Browse the reference ⌄
Decision Guide published

Classify Use Cases into Risk Tiers

The 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.

Version
0.1.0
Owner
DEIXEI
Last reviewed
Reading time
39 min

For Cto · Ai Architect · Platform Engineer · Risk And Assurance · Compliance · Internal Audit

Sources and details
Section
Design
Source works
Book 3 · Book 2 · Book 1 · Research Series
Source chapters
2, 4, 6, 11, 14, 22
Source appendices
04:D, 04:E, 04:G, 01:B
Refresh trigger
regulatory change

Full source list ↓ Source index →

Several factory pages ask you to “assign the risk tier” — at intake in the factory model, in the classification line, when routing a lane in agent readiness, and at the first gate of the agent lifecycle. This page defines what a tier is, how to assign one, what it switches on, and how it is enforced and changed.

The output of using this page is a tier decision: a registered use-case record carrying a tier (T0–T4), the control package that tier requires, the named signatories, and the evidence that each control is enforced by a mechanism rather than declared in a document.

Status of the model. The five-tier taxonomy, the twelve-area matrix and every SLA below are the author’s operating proposal in Book 3 (Chapter 14 and Appendix G), written for EU-regulated banks, insurers and asset managers deploying Claude. Regulators do not name these tiers. Asset Manager X, Bank A and Insurer I are the books’ composite institutions, not real firms. Vendor and product statements are dated to the source and marked where this page has not re-verified them. Legal statements and dates live on the regulatory map; this page only routes to them.

The operating rule

Book 3 reduces the model to three statements. This page adopts them as the acceptance test for the whole taxonomy:

  1. Every use case has a tier, and a use case without one does not run. The tier is set at intake, before any production traffic.
  2. Every tier maps to a control package, and the package is not negotiable per use case. Executive risk governance approves packages at the class level; the wrangler instantiates them per case. The approval body therefore approves a use case into a tier, and the controls follow. The debate in the room becomes “is this T2 or T3?”, not a fresh negotiation of controls.
  3. The tier is enforced at runtime — at the gateway, at the connector broker and in the agent registry — and approved by the Agent Control Board (ACB) for T2 and above. A tier whose controls exist only in policy is, in the source’s framing, a document rather than a control.

A tier is not a description of the use case. It is the contract under which the use case may run, and the registry is the authoritative copy of that contract.

Step 1 — Check the prohibited-use list before the taxonomy

The prohibited-use list is applied before tiering. A use case on the list is not tiered, does not enter the registry, and the gateway refuses its traffic.

The list (Book 3 Chapter 14 §6, as of August 2026)

#Prohibited use (items 1–5 reachable only through the Tier 4 exception route; items 6–8 not at all)What remains permitted, and at which tier
1Autonomous credit decisions affecting customersDrafting a credit memo for a human decision — T3
2Autonomous trading, order placement or portfolio rebalancingResearch and synthesis — T2 or T3; the order is placed by a person
3Autonomous payment execution or money movementDrafting payment instructions for human approval — T3
4Customer communications sent without human approvalDrafting — T3; the named approver sends
5Employment, disciplinary, compensation or promotion decisionsDrafting support for the human decision-maker — may be T2
6Biometric, emotion-recognition or surveillance use casesNothing. No exception route. Book 3 grounds this in the AI Act prohibitions and the provider’s Usage Policy; see the regulatory map for the legal position
7Any use case for which the firm cannot capture enough evidence for supervisory, audit or legal reviewThe test is the audit evidence schema: a use case that cannot meet it does not run, whatever its value
8The provider’s universal prohibitions, restated locallyNothing. Book 3 lists CBRN, critical-infrastructure attack, CSAM, election manipulation, fraud and sexual content per Anthropic’s Usage Policy of 15 September 2025. [FACT-CHECK: confirm the current Anthropic Usage Policy universal-prohibition categories and revision date at anthropic.com/legal/aup]

Item 8 is restated locally because, as Book 3 argues, the provider’s policy is a contract term: the provider enforces it by policy, and enforcement inside the firm is the firm’s own job.

Owning and maintaining the list

AspectRule (Book 3 Ch 14 §6 and §8, unless stated)
ApproverExecutive committee. Book 3 Chapter 2 instead calls the initial list the first artefact the AI Steering Committee approves, while pointing to Chapter 14 for the canonical version. This page follows Chapter 14: the executive committee owns and approves the list. (Recommendation: if your AI Steering Committee drafts it, record that as a proposal step, not as the approval.) Record which body your firm uses.
Review cadenceAnnually, and on regulatory change. Ad-hoc updates on a 30-day cycle after a relevant regulatory publication.
RemovalOnly with executive-committee approval and a documented rationale.
PublicationPublished internally; referenced in onboarding training; on the ACB standing agenda and read out at the start of each meeting.
SizeDeliberately short, so that it can be kept current.
Runtime effectNo registry entry can be created for a listed use case; the gateway refuses traffic for it.

Failure mode — list out of date. A regulator opinion names a use case after the list’s last review; a wrangler proposes it; the ACB has to argue the prohibition from first principles. Recovery: the 30-day update cycle and the standing agenda item.

Recommendation. Keep the list in the same registry that holds use-case records, as a controlled reference list with an owner and a version, so the “not on the list” check is a lookup with an evidence trail rather than a judgement remembered in minutes.

Step 2 — Ask four intake questions under least agency

Two wranglers must reach the same tier for the same use case, or the taxonomy looks arbitrary the first time internal audit samples it. Book 3 uses a four-question instrument that it attributes to Anthropic’s published CISO guidance (Jason Clinton, 17 July 2026, per Book 3 Ch 14 §4.1). [FACT-CHECK: confirm the four-question framing and “least agency” wording in Anthropic’s “Zero risk isn’t the job: a CISO’s guide to agentic AI”.]

The governing principle is least agency: grant the narrowest capability that still completes the task. It is the same rule as the least-agency rule for product shape in choose the shape of an AI product and the Mandate and Means sides of the bounded agent.

Intake questionWhat it sets (Book 3 Ch 14 §4.1)Record on the intake form (this page’s recommendation)
1. What untrusted content does the use case ingest?The data-class boundaryEvery data class read, with the source system; whether any content is externally supplied
2. What actions can it take, and on whose behalf?The action-class boundary and the autonomy ceilingRead / draft / recommend / execute; every write target; the identity it acts under
3. What is the blast radius if the agent is misaligned?The tier’s centre of gravityWorst credible outcome, who is affected (user, firm, customer, market), and whether output reaches a customer
4. What observability do we have?Whether the tier’s evidence requirement is achievableWhich capture points exist (gateway, hook or SDK callback, transcript) and whether the evidence schema can be met

A “not much” answer to question 4 has one honest classification: not yet deployable. Do not tier it lower to make the evidence requirement disappear; that is item 7 on the prohibited-use list.

Question 3 is where tiers diverge. A drafting assistant whose worst outcome is a lost afternoon and a correspondence agent whose worst outcome is a misstated claims decision reaching a policyholder are different classes, regardless of the model behind them.

Step 3 — Place the use case in a tier

The five tiers

Data classes below use Book 3’s data dictionary (Chapter 10): Public, Internal, Confidential — non-customer, Confidential — customer-non-financial, Confidential — customer-financial, Restricted, Strictly Confidential. Book 2 and Research Paper 1 use a coarser four-class scheme (public, internal, confidential, strictly confidential, with flags for personal data, client-identifying data and MNPI). Treat Book 3’s classes as a refinement of the confidential band; keep your own firm’s dictionary as the authority and map to it.

TierNameData-class boundaryAction-class boundaryAutonomy ceiling (envelope)Composite examples
T0Personal productivityPublic or Internal only. No customer data, no regulated material, no employee personal data beyond name and emailDrafting, transformation, summarisation. No tools beyond Read and Web Search. No writes to any internal systemNo registered agentRephrasing a slide title; formatting a JSON example for documentation
T1Internal knowledge assistanceInternal, Confidential — non-customer. No customer-financial, no regulated reporting data, no employee personal data beyond published role contextRetrieval over approved corpora and drafting. Read-only access to named MCP servers on the broker allow-list. No writesrecommendPolicy Q&A over the internal library; meeting summaries from analyst notes
T2Controlled operational assistanceAs T1, plus Confidential — customer-non-financial only with gateway tokenisation and an explicit dictionary entry. Restricted and Strictly Confidential prohibitedRead across allow-listed MCP servers. Writes only to the agent’s own evidence store and approved internal drafts. No external communication. Any production change needs its own T2 approvalrecommend, draft or execute_with_approval; execute_autonomously deniedService-desk triage proposing a runbook step; code-review subagent in the IDE; regulatory-research summary a lawyer reviews
T3Regulated process supportConfidential — customer-financial with tokenisation. Restricted only with named wrangler, named approver and per-output approval. Strictly Confidential prohibitedAs T2, plus: no autonomous external action, and every output crossing the agent’s boundary carries a named human approver in the audit chainrecommend, draft or execute_with_approval; execute_autonomously deniedCustomer-correspondence drafting; AML triage proposing a SAR; credit-memo drafting; regulatory-reporting support; drafting that contributes to a suitability outcome
T4Material autonomous actionBespoke; at least the T3 baseline; Strictly Confidential prohibited absent explicit executive exceptionAutonomous action with material effect. Out of scope by default; only via the executive exception route belowexecute_autonomously only with executive-committee approval, board notification and an independent legal opinionAutonomous credit, trading, payment, customer-affecting or employee-binding decisions

T2 is the first tier with a registered agent, a named wrangler and a written bounded-agency envelope. T3 is the tier Book 3 calls the line in the carpet.

Boundary tests

Most disputes are about two boundaries. Apply these tests and record the answer in the tier rationale.

T2 or T3? — the customer-output rule. If the output reaches a customer, even through human approval, the use case is T3. Book 3 also re-tiers into T3 when a T2 use case touches the provider’s high-risk finance categories (investment advice, loan approval, creditworthiness, insurance underwriting) or becomes biometrics-adjacent (for example voice-biometric KYC). Composite: Asset Manager X’s research-synthesis assistant stays T2 because it informs rather than makes an investment decision, never reaches a client without an analyst’s named approval and generates no recommendations; its regulatory-change analyst is T3 because its output shapes the firm’s compliance posture on retail-investment suitability.

T3 or T4? — material and autonomous. The question is whether the agent’s action is material and autonomous in the regulatory sense, not how narrowly the wrangler has drawn the boundary on the diagram. A use case that would be T4 cannot be scoped into T3.

The burden of proof runs downward. The ACB’s first question on every new use case is what tier is this, and why is it not the tier above? The proposer argues the case down; the second line challenges the down-tier rationale; the minutes record both the tier proposed and the tier approved, so that tier shopping becomes visible across decisions.

Classify by harm, not by role, model or readiness

A tier is one value among several on the use-case record. Keep it separate from the others.

Do not tier byWhyWhere that value lives instead
The vendor’s product categoryBook 3 notes that Anthropic shipped role-named finance plugins for Claude Cowork on 24 February 2026 (Financial Analysis, Investment Banking, Equity Research, Private Equity, Wealth Management). [FACT-CHECK: plugin names and launch date per claude.com/blog/cowork-plugins-finance.] Those names are organised by job; tiers are organised by harm. Wealth Management output that runs toward a client is T3; Equity Research used for internal drafts an analyst reviews is T2. Same plugin family, two tiers apart.The intake answers to questions 1–3
The model”Sonnet = T2, Opus = T3” turns the taxonomy into a model-selection rule, and the same use case gets a different tier depending on the model picked. Model choice is a cost, latency and accuracy decision taken inside the tier. A T3 use case may use a cheaper model; it cannot drop a tier because it did.Model field on the use-case record; model governance
Station readinessBook 2 Chapter 4: readiness routes the delivery lane; the risk tier authorises the autonomy. The two are orthogonal and have different owners. A high readiness grade is never permission to lower the tier.Agent readiness
Agency classAgency class (class 0–4 in Book 2 Ch 22; A0–A4 in Research Paper 1) says who decides the next step; it is a design property. Book 2 sets the autonomy ceiling by risk, separately from mechanics, so the two vary independently — for illustration, a bounded agent (class 3) may be T1 while a workflow (class 2) is T3.Classification line
DORA CIF statusThe tier does not decide whether a use case supports a critical or important function. That is a parallel second-line determination. Most T3 use cases are CIF-supporting, not all; some T2 ones are.A separate CIF flag on the record. The tier carries the control package; the CIF flag carries the third-party contract surface. See the regulatory map.
Regulatory classBook 1 Appendix B records a regulatory_class per usage (prohibited / high-risk / limited / minimal, with sectoral tags) — the same model endpoint may be high-risk in one row and minimal in another. That is a legal classification; the tier is the firm’s control classification informed by it.Model usage registry regulatory_class

Licensed data is a tiering input too. Book 3 flags an open contract question (as of August 2026): whether a model session that ingests licensed market data (its examples are LSEG, FactSet and MSCI) and emits derived content counts as redistribution. Until your market-data team and general counsel have an answer, put licensed-feed access behind the connector broker with a named entitlement owner, so that a T2 research assistant does not become an unlicensed redistribution channel. [FACT-CHECK: whether any market-data vendor has published licence terms addressing LLM or agent session use.]

Step 4 — Write the use-case record

Book 3 Chapter 2 makes the use-case record — not the model — the unit of governance. The model becomes one field; the record carries the risk tier, data classification, human-oversight design and the action surface (read / draft / recommend / execute). The model can change without the record changing. The record cannot change without ACB review (routed through the Virtual CAB; see re-tiering below). That is what lets governance survive the next model release or the next provider, and it is the basis of the portability objective: the firm can route, suspend or migrate any use case without losing evidence or accountability.

Minimum fields, taken from the Book 3 Appendix D envelope Mandate (which requires them as enumerated, version-controlled fields rather than prose) and the Chapter 14 control notes:

FieldTypeNotes
use_case_classenumTwo meanings in Book 3. In the App D envelope it is a functional category (research_synthesis, regulatory_drafting, incident_summarisation, evidence_preparation, code_review, claims_triage, customer_correspondence, other). On the gateway record (Ch 6, App E) it is class0–class3, read from the request header. Recommendation: keep both, under distinct names in your registry; see the vocabulary note under runtime enforcement
risk_tierenum T0–T4Set at intake; changed only through re-tiering
business_owner_namestringAccountable business-line owner; first-line accountability for value
wrangler_namestringMust resolve to an active directory principal (Entra ID, AWS IAM Identity Center or GCP Cloud Identity) at registration and at each quarterly review
named_approver_rolestringIn App D, the role authorised to approve Mandate amendments — second line for T2 and above. Do not confuse it with the T3 per-output approver, whose identity (UPN) is captured per action in the approval-queue record
autonomy_ceilingenumrecommend, draft, execute_with_approval, execute_autonomously; immutable without a Virtual CAB change
customer_facingbooleanWhether output reaches an identifiable customer; true forces T3 or above
data_classes_permittedlistFrom the data dictionary; the gateway DLP check compares observed classes to this list
fria_referencedocument IDNon-null when the class maps to the Annex III categories Book 3 names; legal applicability per the regulatory map
usage_policy_alignmentobjectConfirms alignment with the provider’s Usage Policy; for the provider’s high-risk categories it carries the human-in-the-loop control reference and the AI-disclosure path
CIF flagboolean + referenceSeparate from the tier (see above)
Tier rationaletext + signaturesTier proposed, tier approved, why not the tier above; wrangler, DPO and Model Risk lead, business owner; for T3 the named second-line signatory

Book 2 Chapter 22’s six-value classification line (anatomy tier / pattern family / agency class / risk tier / data class / reuse generation) carries the same tier into the product’s bill of material, so policy-as-code can attach the right policy pack without a meeting. Use the same tier value in both places.

Step 5 — Apply the tier’s control package

The twelve control areas

Chapter 14 shows a compact six-area matrix (identity, evidence retention, human oversight, evaluation, approval, reporting). Appendix G expands it to twelve areas for the second-line signatory’s ACB pack. The two are largely consistent; the material difference is retention (see the retention section). This page uses the twelve-area form.

Control areaT0T1T2T3T4 (exception only)
1. Identity and accessSSO to the managed chat surface; no personal accounts; named role group; revoke on leaverSSO + conditional access; named role group; SCIM-provisioned; revoke on leaverWorkload identity + agent identity; gateway rejects a wrong identity pair; wrangler identity capturedWorkload identity + agent identity + named human approver in every evidence record; gateway enforces the trioExecutive-approved identity model; at least the T3 trio plus a board-notified naming convention
2. Gateway routing and DLPStandard gateway log; correlation ID; outbound DLP class checkDLP class check + retrieval-source log; correlation ID across the chain; tokenise only on accidental sensitive entriesFull DLP + tokenisation; outbound class check enforced; policy decision pointFull DLP + tokenisation; outbound class + content-grounding check; policy decision point enforces named-approver presenceBespoke; at least T3 plus AI Steering Committee-approved additions
3. MCP allow-list class (connector tiers R0–R2, see broker)None or R0 (read-only, public/internal)R0, or R1 (read on confidential/regulated data) with DPO + data-class owner approvalR0–R2; R2 (write or external HTTP on regulated data) needs Virtual CAB approvalR0–R2; R2 mandatory for any write or external HTTP; named wrangler + approver per call where the action warrantsDefault deny on any new R2 server without executive approval; suspension rehearsal before admission
4. Workstation managed settingsStandard SSO posture; no coding agent requiredStandard managed-settings baseline if a coding agent is used; hooks emit standard telemetryFull baseline; hook bundle referenced in the envelope; fail-closed remote settings refresh; cohort gate evidenceFull baseline; cohort review on every model upgrade; hooks capture extended-thinking traces where the route supports itWorkstation use prohibited; runs server-side only; workstation sessions may not invoke T4 envelopes
5. Envelope autonomy ceilingNo registered agentrecommend; risk_tier: T1 recorded where an envelope existsEnvelope mandatory; up to execute_with_approvalEnvelope mandatory; up to execute_with_approval; every boundary-crossing output has a named approverexecute_autonomously only with executive approval, board notification, independent legal opinion
6. Data classesSee tier tableSee tier tableSee tier tableSee tier tableSee tier table
7. Evidence retention classSee retention sectionSee retention sectionFull audit evidence schemaFull schema + thinking trace + citations chain + approver identity and decisionAt least 10 years; legal hold on by default
8. Approval bodyLine manager at onboarding; training recordACB notificationACB full approvalACB + named second-line signatory; AI Steering Committee notified; FRIA where Annex III appliesExecutive committee with board notification; CRO + CEO sign-off
9. Evaluation cadenceLight-touch; user feedbackWeekly pass on a representative sample; hallucination-rate threshold; source-grounding checkPre-deployment suite + monthly re-runs; drift thresholdAs T2 + Model Risk validation report on the firm’s quantitative-model cadenceBespoke; at least quarterly Model Risk re-validation; independent assurance every six months
10. Incident-reporting thresholdInternal log unless individual harm; SOC ticket on a security eventSOC ticket on any control denial above the noise floorSOC ticket on every control denial; regulatory incident classification considered on any envelope breach or data-class crossingRegulatory incident classification on every material breach; parallel personal-data breach gate; supervisor-ready pack on demandExecutive notification on any breach; all incident channels default-on
11. Red-team coverageNone (sampling if usage warrants)Sample-based; annual programme above a usage thresholdQuarterly mini-exercise, OWASP Agentic Top 10 categories ASI01–ASI03 minimumQuarterly full scenario against OWASP Agentic Top 10 plus MITRE ATLAS agent techniques; TLPT inclusion where the entity is in scopeContinuous adversarial monitoring; bespoke programme; TLPT where in scope
12. Reporting cadenceMonthly aggregate usage to ACBMonthly to ACB: usage, accuracy spot-checks, incidentsMonthly to ACB; quarterly to AI Steering Committee (incidents, control effectiveness, value)Monthly to ACB; quarterly to AI Steering Committee + risk committee; annual to the management bodyQuarterly to executive committee and board; ad hoc on any breach; annual independent assurance report

Additional human-oversight rules from Chapter 14 that sit across rows 1, 5 and 8:

  • T0–T1: the user is the oversight; every output is read before reuse.
  • T2: human-in-the-loop on every action that crosses the agent’s boundary; the wrangler handles escalations within one business day.
  • T3: the review must be meaningful. The interface, the latency budget and the reviewer’s override authority are documented and validated. Book 3 grounds this in the CJEU SCHUFA judgment and AI Act human-oversight duties; the legal reading belongs on the regulatory map.

Vendor-specific cells (as of May 2026, per Book 3 Appendix G). The source’s identity cells name Microsoft Entra ID, Conditional Access and Entra Agent ID; its workstation cells name the Claude Code managed-settings baseline with forceRemoteSettingsRefresh: true; its retention cells name Microsoft Sentinel and Azure immutable Blob storage. Substitute equivalents on AWS or GCP estates (Book 3 Appendix D allows AWS IAM Identity Center or GCP Cloud Identity principals for the wrangler). [FACT-CHECK: Entra Agent ID availability and naming; forceRemoteSettingsRefresh behaviour in current Claude Code managed-settings documentation.]

Read the matrix two ways, and prove every cell

Across a row is the firm’s contract with itself for one use case. Down a column is the scaling rule for one control. Book 3 sets two invariants, which this page adopts as acceptance checks:

  1. Monotone scaling. No cell may be weaker than the cell in the tier below it. A cell that loosens against the lower tier is a registration error, not a design choice.
  2. Enforced, not declared. Every cell names the mechanism that enforces it and points to evidence of its last successful enforcement. A cell that cannot be evidenced is a control gap.
Control areaEnforcing mechanism (Book 3 App G §3.3)Evidence to ask for (this page’s recommendation)
Identity and accessNot named in App G §3.3; this page proposes the identity provider plus the gateway identity-pair check (App G row 1)Last rejected wrong-pair call; leaver revocation record
Gateway routing and DLPModel gatewayDLP decisions and tokenisation events for the use case
MCP allow-list classConnector brokerBroker allow-list version and last refused call
Workstation postureDevice management pushing managed settingsCohort compliance report
Autonomy ceilingAgent registryRegistry record and drift report
Data classesData dictionary (checked at the gateway)Dictionary entries and DLP class matches
Retention classThe writer that routes records to the locked immutable container by retention_tier (in Book 3, a Logic App writing to immutable Blob)Container policy and a sample record per tier
Approval bodyStanding ACB packSigned minutes and named signatures
Evaluation cadenceWranglerLast evaluation run and threshold result
Incident thresholdSOCDetection rule and last ticket
Red-team coverageSecond lineLast exercise report
Reporting cadenceACB secretariatLast report issued

The question to ask of each cell, at every approval: which mechanism enforces this, and where is the evidence of its last success? The answer should be the same mechanism, evidence type and retention for every use case in that tier.

Retention and payload truncation by tier

What the sources say — and where they disagree

Book 3 gives retention values in four places, and they do not agree. Do not blend them; pick a rule and record it. (The Ch 6 column is placed on the assumption that classN corresponds to TN, which Book 3 does not state.)

TierCh 14 §4 and OD-3Ch 6 PR-3 noteCh 11 PR-3 tableApp E §5 table (and App G row 7)
T013 months gateway log—Sentinel hot 90 days; immutable 1 year, labelled class1-12mclass0-12m: hot 90 days → immutable 9 months; total 1 year. App G’s cell says “13-month gateway log” beside the same class0-12m label
T113 months (gateway + retrieval-source log)class1 12 monthsHot 365 days; immutable 1 year (class1-12m)class1-2y: hot 180 days → immutable 18 months; total 2 years
T23 years, full schemaclass2 12 monthsHot 365 days; immutable 3 years (class2-3y); cool tier years 2–3class2-3y: hot 1 year → immutable 2 years; total 3 years
T36 years; 10 where sectoral rules exceedclass3 6 yearsHot 365 days; immutable 6 years (class3-6y); cool tier years 2–6class3-7y: hot 2 years → immutable 5 years; total 7 years. App G calls this a 6-year minimum, with sectoral overrides (its example is 10 years for German commercial and tax records)
T4Use-case specific; minimum 10 years—— (out of scope)Not covered (App E); App G: minimum 10 years, legal hold default-on

Four specific inconsistencies to know about:

  1. T0: 1 year or 13 months. Chapter 14 and the App G cell say 13 months; the App E table says 1 year, and the label class0-12m encodes 12 months.
  2. T1: 1 year, 13 months or 2 years. Chapters 6, 11 and 14 give roughly a year; App E and App G give 2 years.
  3. T3: 6 or 7 years. Chapters 6, 11 and 14 give 6 years (aligned, the source says, to DORA contractual minima and Anthropic’s Compliance API window); App E and App G use class3-7y. The App E field definition also lists the enum as class0-12m, class1-12m, class2-3y, class3-7y — which contradicts its own §5 table (class1-2y).
  4. Labels that do not match durations. Chapter 11 labels Tier 0 class1-12m; App G describes class3-7y as a 6-year minimum. A label that misstates its duration will eventually be trusted by someone.

Rule this page follows

Source rule (Ch 14 OD-3): retention is the longest of the regulatory minimum, the contractual minimum and the firm’s own evidence-replay requirement. Recommendation: apply that rule to the conflicting book values too, and take the longest stated figure per tier as the default, until legal and the DPO set the firm’s values:

TierDefault total retention (this page)Payload rule (Ch 11 §9 recovery; App E)
T013 monthsTruncate at 4 KB; full payload not retained; metadata only beyond the SIEM hot path
T12 yearsTruncate at 4 KB; full payload not retained
T23 yearsTruncate at 16 KB in the hot path; full payload to the immutable store on tag
T37 years (never less than 6), plus any longer sectoral periodFull payload to the immutable store always; truncation only on the SIEM hot path
T4At least 10 years; legal hold on by defaultBespoke; at least the T3 rule

Then make the enum name encode the duration you actually configure. App E’s hook-record fields show the split: tool_input_truncated_4kb and tool_result_truncated_4kb carry the hot-path copy, and tool_result_full_16kb_ref points to the full payload in the immutable store. The event stream fans out: the SIEM takes the truncated stream, the immutable store the full one.

Why the decision cannot wait. Book 3 relies on time-based retention on immutable storage, which once locked can be extended but not shortened. A record written to the wrong tier’s container costs a full retention window the firm cannot recover. Decide the values, lock one container per tier at creation, and have the writer route by retention_tier. [FACT-CHECK: Azure immutable Blob locked time-based retention semantics per Microsoft Learn, as cited by Book 3 App E §5 as of May 2026.] Retention and erasure interplay (legal hold, data-subject erasure versus record-keeping duties) is a DPO decision documented per record class; see audit evidence.

Failure mode — truncation defeats reconstruction. Book 3’s workstation hook truncates tool results to 4 KB, which suits cost but fails a T3 use case whose decisive response is 12 KB. Acceptance check: the quarterly replay exercise reconstructs one T3 action from the full payload, and any truncation that blocked a reconstruction is a finding with an owner.

Vendor detail — telemetry content is a tier decision (as of August 2026, per Book 3 Ch 11). Claude Code and the Agent SDK emit OpenTelemetry that is structural by default (durations, tool and model names, no content). Content export is opt-in through OTEL_LOG_USER_PROMPTS, OTEL_LOG_TOOL_DETAILS, OTEL_LOG_TOOL_CONTENT and OTEL_LOG_RAW_API_BODIES. Record which are enabled per tier as a DPO-approved setting. Export fails silently by default: set CLAUDE_CODE_OTEL_DIAG_STDERR=1 and alert on collector liveness. [FACT-CHECK: variable names and defaults in current Claude Code / Agent SDK observability documentation.]

Enforce the tier at runtime

The tier is only a control if the runtime refuses traffic that does not match it. Book 3 places enforcement in four mechanisms.

Gateway

  • Class header. The gateway reads X-Use-Case-Class, validated against the caller’s token scope, and records it as use_case_class on every evidence record (Book 3 Ch 6). The class selects the route and the policy stack.
  • Two policy stacks. Customer-facing, regulated or CIF-supporting workloads get the full stack: inbound DLP, tokenisation, budget enforcement, a policy decision call on consequential actions, response-side DLP and paired-region failover. Internal T0–T1 workloads without customer data get a reduced stack: logging, budget, rate limit and response-side DLP. No workload bypasses the gateway, including T0–T1.
  • Policy decision point. Book 3 Ch 6 has the gateway call the Policy Decision API for actions that require approval — writes against regulated systems, customer-affecting decisions, autonomous payments — and the App G matrix makes it part of the T2 gateway cell and, at T3, the point that enforces named-approver presence. Book 3 gives no finer per-tier invocation rule; define one in your policy contract. Its verdicts are named allow, allow-with-evidence, escalate-for-approval and deny in Chapter 6, while Chapter 11 calls the approval verdict allow-with-approval. Pick one name in your policy decision contract and use it in the evidence schema.
  • Identity. At T2 the gateway rejects calls bearing the wrong workload/agent identity pair; at T3 it enforces the trio including the named approver.

See model gateway for the gateway contract and bypass controls.

Vocabulary note — class0–class3 versus T0–T4. Book 3 uses two identifier sets: the gateway’s use_case_class (class0–class3) and the envelope’s risk_tier (T0–T4). (The envelope also has a field named use_case_class, but there it is a functional category, not a class number.) There is no class4, consistent with T4 not running without an exception. The books do not publish the mapping explicitly. Recommendation: keep a one-to-one mapping table (T0→class0 … T3→class3) in the registry, have the gateway reject any request whose header class differs from the registered tier, and treat a T4 exception as a separately named class rather than reusing class3.

Connector broker

The broker admits only MCP servers whose connector tier (R0–R2) is permitted for the use case’s tier (row 3 of the matrix). Connector tiers classify the connector; use-case tiers classify the use case — do not mix the identifiers.

Agent registry

  • The envelope’s risk_tier and autonomy_ceiling are registered, machine-readable and immutable without a Virtual CAB change.
  • An envelope whose actual conditions (data classes observed, servers called, customer-facing status) have drifted from its registered tier is in registration error, and the gateway refuses traffic until the re-tier is resolved.
  • A use case with no tier has no registry record, and therefore no traffic.

Source conflict — the envelope validation rule. Book 3 Appendix D’s validation rule permits execute_autonomously only for T0 and T1, and its enum glosses T2 as “regulated process support” and T3 as “customer-affecting”. Appendix G and Chapter 14 instead give T0 no registered agent, cap T1 at recommend, deny execute_autonomously at T2 and T3, and admit it only at T4 under the exception route. Appendix G states that the Chapter 14 vocabulary is authoritative. This page follows Chapter 14 and Appendix G. Implement the validator as: execute_autonomously denied for T0–T3; permitted for T4 only when an executive exception record with an unexpired review date is attached.

Named approver on every T3 output

This is the T3 control that most often decays, so specify it precisely.

  1. Policy: for any T3 output crossing the agent’s boundary, the policy decision must be the approval verdict, never plain allow.
  2. Join: an approval-queue entry must exist for the same correlation ID. Book 3 Chapter 11 lists its content: approver UPN, approval timestamp, justification text, any edits applied to the agent’s output, and the final approve or reject decision.
  3. Enforcement: the gateway or policy point blocks the outbound action if the approver identity is absent (matrix rows 1–2).
  4. Meaningfulness: the reviewer has the documented interface, time and authority to override; rubber-stamping defeats the control (see the autonomy-creep audit).
  5. Automatic re-tier: if customer_facing changes to true on a T2 use case, it moves to T3 and this control applies before the next customer-bound output.

Where no approval was required, the evidence must reference the policy that says so — never silence.

Approval gates, evidence packs and SLAs

Book 3 Chapter 14 OD-2, a proposal; SLAs are in business days to first decision.

TierApproval bodyEvidence packSLAIf denied
T0Line manager at user onboardingTraining completion; SSO group membershipSame dayRe-propose with adjusted scope
T1ACB, notification onlyEnvelope template; data-classification record; MCP server list10Revise envelope and resubmit
T2ACB, full approvalEnvelope (Mandate, Means, Memory, Monitoring); evaluation-suite plan; identity pair confirmed; data-dictionary entries; wrangler RACI15Revise; if denial is sustained, escalate to the AI Steering Committee with written rationale
T3ACB with a named second-line signatory; AI Steering Committee notifiedAll of T2 plus: FRIA where Annex III applies; Model Risk validation report; meaningful-review design; 6-year (or longer) retention plan; CIF assessment; DORA third-party contract confirmation25Escalate to the AI Steering Committee, which may overturn by majority with CRO concurrence
T4Executive committee with board notificationAll of T3 plus: regulator-engagement plan; risk-appetite statement; CIF assessment accepted at board level; independent legal opinion60No automatic escalation; the decision stands for the year

Each tier decision carries three validators: the wrangler (first line), the DPO and Model Risk lead (second line), and the business owner (first line, accountable for value). The ACB ratifies; at T3 a named second-line signatory countersigns. Names, dates and rationales — not committee membership — are what a supervisor reads. A denial returns named gaps and remediation; most proposals are re-scoped (different tier, narrower data class, held for an evaluation) rather than blocked. The AI Steering Committee is appellate, not first-instance. Size the T3 gate for the throughput the business needs, rather than letting throughput pressure argue a use case down a tier — Book 3’s opening composite turns on exactly that trade.

Tier 4 — the executive exception

Tier 4 is named so that the prohibition is explicit. Book 3’s default answer for the twelve months after publication is no. Items 1–5 of the prohibited-use list can only be reached through this route. Item 6 has no exception route (App G §6), item 7 does not run “regardless of business value” (App G §6), and item 8 restates the provider’s universal prohibitions — none of the three can be reached at all.

The exception procedure (Book 3 App G §6):

  1. Proposal to the AI Steering Committee (App G; Chapter 14 §4.6 sends the proposal to the executive committee — this page follows App G for the entry point and Chapter 14 for the deciding body) with a Model Risk classification, a DORA CIF assessment, an Annex III classification with FRIA where applicable, a regulator-engagement plan for the relevant supervisor, and an explicit risk-appetite decision.
  2. Executive review by the CRO and CEO, with the board risk committee adding independent challenge for customer-affecting cases. Chapter 14 states that the board, not the ACB, takes the decision.
  3. Independent legal opinion where customer impact is significant (autonomous communications, decisions affecting credit, account or trading status).
  4. Time-bound admission: twelve months, with exit criteria — the conditions for returning to T3 with a human restored, or for decommissioning — signed at admission, not at review.

A T4 use case also runs server-side only, keeps legal hold on by default, and reports quarterly to the executive committee and board.

Route each tier to its regulatory questions

Book 3 maps each tier to the regimes the second line walks at each gate. This is a routing summary of the author’s mapping, not a legal determination; the article-level obligations, their dates and their current status are on the regulatory map, and planning dates on its calendar section.

TierAI Act questions to raiseProvider Usage PolicyDORA CIFAutomated-decision (GDPR Art 22)Sectoral
T0Expected minimal-risk; transparency only if output is published externally without redraftingUniversal prohibitions onlyNot CIF-supportingNot engagedInventory only
T1Expected limited-risk; transparency only if content reaches a third partyUniversal prohibitions onlyTypically notNot engaged; DPIA threshold may be crossed by volumeProportional review
T2Limited-risk unless biometrics-adjacent (then re-tier to T3)High-risk finance categories only at the boundary (then re-tier to T3)Depends; a T2 use case supporting a CIF takes the full third-party clause setNot engaged by designPeripheral (conduct, insurer governance)
T3High-risk under Annex III creditworthiness or life/health pricing is likely; FRIA; deployer obligationsHigh-risk categories engaged: human in the loop and AI disclosureFrequently CIF-supporting; Register of Information entry; exit strategyFrequently engaged; meaningful human review is the operative defenceFull sectoral governance (e.g. trading self-assessment, insurer governance and product oversight)
T4Bespoke; executive risk-appetite decisionPresumption against without exceptionCIF by defaultEngaged by defaultEngaged by default

The Usage Policy column is a vendor contract term as of Book 3’s 15 September 2025 reference. [FACT-CHECK: current Anthropic Usage Policy high-risk use-case categories and their human-in-the-loop and disclosure requirements.] Book 3 also maps FINMA materiality for Swiss entities; route that through the regulatory map’s sectoral diagnostic.

Re-tier when the use case changes

A tier fixed at registration ages badly: six months later the wrangler has widened the Means, the agent reads customer-financial data, its output reaches a customer email, and the registry still says T2. Re-tiering is the release valve.

Triggers

Book 3 gives two trigger lists that overlap. Use Appendix G’s six categories as the policy, and Chapter 14 OD-1’s checklist as the wrangler-facing form, published where wranglers work rather than in a policy repository.

App G §5 triggerSignal and sourceChapter 14 OD-1 checklist items it covers
1. Change in data classes touchedNew class in data_classes_permitted, or the gateway DLP check sees a class not in the envelopeData class changes
2. Change in autonomy ceilingAny change to autonomy_ceilingAction class expands; a human-in-the-loop step is removed
3. Change in customer-facing statuscustomer_facing flips; becoming customer-facing forces T3Customer-affecting boundary changes
4. Incident at the current tierMajor ICT incident classification; incident touching automated-decision scope; evaluation drift beyond threshold; tier-relevant SOC near-missEvaluation drift threshold breached two consecutive months
5. New regulatory designationNew CIF designation; new Annex III classification; new national critical-data designation; new CTPP designation touching a consumed routeRegulatory change names the use case
6. Materially different modelChange of model family (e.g. Sonnet ↔ Opus), major version, or a change Model Risk flags as material on the evaluation suiteModel upgrade crosses the change threshold
(OD-1 only)Broker allow-listMCP allow-list gains a server above its current risk class
(OD-1 only)PeopleNamed wrangler or approver changes

Tighten freely, loosen only with approval

  • Tightening (App G’s example: draft → recommend) is permitted with notification only. This page extends the same treatment to removing a data class or adding a human step.
  • Loosening (any move up from the registered value, App G; any reduction in human-in-the-loop scope, Ch 14 §8) requires Virtual CAB approval; App D adds a separate change request, a second-line co-signature and a refreshed risk-tier review for every autonomy change.

This asymmetry is what stops a use case reaching T4 by a series of small T2 amendments.

Process and SLA

The wrangler raises a Virtual CAB ticket; the DPO and Model Risk lead validate; the business owner attests; the ACB decides. The registry logs named signatures and rationale; the envelope updates; the gateway re-syncs; the revalidation cadence moves to the new tier. See change governance for the Virtual CAB.

The two Book 3 sources give different SLAs:

CaseApp G §5Ch 14 OD-1
Non-urgent / envelope change10 working days5 business days (low-risk) or 15 (material)
Incident-drivenSame business day; agent status set to suspended until the re-tier is resolved and the new envelope signed5 business days from incident classification

This page follows both as layers: suspend on the same business day (App G), and decide the re-tier within 5 business days of incident classification (Ch 14). Regulatory incident-reporting clocks run in parallel with the re-tier, not after it.

Revalidation cadence (Ch 14 OD-1)

EventOwnerOutputSLA
Initial classificationWrangler, with DPO and Model Risk inputTier, control package, envelope draft10 business days from intake to ACB agenda
Quarterly revalidation (Book 3 uses the last Tuesday of the quarter)Wrangler attests; ACB reviews exceptionsTier confirmed in the registry; evidence pack updatedOne ACB meeting per quarter
Annual review, aligned with the Register of Information submissionACB chair and second-line signatoriesAnnual tier and control-effectiveness reportOne executive-committee meeting per year

Autonomy-creep audit

The quiet path to T4: a drafting assistant loses its human step for “trivial” cases, then becomes autonomous on low-risk classes, and now operates in T4 territory without ever being approved into it. Run this audit quarterly against the registry:

  1. Approval ratio. For each T2/T3 use case, read the human-approval ratio from the evidence store and flag drift below threshold (Ch 14 §8). Book 3 defines neither the ratio’s formula nor the threshold. Recommendation: define it as boundary-crossing outputs carrying an approval-queue decision over all boundary-crossing outputs, set the threshold per use case at registration, and record both in the envelope.
  2. Envelope drift report. Diff current state against registration state across Means, Memory and Monitoring, alongside the autonomy_ceiling (Book 3 App D). The warning sign App D names: a ceiling unchanged for twelve months while the built-in tools list and the allowed MCP servers list have doubled.
  3. Approver behaviour (this page’s addition). Sample T3 approvals for evidence of meaningful review (edits, rejections, time spent). An approver who never edits or rejects is a finding, not a success.
  4. Outcome. Every finding is a re-tier trigger routed through the Virtual CAB.

Failure modes and recovery

FailureRecognise itRecover
Tier shoppingA proposer argues for the lower tier because the higher gate is “too heavy”Downward burden of proof; minutes record proposed and approved tier; size the T3 gate’s SLA instead of granting a carve-out
No re-tier triggerRegistry tier unchanged while Means and data reach have grownPublished trigger list; Virtual CAB enforcement; gateway refusal on drift
Prohibited-use list out of dateThe ACB debates a prohibition from first principlesAnnual review, 30-day ad-hoc cycle, standing agenda item
Tier mapped to modelSame use case, different tier depending on modelTier on the use case; model chosen inside the tier
Readiness read as safety”That repo is Level 4” in a risk conversationBoth axes on one page with different owners; autonomy changes cite the tier
Autonomy creepApproval ratio falling; envelope Means growing under a static ceilingQuarterly audit; loosening requires Virtual CAB
Retention mislabelled or mis-tieredEnum label disagrees with configured duration; record in the wrong locked containerSingle enum set encoding real durations; route by retention_tier; replay test per tier
Truncation hides the decisive payloadReplay of a T3 action cannot reproduce the responseT3 full payload always to the immutable store
Tier declared, not enforcedA matrix cell has no mechanism or no evidence of last enforcementTreat as a control gap; block approval until evidenced

Acceptance checks

A tiering implementation is ready when all of these hold:

  • The prohibited-use list exists, has a named owner and version, was reviewed in the last 12 months, and the gateway refuses a test request for a listed use case.
  • Every production use case has a registry record with risk_tier, autonomy_ceiling, customer_facing, data_classes_permitted, a CIF flag and signed tier rationale; a request for an unregistered use case is refused.
  • The intake form asks the four questions and records the answers; two wranglers classifying the same sample use case reach the same tier.
  • Each matrix cell for each tier in use names its enforcing mechanism and links evidence of its last enforcement; no cell is weaker than the tier below.
  • The gateway rejects a request whose header class differs from the registered tier, and (at T2+) a request with the wrong identity pair.
  • No T0–T3 envelope validates with execute_autonomously.
  • A T3 boundary-crossing action without an approval-queue entry on the same correlation ID is blocked in a test.
  • Retention values per tier are signed by legal and the DPO, the enum names encode those values, and one locked immutable container exists per tier.
  • The quarterly replay reconstructs a T3 action from its full payload.
  • The last quarterly autonomy-creep audit ran, with approval ratios and drift reports per T2/T3 use case.
  • Re-tier tickets show loosening changes carrying Virtual CAB approval and tightening changes carrying notification.

Provenance and refresh

The taxonomy, matrix, SLAs, trigger lists and exception procedure are Book 3’s proposals (Chapter 14, Draft 2 of 3 August 2026; Appendix G, Draft 1 as of 14 May 2026), synthesised here with Book 2’s separation of readiness, agency class and risk tier, and Research Paper 1’s descriptive attributes (§4.4). Where the sources conflict — retention values, envelope validation, re-tier SLAs, the list’s approving body, the policy-decision verdict name — this page states which it follows and why. Recommendations marked as this page’s are the editor’s, not the books’. Refresh when the books change, when any regulatory designation in the routing table changes (via the regulatory map), and when vendor product claims marked [FACT-CHECK] are verified or overtaken.

Source material

  • Book 3 (Deploying Claude Safely), Chapter 14, Risk Tiers and the Use-Case Taxonomy (Draft 2, 3 August 2026)
  • Book 3, Appendix G, Risk-Tier Control Matrix (Draft 1, as of 14 May 2026)
  • Book 3, Chapter 2, Why Claude Becomes the Default (use-case record, prohibited-use list, portability)
  • Book 3, Chapter 6, The LLM Gateway and Policy Decision Point (use-case class header, policy stacks, policy decisions)
  • Book 3, Chapter 11, Evidence Capture and the Audit Schema (PR-3 storage and retention table; truncation recovery)
  • Book 3, Appendix D, Bounded-Agency Envelope Template (risk_tier, autonomy_ceiling, customer_facing; autonomy-creep drift report)
  • Book 3, Appendix E, Audit Evidence Schema for Claude Actions (§5 storage and retention; truncation policy)
  • Book 2 (The AI Delivery Factory), Chapter 4, Agent Readiness: Scoring the Estate (readiness routes the lane, risk tier authorises the autonomy)
  • Book 2, Chapter 22, The Shape of an AI Product (classification line, least-agency rule, consequence check)
  • Book 1 (Building Organizations for the Age of AI), Appendix B, Operating Schemas: AES, Control Plane, Model Usage, Data Dictionary, and Change Dossier (model usage registry regulatory_class, assigned per usage)
  • Research Paper 1, A Taxonomy for a Reuse-Driven AI Factory, §4.4 Descriptive attributes

← All AI Delivery Factory guidance

On this page

  1. The operating rule
  2. Step 1 — Check the prohibited-use list before the taxonomy
    1. The list (Book 3 Chapter 14 §6, as of August 2026)
    2. Owning and maintaining the list
  3. Step 2 — Ask four intake questions under least agency
  4. Step 3 — Place the use case in a tier
    1. The five tiers
    2. Boundary tests
    3. Classify by harm, not by role, model or readiness
  5. Step 4 — Write the use-case record
  6. Step 5 — Apply the tier’s control package
    1. The twelve control areas
    2. Read the matrix two ways, and prove every cell
  7. Retention and payload truncation by tier
    1. What the sources say — and where they disagree
    2. Rule this page follows
  8. Enforce the tier at runtime
    1. Gateway
    2. Connector broker
    3. Agent registry
    4. Named approver on every T3 output
  9. Approval gates, evidence packs and SLAs
  10. Tier 4 — the executive exception
  11. Route each tier to its regulatory questions
  12. Re-tier when the use case changes
    1. Triggers
    2. Tighten freely, loosen only with approval
    3. Process and SLA
    4. Revalidation cadence (Ch 14 OD-1)
    5. Autonomy-creep audit
  13. Failure modes and recovery
  14. Acceptance checks
  15. Provenance and refresh

All pages in Design →