed contract, with a verification date;
- a close-out report from the last twelve months is filed against the Register of Information row, with timings, gaps, owners and a business-owner continuity signature;
- the 15-item checklist was run before and after the rehearsal and the delta is recorded;
- each production use case has a recorded portability cost and passes the “no vendor names” test.
Provenance and refresh
This page synthesises Book 3 Appendix H (the concentration framework, cleavage points, rehearsal playbook, portability checklist and model-portability stance), Book 3 Appendix A §6 and Chapter 20, Book 1 Chapters 6 and 10, and Research Paper 7 §6.3. Thresholds, step counts and checklist items are the authors’ proposals, not regulatory requirements. Vendor and contract facts are as of 14 May 2026 unless stated and have not been re-verified against vendor documentation; each carries a [FACT-CHECK] marker. Book 3 is internally inconsistent on the Foundry data processor and EU data-zone status; this page records both readings and follows the later App A instruction to re-derive from current vendor text.
Review trigger: quarterly, and on any trigger above. Before publication, have procurement and Legal confirm each route’s cleavage points against the executed agreements, and have the platform owner re-verify every vendor fact marked [FACT-CHECK].
Source material
- Book 3 (Deploying Claude Safely), Appendix H, Vendor Concentration and Exit Playbook (§2 concentration assessment; §3 contract-cleavage analysis; §4 exit-rehearsal playbook; §5 evidence-portability checklist; §6 model portability; §7 maintenance)
- Book 3 (Deploying Claude Safely), Appendix A, Anthropic Deployment Options Matrix (§6 quarterly review checklist)
- Book 3 (Deploying Claude Safely), Chapter 20, The 18-Month Roadmap and Reporting to the Board (§1 and §6 composite; §3; §4.6 Phase 5 exit rehearsal; §5 annual board paper and exit-rehearsal report; §7 anti-pattern 'no exit rehearsal')
- Book 1 (Building Organizations for the Age of AI), Chapter 6, DORA in Depth (§7 CTPP exit playbook refresh; §8 anti-pattern four, the register that ignores hyperscaler concentration)
- Book 1 (Building Organizations for the Age of AI), Chapter 10, The Convergent Platform Pattern (§3 thesis; §4.5 macroprudential implication; §6 worked example and contract spine; §8 anti-patterns)
- Research Paper 7, Build, Configure or Consume (§6.3 exit under DORA)
Related research
Related guidance
-
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. -
Regulatory Calendar for AI in EU Financial ServicesThe stop-the-press calendar rebuilt as a dated register: AI Act stages after the AI Omnibus, DORA and its RTS/ITS, CTPP designations, ESA and supervisor guidance, national laws such as Italy's Law 132/2025 with corrected dates, and harmonised standards, each with status, as-of date, official source and the reference pages it moves, plus the change-trigger procedure for when an item moves. -
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. -
Choose and Contract Claude Deployment RoutesDecide, 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. -
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. -
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. -
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. -
Accountability and OwnershipHold provenance and accountability apart on purpose: four named accountabilities per product, a human owner for every agent change, seven named seats, a signed Build/Run RACI and a first-line/second-line boundary that survives consolidation. -
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. -
Evaluate, Assure and Test Agent ResilienceDefine a versioned evaluation contract, test the trajectory and the execution boundaries, run three assurance layers into one quarterly pack, map the OWASP Agentic Top 10 to controls that demonstrably fire, bring agents into threat-led penetration testing with a five-section scope, keep a continuous cadence between formal tests, land every finding as a live control-plane artefact, and engineer recovery down to a minimal AI service level on models you host. -
The Enterprise Agent Control PlaneThe firm-owned services that make every agent action identifiable, decidable, approvable and reconstructable: the ownership boundary, six planes and their contracts, the Policy Decision API with its latency budget, the persistence schema, the build sequence and the tests that show it works. -
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. -
Build a Unified Control MapJoin the firm's AI and technology controls in one traceable operating view: rows are operating concerns, columns are frameworks read at their own altitude, every cell points to live evidence, and existing ITIL, COBIT and CIS artefacts are extended to agents with a five-question test instead of a parallel AI governance track.