High-Tech & Software
Continuous Technical Documentation Generation
Minimal RiskUnverifiedDiscuss / dispute
Codebase, telemetry and training metadata parsed continuously to auto-generate and maintain Annex IV technical documentation, model cards and change histories for deployed AI systems.
Consensus classification rationale: The output is the evidence other systems are judged on, which inverts the risk: a fluent but unverified dossier is worse than none, because Art. 11 documentation is what the market-surveillance authority inspects. Lead-engineer verification of each Annex IV disclosure, plus lineage back to the artefact that produced each claim, is non-negotiable.
Indicative decision support, not legal advice. Risk classification depends on your concrete deployment context and can change with scope drift — validate the result with qualified counsel.
pricingPer generated technical compliance dossier
oversightLead AI engineer verifies every Annex IV disclosure before the dossier is released
Target market(s)European UnionUnited States (federal)change
Changes which instruments below count as in scope for this profile.
Target markets: European Union, United States (federal)
Regulatory footprint
5 instruments across 3 of 7 regulatory domains, plus 12 standards references- AI law1 instrument
- Data protectionnone triggered
- Cyber & resilience1 instrument
- Online safety & platformsnone triggered
- Product safetynone triggered
- Financial servicesnone triggered
- Sector & employment3 instruments
- Standards12 references
By jurisdiction
- EU5European UnionArt. 11 — Technical Documentation, Art. 12 — Record-Keeping / Logging, Art. 17 — Quality Management System, Cyber Resilience Act, EU AI Act
The AI Act is one dimension of this footprint, not the whole of it — every domain above carries its own obligations and deadlines. See the instruments in the graph →
Confidence in this chain of evidenceConfidence: Check-worthy
The chain holds, but at least one hop rests on a secondary source, an ageing verification or a practice-derived step. Check the flagged hops before you rely on them.
Computed weakest-link over 20 evaluated hops across 1 target market: a chain is only as strong as its weakest step, so the band follows the worst hop rather than an average that would hide it. Five factors per hop — source tier, verification age, status certainty, community hardening, derivation kind — all read from graph data, never from a hand-set score.
Why this band12 factors lowered the band — each links to the claim behind it
- Source tier: JTC 21 Technical Package (prEN 18228/18229/18281–83) rests on a secondary source (tracker or summary), not on the primary text. open node → primary source →
- Source tier: IEEE CertifAIEd™ carries no resolvable citation — the claim is uncited. open node →
- Source tier: prEN 18229-1 (Trustworthiness Framework, part 1) rests on a secondary source (tracker or summary), not on the primary text. open node → primary source →
- Source tier: ENISA Multilayer Framework & AI Threat Landscape carries no resolvable citation — the claim is uncited. open node →
- Source tier: NIST SP 800-218 (SSDF) carries no resolvable citation — the claim is uncited. open node →
- Source tier: TAGOF (Audit-as-Code) carries no resolvable citation — the claim is uncited. open node →
- Status certainty: JTC 21 Technical Package (prEN 18228/18229/18281–83) is "draft", not settled in-force law. open node → primary source →
- Status certainty: prEN 18229-1 (Trustworthiness Framework, part 1) is "enquiry", not settled in-force law. open node → primary source →
- Verification age: IEEE CertifAIEd™ has no recorded verification date. open node →
- Verification age: ENISA Multilayer Framework & AI Threat Landscape has no recorded verification date. open node →
- Verification age: NIST SP 800-218 (SSDF) has no recorded verification date. open node →
- Verification age: TAGOF (Audit-as-Code) has no recorded verification date. open node →
Compliance brief
This use case is minimal-risk under the EU AI Act (Minimal Risk); no product-specific obligations beyond general AI literacy apply.
What is owed
- Art. 4. Providers and deployers must ensure sufficient AI literacy of staff dealing with AI systems.
- Art. 25. A deployer becomes the provider (full Art.
Dates that bind
- 2024-08-01 — AI Act enters into force. Regulation (EU) 2024/1689 in force; countdown for all staged obligations starts.
- 2025-02-02 — Prohibitions + AI literacy. Art. 5 prohibited practices ban applies (manipulation, social scoring, untargeted face scraping, workplace emotion recognition); Art. 4 AI literacy duty.
Maximum exposure
- EU AI Act: Tiered: €35m / 7% (prohibited practices); €15m / 3% (Art. 9–15 high-risk obligations incl. data governance, documentation, logging); €7.5m / 1% (Art. 99(5) — incorrect, incomplete or misleading information to notified bodies or national competent authorities)
- Cyber Resilience Act: Up to €15m or 2.5% of worldwide annual turnover
First five actions
- Confirm in writing whether this organisation builds/places the system on the market (provider) or only operates it (deployer), since the role is not yet established.
- Commission and confirm the Art. 4, Art. 25 obligations named above as active workstreams with an accountable owner.
- Stand up the named oversight design — Mode 1 — with a documented human-review procedure.
- Produce the technical documentation and evidence artefacts already mapped to this use case (Zero-Data-Retention Vendor Binding, Agent Discovery & Registry Endpoint, SBOM & Dependency Management) before they are requested.
- Put 2024-08-01 — AI Act enters into force — into the compliance calendar with an owner and lead time.
Terms used above: · · ·
Classification precedent
Consensus reading: Minimal Risk open in the graph →
The output is the evidence other systems are judged on, which inverts the risk: a fluent but unverified dossier is worse than none, because Art. 11 documentation is what the market-surveillance authority inspects. Lead-engineer verification of each Annex IV disclosure, plus lineage back to the artefact that produced each claim, is non-negotiable.
What the reading rests on — the provisions this classification actually pulls in:
- Art. 4 — AI Literacy
- EU AI Act (Regulation (EU) 2024/1689)
- Cyber Resilience Act (Regulation (EU) 2024/2847)
- Art. 11 — Technical Documentation
- Art. 17 — Quality Management System
- Art. 12 — Record-Keeping / Logging
No dissenting reading is recorded for this case. That means nobody has filed one yet — not that the classification is beyond argument. file a dissent with a source →
Baseline: of 100+, 40% were not definitively classifiable (18% clearly high-risk, 42% clearly low-risk). appliedAI Institute — AI Act risk classification of AI systems from a practical perspective
Applicable Regulations (5)
EU AI Act (Regulation (EU) 2024/1689)
unverified · verified 2026-08-12 source (as amended) EUR-LexAmended by Regulation (EU) 2026/1744. Verified 16 Aug 2026: the popular mirrors have not yet been updated — artificialintelligenceact.eu still serves the unamended 13 June 2024 text with no disclaimer, and the Commission's AI Act Service Desk pages still show pre-omnibus text with a visible omnibus disclaimer. Read the OJ or consolidated text on EUR-Lex. in force EU
Horizontal, risk-based product-safety law for AI systems and GPAI models. Extraterritorial market-place principle. Staged applicability 2025–2030 (Digital Omnibus: Art. 50 → 2 Aug 2026, Annex III → 2 Dec 2027, Annex I → 2 Aug 2028). (Digital Omnibus: Regulation (EU) 2026/1744, in force 27 July 2026).
Sanctions: Tiered: €35m / 7% (prohibited practices); €15m / 3% (Art. 9–15 high-risk obligations incl. data governance, documentation, logging); €7.5m / 1% (Art. 99(5) — incorrect, incomplete or misleading information to notified bodies or national competent authorities)
Cyber Resilience Act (Regulation (EU) 2024/2847)
Security-by-design for products with digital elements over the full lifecycle: vulnerability management, patching, SBOM. Complements AI Act Art. 15 at product level.
Sanctions: Up to €15m or 2.5% of worldwide annual turnover
Art. 11 — Technical Documentation
Annex IV technical file before placing on market: system description, architecture, capabilities/limitations, risk measures — kept up to date.
Art. 17 — Quality Management System
Product-focused QMS for providers: strategy, design controls, data management, post-market monitoring — target of EN 18286:2026, the first AI Act harmonised-standard candidate to be published; presumption of conformity applies only once it is cited in the OJEU, which is still pending.
Art. 12 — Record-Keeping / Logging
Requires logging CAPABILITY over the system's lifetime, recording events relevant to identifying situations that may present an Art. 79(1) risk or lead to a substantial modification, and to post-market monitoring (Art. 72) and deployer monitoring (Art. 26(5)). SCOPE: the itemised minimum log content of Art. 12(3) — period of each use with start and end date and time; the reference database against which input data was checked; the input data for which the search led to a match; the identification of the natural persons involved in verification of the results per Art. 14(5) — applies ONLY to Annex III point 1(a) remote biometric identification systems, not to all high-risk systems. RETENTION: providers (Art. 19) and deployers (Art. 26(6)) must keep the logs under their control for a period appropriate to the intended purpose, at least six months, unless other Union or national law — in particular data-protection law — provides otherwise. A bolted-on logging wrapper does not satisfy the requirement: logging must be core architecture.
Legal Obligations (2)
density
Art. 4 — AI Literacy
Providers and deployers must ensure sufficient AI literacy of staff dealing with AI systems. In force since 2 Feb 2025.
unverified · no verification date source (as amended) EUR-Lexconvenience mirror — not updated artificialintelligenceact.euAmended by Regulation (EU) 2026/1744. Verified 16 Aug 2026: the popular mirrors have not yet been updated — read the OJ or consolidated text on EUR-Lex.
Art. 25 — Value Chain / Role Flip
A deployer becomes the provider (full Art. 8–17 duties) by re-branding, changing intended purpose, or making a substantial modification — e.g. deep fine-tuning or wiring a model into autonomous agent toolchains.
unverified · no verification date read the article artificialintelligenceact.eu
Control Objectives (0)
obligation (article) → operationalized_by → control objective → satisfied_by → component/pattern; control objective → evidenced_by → evidence artifact
Art. 4
control layer: community mandate — propose objectives
Art. 25
control layer: community mandate — propose objectives
Take this into your GRC tooling
A control mapping your ISO/IEC 42001 or CSA AICM workbook can ingest, and an Annex IV skeleton to start the technical file from. Indicative mappings only — cells we are not confident about are exported empty rather than filled in.
Standards & Evidence
JTC 21 Technical Package (prEN 18228/18229/18281–83)
CEN-CENELEC JTC 21 technical package under standardisation request M/593 (prEN 18228 trustworthiness, 18229 risk management, 18281–83 CV/NLP evaluation et al.); staged drafts, none OJEU-cited yet — Annex III applicability (Dec 2027) is Omnibus-coupled to their availability.
evidence for: EU AI Act
IEEE CertifAIEd™
Ethics certification (transparency, accountability, algorithmic bias, privacy) for products and professionals; interfaces with the EU ALTAI assessment list.
unverified · no verification date
evidence for: EU AI Act
prEN 18229-1 (Trustworthiness Framework, part 1)
Part 1 of the JTC 21 trustworthiness deliverable — the framework layer other prEN 18xxx documents build on.
evidence for: EU AI Act
ENISA Multilayer Framework & AI Threat Landscape
Three-layer good-practice model (cyber foundations → AI-specific → sectoral) and lifecycle threat landscape — the operational base for Art. 15 and CRA.
unverified · no verification date
evidence for: Cyber Resilience Act
NIST SP 800-218 (SSDF)
Secure Software Development Framework: practices for provenance, review and vulnerability handling of generated and third-party code; SSDF-AI companion covers AI-assisted development.
unverified · no verification date
evidence for: Cyber Resilience Act
ISO/IEC 27001:2022 + A.8.28
Information-security management; control A.8.28 (secure coding) is the natural anchor for AI code-generation and QA workflows alongside ISO 42001.
unverified · no verification date publisher ISO
evidence for: Cyber Resilience Act
EN ISO/IEC 22989 (AI Concepts)
Published terminology and concepts standard — the shared vocabulary layer for documentation and audits.
unverified · no verification date publisher ISO
evidence for: Art. 11 — Technical Documentation
EN 18286:2026 (QMS for AI Act)
Harmonised-norm candidate translating Art. 17 QMS into a product-focused governance framework; mappings to ISO 9001 and ISO/IEC 42001 Annex A (Annexes C & D); published as EN 18286:2026 in July 2026, OJEU citation (and with it the presumption of conformity) still pending.
evidence for: Art. 17 — Quality Management System
ISO/IEC 42001:2023 (AIMS)
ISO/IEC 42001:2023 — certifiable AI management system (Annex SL harmonized structure, PDCA logic, synergy discount when an ISO 27001 ISMS exists). Clauses 4–10 plus Annex A controls (control count 38 vs 39 is a live community dispute — counting method differs by edition/guide). Covers an estimated 40–50% of AI Act organizational duties; organizational certificate, no product presumption of conformity.
evidence for: Art. 17 — Quality Management System
IEC 62304:2006+AMD1:2015 (Medical device software life cycle)
Medical device software — software life-cycle processes: software safety classification (A/B/C), development planning, architecture, unit verification, integration and system testing, release, maintenance and problem resolution, and management of SOUP/off-the-shelf components. The recognised life-cycle spine for MDR software, and the process framework a notified body expects an AI-based diagnostic to be built inside.
evidence for: Art. 17 — Quality Management System
FprEN ISO/IEC 24970 (AI Logging)
Specifies event logging in AI systems — the concrete implementation target for Art. 12 record-keeping.
evidence for: Art. 12 — Record-Keeping / Logging
TAGOF (Audit-as-Code)
Operationalizes governance as code in CI/CD: policy-as-code enforcement, continuous runtime telemetry and automatically generated audit evidence — the execution layer that replaces periodic audits with continuous assurance.
unverified · no verification date
evidence for: Art. 12 — Record-Keeping / Logging
Evidence you will need (12)
The concrete deliverables this use case's obligations ask for — grouped by what kind of artifact they are. Documentation is the largest single conformity cost block, so the list is a work plan, not a reading list. Full evidence matrix →
Documents & files (5)
Written deliverables an authority or auditor can request as a file.
AI Bill of Materials (AI-BOM) & Factsheetspractice-derived — dispute welcomeserves 2 obligations
Machine-readable composition manifest per AI application: base-model metadata (identifier, version, parameters, supplier tag), dataset provenance (fine-tune/RAG sources, scrubbing logs, consent records), vector-namespace bindings and access rules, active runtime-policy ruleset versions and thresholds, performance & safety verification history (bias scores, accuracy benchmarks, red-team reports). Complements the SBOM (software dependencies) with the AI-specific supply chain; auto-published into the register on every change. Feeds the Art. 11 technical file, vendor due diligence (contractually demanded from providers) and Colorado/LL144-class disclosure duties. Factsheets are its human-readable projection for auditors.
verifiability: tamper-evident
chain: Art. 11 — Technical Documentation · Art. 26 — Deployer Obligations → CO: Embedded-AI Vendor Governance · Continuous Technical Documentation Generation
AI System Model Cardpractice-derived — dispute welcomeserves 2 obligations
Model lineage, architecture, pre-training data sources, context limits, evaluation benchmarks and known failure modes.
verifiability: tamper-evident
chain: Art. 11 — Technical Documentation · Art. 13 — Transparency to Deployers → CO: System Traceability & Decision Transparency · Algorithmic Portfolio Execution & Advisory · Automated Financial Forecasting & Audit Trails · Continuous Technical Documentation Generation · Enterprise SDLC Code Automation & QA · +1 more
SBOM & Vulnerability Management Recordspractice-derived — dispute welcomeserves 2 obligations
CRA evidence: software bill of materials incl. model weights and datasets, vulnerability handling and patch history — reused for NIS2 supply-chain security and DORA third-party registers.
verifiability: documented artefact — verifiable on inspection
chain: Cyber Resilience Act · NIS2 Directive · Enterprise SDLC Code Automation & QA
Technical Documentation (Annex IV)text-derivedserves 2 obligations
The Art. 11 technical file: system description, architecture, capabilities/limitations, risk measures, development process. Built incrementally during development — retro-reconstruction is an audit red flag. Reviewed by a notified body where the Annex VII route applies.
verifiability: self-asserted
chain: Art. 11 — Technical Documentation · Art. 43 — Conformity Assessment · Clinical Imaging Triage & Patient Follow-Up · Continuous Technical Documentation Generation
QMS Documentation (Art. 17 / EN 18286:2026)text-derived
Documented quality management system: design controls, development testing, validation, supplier management, post-market processes — the documentation set that EN 18286:2026 is expected to address once the standard is cited in the OJEU.
verifiability: documented artefact — verifiable on inspection
chain: Art. 17 — Quality Management System
Assessments (2)
A structured judgement about risk, rights or a management system.
FRIA / AI Impact Assessment (AIIA)text-derivedserves 3 obligations
Fundamental-rights impact assessment (Art. 27, deployer-side) generalized to the AI Impact Assessment: societal, legal and operational risk evaluation per ISO/IEC 42005 and ISO 42001 Clause 8.2, defining HITL intervention parameters and acceptable-use bounds. Cadence: pre-deployment, refreshed annually and on major model updates — a stale AIIA is a finding, not a document.
verifiability: documented artefact — verifiable on inspection
chain: Art. 26 — Deployer Obligations · Art. 27 — Fundamental Rights Impact Assessment · EU AI Act · Clinical Imaging Triage & Patient Follow-Up
Third-Party AI Data & ZDR Certificatepractice-derived — dispute welcome
Binding vendor terms on zero data retention, non-training use, sub-processor list and security boundary, with technical verification records.
verifiability: independently-attested
chain: Art. 25 — Value Chain / Role Flip · Generative Asset Production & Virtual Try-On
Log records (3)
Machine-generated records produced while the system runs.
Event Logs & Decision Tracestext-derivedserves 17 obligations
The single highest-leverage artifact: hash-chained, WORM-stored logs with structured decision traces. Required capability fields per FprEN ISO/IEC 24970: input/output traces, execution timestamps, acting user/agent identity, referenced sources, human overrides. Audit-packet spec per event: model version, system-prompt/context hash, hyper-parameters (temperature, top-p), output payload, confidence score, active policy-ruleset versions, human override record. Simultaneously serves AI Act Art. 12, GDPR accountability, DORA incident reporting, NIS2 logging, PLD disclosure duties and its rebuttable defect presumption; financial-sector regimes push retention to 7 years (SEC 17a-4-class WORM rules). Credibility bar: anchor hash-chain heads externally (qualified timestamp / eIDAS ledger) so integrity survives an insider with admin rights.
verifiability: externally-anchored
chain: Art. 12 — Record-Keeping / Logging · CRA Art. 14 — Vulnerability & Severe-Incident Reporting · DORA Art. 19 — Major ICT-Incident Reporting · GDPR Art. 33/34 — Personal-Data Breach Notification · HIPAA Breach Notification Rule · NIS2 Art. 23 — Significant-Incident Reporting · +21 more
Guardrail Telemetry & Sanitization Recordspractice-derived — dispute welcomeserves 3 obligations
Control-level evidence for the OWASP mappings: guardrail trigger records, blocked-prompt statistics (LLM01), runtime output-sanitization logs (LLM05), groundedness-check outcomes — the empirical proof that declared controls actually execute.
verifiability: tamper-evident
chain: Art. 15 — Accuracy, Robustness, Cybersecurity · Art. 50 — Transparency Duties → CO: AI Interaction & Content Disclosure · Art. 15 — Accuracy, Robustness, Cybersecurity → CO: Runtime Injection Defense · Dynamic Deal Desk & Quoting Engine · Enterprise Marketing Disclosure Compliance · LLM01 Prompt Injection · +3 more
Immutable Decision Ledger (WORM)practice-derived — dispute welcomeserves 3 obligations
Per-execution audit packet: timestamp, model version, system prompt, input-context hash, hyper-parameters, output payload, confidence score and human override record.
verifiability: externally-anchored
chain: Art. 12 — Record-Keeping / Logging · Art. 12 — Record-Keeping / Logging → CO: Log Integrity & Non-Repudiation · IFRS / US GAAP Reporting Assurance · Algorithmic Portfolio Execution & Advisory · Automated Financial Forecasting & Audit Trails · Cross-Border Statutory Tax & Wealth Filing · +3 more
Process records (1)
Traces that a process actually happened, and who did it.
AI Literacy Training Recordspractice-derived — dispute welcomeserves 2 obligations
Art. 4 evidence: role-based training curricula and completion records for staff dealing with AI systems — the one obligation that applies at every risk level.
verifiability: documented artefact — verifiable on inspection
chain: Art. 4 — AI Literacy · Art. 14 — Human Oversight → CO: Oversight Competence & Authority
Registry entries (1)
An entry in a register — internal inventory or public registry.
AI System Inventory / Registry Entrypractice-derived — dispute welcomeserves 3 obligations
The organizational register of AI systems in use — role (provider or deployer), classification, owner, vendor and lifecycle state — from which per-system obligations are assigned.
verifiability: self-asserted
chain: Art. 17 — Quality Management System · Art. 26 — Deployer Obligations · Art. 26 — Deployer Obligations → CO: Embedded-AI Vendor Governance
Architecture Blueprint
Agentic RDA Stack (6 Layers)
Regulatory Design & Architecture framework for agentic systems: (1) isolated ephemeral execution (gVisor/Firecracker, read-only root, egress allowlists) → (2) agentic zero-trust identity (per-agent ID, short-lived OBO OAuth 2.1 tokens) → (3) MCP gateway with default-deny ACLs & credential vault → (4) reasoning + guardrail interception (prompt shields, sidecar alignment checks) → (5) human-oversight & durable state persistence (propose-action objects, checkpoint store) → (6) continuous observability & signed audit logs (≥ 6 months, SIEM).
Mode 1 — Assistant (HITL)
Agent proposes, human disposes: every consequential action reviewed before execution. Default for first deployments, irreversible or legally sensitive actions.
Required Technical Components (21)
Zero-Data-Retention Vendor Binding
Sensitive inference is contractually and technically restricted to endpoints under zero-data-retention and non-training terms, evidenced per vendor and re-validated annually.
from: Art. 25
Agent Discovery & Registry Endpoint
The marketplace/discovery API through which external agents find, authenticate against and transact with your agents: published capability descriptors, counterparty authentication, per-counterparty rate and value limits, and a resolvable record of which external principal initiated which transaction. Without it, business-to-agent traffic is anonymous inbound automation.
from: Art. 25
SBOM & Dependency Management
Software bill of materials incl. model weights and datasets; automated vulnerability patching pipeline.
from: Cyber Resilience Act
Secure Boot & Hardened Runtime
Verified boot chain and hardened runtimes for edge/IoT deployments per CRA security-by-design.
from: Cyber Resilience Act
Ephemeral Execution Isolation
gVisor/Firecracker microVMs, read-only root, egress allowlists; container discarded after each task to prevent persistence of exploits.
from: Agentic RDA Stack (6 Layers)
Agentic Zero Trust
Unique cryptographic identity per agent; short-lived, finely-scoped tokens (OAuth 2.1 + PKCE); On-Behalf-Of flow so an agent can never see more than its triggering user.
from: Agentic RDA Stack (6 Layers)
MCP Gateway / Proxy
Central chokepoint for agent tool traffic: default-deny tool ACLs (tools/list vs tools/call), schema & argument inspection, credential injection from vault, rate limits, full audit mirror. Regulatory root cause: the base MCP protocol enforces no authentication or authorization at protocol level — Host/Client/Server topology with Tools/Resources/Prompts primitives ships without an identity layer, so a policy-enforcing gateway is not optional hardening but the only place Art. 12/15 duties can be enforced for tool calls.
from: Agentic RDA Stack (6 Layers)
Guardrail Sidecar / Interception
Rule-based (NeMo/Colang), model-based (alignment checkers) and structural validators deployed as sidecar or gateway plugin (<50 ms), decoupling safety scaling from inference scaling.
from: Agentic RDA Stack (6 Layers)
Propose-Action Objects
Agents never call target APIs directly: they emit typed proposal objects (endpoint, params, risk estimate, rationale) validated by the governance layer before execution; idempotent execution layer.
from: Agentic RDA Stack (6 Layers)
Durable Checkpointing (Pause & Resume)
At oversight gates the complete operational state — working memory, conversation history, tool arguments, intermediate artifacts — is serialized into a durable checkpoint (fast KV store for sub-ms lookups, transactional backend as recovery anchor, vector store for semantic caching of past human decisions). On approval the agent deserializes and resumes at the exact step; matched precedents can shortcut re-planning entirely.
from: Agentic RDA Stack (6 Layers)
Per-Action Autonomy Tiering
Tools tagged read-only / reversible-write / irreversible-write; controls layer routes each action to the matching oversight mode. Mode selection is per action type, never per agent.
from: Agentic RDA Stack (6 Layers)
Agent Identity & Access (IdP)
Per-agent identities, short-lived scoped tokens, OBO flow enforcement — the identity substrate of agentic zero trust.
from: Agentic RDA Stack (6 Layers)
Central Credential Vault
Agents never hold target-system keys; the gateway injects centrally managed credentials after policy checks.
from: Agentic RDA Stack (6 Layers)
OpenTelemetry / FCoT Tracing
Hierarchical trace spans for every sub-task, prompt, retrieved document and API call — the reconstructible decision path for Art. 12/14 and PLD disclosure.
from: Agentic RDA Stack (6 Layers)
WORM / Immutable Audit Vault
Append-only, hash-chained audit vault (WORM object-lock storage, AES-256 at rest, TLS 1.3 in transit). Guarantees tamper-evidence within the organization's trust domain — which stops your own team, but not an admin who can rebuild the vault. Pair with an external trust anchor and key ceremonies outside the operating team for evidence that holds against the insider scenario.
from: Agentic RDA Stack (6 Layers)
Watchdog Supervisor & Rate Limiting
Cost/iteration caps, loop detection, anomaly-triggered mandatory approval (CodeBuddy 'suspicious command override').
from: Agentic RDA Stack (6 Layers)
Trinity Defense (TCB + Command Gates + IFC)
Treats the LLM as an untrusted proposal engine behind a hardened non-LLM Trusted Computing Base. Three pillars: (1) command gates — actions only via a Finite Action Calculus, authorized by a deterministic policy checker before any execution; (2) information-flow control — lattice labels stop confidential data flowing to low-trust sinks without audited declassification; (3) privilege separation — sandboxed low-privilege planner ingests untrusted input, isolated high-privilege worker executes only gate-approved, TCB-normalized actions. Grounded in the impossibility result: token content alone can never unforgeably separate commands from data.
from: Agentic RDA Stack (6 Layers)
Guardian Agents (Runtime Policy Enforcement)
Autonomous supervisory agents outside the supervised agent's reasoning loop: stateful threat engines with graph-based cross-session history (catch multi-turn injection, gradual exfiltration, incremental privilege escalation), event-driven exposure visibility (permission drift, new connectors), and contextual risk correlation into unified issues — interception before execution, not post-hoc logging.
from: Agentic RDA Stack (6 Layers)
Deterministic Policy Engine (OPA / Cedar)
Policy-as-code decision point (PDP) with enforcement points (PEP) in front of every tool call: versioned policies in Git, microsecond evaluation, typed action schemas — authorization decided outside the model's reasoning space, never in the prompt.
from: Agentic RDA Stack (6 Layers)
Shadow-Mode Execution
Run governance controls in observe-and-score mode before enforcement: the policy engine and guardrails evaluate every agent action and log verdicts without blocking, yielding empirical false-positive/negative rates and calibrated thresholds. De-risks the enforcement cutover, produces baseline evidence for Art. 9 risk estimation, and is the standard migration path when retrofitting controls onto a live workflow.
from: Agentic RDA Stack (6 Layers)
Multi-Model Router & Fallback Abstraction
Abstraction layer decoupling application logic from model providers: dynamic routing on capability, cost, latency SLA and regulatory constraint (sensitive-data classes pinned to ZDR private/VPC endpoints or on-prem open-weight instances); real-time health monitoring with automatic fallback to secondary endpoints or local fine-tuned models on outage/latency spikes. Discharges resilience duties (DORA-class), prevents provider lock-in, and makes model deprecations a routing-table change instead of a re-architecture. Router decisions are logged into the decision trace — model version per event is an audit-packet field.
from: Agentic RDA Stack (6 Layers)
Delivery Stack & Pipeline Stage (8)
Service-as-a-Software delivery: the engines, patterns and artifacts this workflow needs on top of the generic obligations. See the full pipeline
Document Intelligence Engine
OCR, layout parsing and semantic clause extraction over filings, contracts and invoices, emitting structured records with span-level source references.
Data Lineage & Versioning
Provenance tracking of datasets, features and embeddings; write-time attribution (source, actor, timestamp, confidence).
AI Register & Model Registry / Factsheets
AI register & model registry: central inventory of every model, agent, RAG pipeline and embedded third-party SaaS AI across the estate, with factsheets per asset. v2.0 duty: every application — internal, open-source or procured — continuously publishes a machine-readable AI-BOM and Factsheet into the register; an asset without a current AI-BOM is an inventory gap, not a formality. Feeds Colorado AIA/ LL144 disclosure duties and the Art. 11 technical file; the enforcement backstop is Shadow-AI discovery on the risk register.
OpenTelemetry / FCoT Tracing
Hierarchical trace spans for every sub-task, prompt, retrieved document and API call — the reconstructible decision path for Art. 12/14 and PLD disclosure.
SBOM & Dependency Management
Software bill of materials incl. model weights and datasets; automated vulnerability patching pipeline.
Cognitive Orchestrator
The reasoning and control plane of an agentic workflow: goal decomposition, tool selection across enterprise APIs, confidence scoring per step, and a human-machine interface exposing progress, limitations and a global halt. It is the architectural home of AI Act Art. 14 oversight — oversight that lives only in a downstream UI cannot stop an executing agent.
Propose-Action Objects
Agents never call target APIs directly: they emit typed proposal objects (endpoint, params, risk estimate, rationale) validated by the governance layer before execution; idempotent execution layer.
Bitemporal Memory (GDPR×Art.12)
valid_from/valid_to + transaction time on every record: GDPR erasure removes data from the active retrieval path while the HMAC-chained immutable log survives for Art. 12 / PLD defence; tenant-scoped partitions allow physical scrub of PII.
Build or Buy — Vendor Layer (9)
The graph models vendor CATEGORIES as first-class nodes and keeps named vendors as community-maintained, disputable desc content with lastVerified dates. A category is stable; a vendor list is a currency-layer object like any standard node.
Agentic Execution Governance
The youngest tier: governance of what an agent is allowed to do at execution time — non-human identity, per-task scoping, action approval, agent inventory and agent-level red-teaming. Named products live in marketExamples; prose here describes the class. Because the category is new, capability claims outrun deployments: ask for a reference in your own regime before believing a control is covered, and treat entries with limited public verification as unconfirmed. Selection metrics: see meta.marketLandscape.selectionMetrics.agentgov.
unverified · verified 2026-08-18 community-maintained
selection metrics: non-human identity inventory completeness, credential time-to-live and revocation latency, per-action approval hooks, agent-level red-team coverage, evidence export a 2nd line can read, deployment references in your regime
supplies: Agent Discovery & Registry Endpoint · Agent Identity & Access (IdP)
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| Pillar Security | agent security & inventory | Discovery, inventory and runtime policy for agents in the estate. Typical: agent registry, policy enforcement. | not checked | agent-inventory positioning |
| Lyzr | agent governance & observability | Agent platform with governance, approval and observability features. Typical: agent approval, agent analytics. | not checked | vendor-stated security posture |
| Astrix Security | non-human identity | Lifecycle governance of machine and agent identities and their grants. Typical: credential scoping, NHI inventory. | not checked | SOC 2 (claimed)NHI governance positioning |
| Britive | just-in-time access | Ephemeral, per-task privileges instead of standing credentials. Typical: JIT credentials, privilege reduction. | not checked | SOC 2 (claimed)least-privilege positioning |
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
AI Supply-Chain Security & AIBOM
Bills of materials for AI: base architecture, training-data dependencies, fine-tuning history and licence lineage of a model, plus scanning of third-party pretrained weights and model artifacts for backdoors, poisoning and tampering before ingestion. Adjacent to, but not the same as, software SBOM tooling — the unit of analysis is a weights artifact and its provenance. Selection metrics: see meta.marketLandscape.selectionMetrics.supplychain.
unverified · verified 2026-08-17 community-maintained
selection metrics: Whether the AIBOM records training-data and fine-tuning lineage or only package dependencies; artifact formats scanned (safetensors, pickle, GGUF, container images); detection basis for tampering and poisoning (signature, behavioural, provenance attestation) and its false-positive rate; support for signing and verifying weights in your own pipeline; whether ingestion can be blocked, not just reported.
supplies: SBOM & Dependency Management
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| Cranium AI | AIBOM & model provenance | AI bill-of-materials generation, model-provenance capture and third-party model risk scanning. Typical: AIBOM, third-party model ingestion. Scope overlap: Its AI-governance reporting scope overlaps this platform's own; we have a commercial interest in the comparison. | SaaS (vendor cloud) | NIST AI RMF alignment (claimed)EU AI Act readiness positioning |
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
Runtime Security & Guardrail Vendors
First-line inline enforcement: single-pass parallel input/output evaluation proxies, injection and exfiltration defense, PII masking, grounding checks, SecOps routing. Named products live in marketExamples; prose here describes the class. What the class buys you: a policy decision point in the request path that fails closed and emits telemetry an auditor can read. Selection metrics: single-pass latency (<20 ms class), catch rates, policy-version telemetry into the AI-BOM. Consolidation matters commercially: a guardrail acquired by a platform vendor tends to follow that platform's roadmap, which is a lock-in question rather than a security one — reported acquisitions are recorded per entry as reported, not asserted here.
unverified · verified 2026-08-18 community-maintained
selection metrics: single-pass parallel evaluation latency (<20 ms class), injection/hallucination catch rates, SecOps/SIEM routing, policy versioning surfaced into the AI-BOM
supplies: MCP Gateway / Proxy · Guardrail Sidecar / Interception
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| Lakera | guardrail proxy | Inline prompt-injection and content detection at request time. Typical: injection defence, content filtering. | not checked | SOC 2 (claimed)supports Art. 15 robustness measures (claimed) |
| HiddenLayer | model/agent detection & response | Model-layer detection and response with adversarial-attack telemetry. Typical: model threat detection, red-team telemetry. | not checked | SOC 2 (claimed)supports Art. 15 robustness measures (claimed) |
| Palo Alto Prisma AIRS | network-integrated AI security | AI runtime security folded into an existing enterprise network security estate. Typical: enterprise rollout, egress control. | not checked | SOC 2 (claimed)enterprise security integration (claimed) |
| Cisco AI Defense | network-integrated AI security | Discovery of AI usage plus inline enforcement across the corporate network. Typical: shadow-AI discovery, inline enforcement. | not checked | enterprise security integration (claimed) |
and 4 more in the stack advisor →
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
Agent Observability & Model Risk Management
Tracing, evaluation, drift monitoring and model-validation records. This layer is where Art. 12 record-keeping becomes technically real (step-level traces, prompt/response records, retention control) and where model-risk practice in the SR 11-7 tradition — validation evidence, performance and drift monitoring, challenger comparison — is operated. Gateways and tracing tools produce the logs; the retention, integrity and access regime around them is still yours.
unverified · verified 2026-08-18 community-maintained
selection metrics: Trace completeness per agent step; log retention and immutability options; drift/quality metrics available out of the box; evaluation dataset support; export into your audit vault; self-host option.
supplies: OpenTelemetry / FCoT Tracing · Shadow-Mode Execution
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| LangSmith | agent tracing & evaluation | Trace capture and evaluation over LangChain/LangGraph runs with dataset-based scoring. Typical: step tracing, regression evaluation. | not checked | SOC 2 (claimed)supports Art. 12 record-keeping (claimed) |
| Langfuse | agent tracing & evaluation | Open-source tracing, prompt management and evaluation; self-hostable for retention control. Typical: self-hosted tracing, cost/latency analytics. | open source | GDPR-positionedsupports Art. 12 record-keeping (claimed) |
| Arize AI / Phoenix | ML & LLM observability | Production monitoring with drift and performance analysis; Phoenix is the open-source tracing side. Typical: drift monitoring, production analytics. | not checked | SOC 2 (claimed)drift-monitoring positioning (SR 11-7 style, claimed) |
| Helicone | LLM gateway & logging | Proxy-level logging of prompts, costs and latency across providers. Typical: gateway logging, cost control. | not checked | SOC 2 (claimed)supports Art. 12 record-keeping (claimed) |
and 11 more in the stack advisor →
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
Cryptographic Evidence & Audit Ledger
Tamper-evident recording of what a system did: content-addressed decision records, hash chains and external anchoring, so a log can be shown not to have been rewritten after the fact. This is the layer that turns Art. 12 logging and Art. 19 retention from a storage question into an evidentiary one. AI Verify is carried in RAIN as a STANDARD node (sg-ai-verify), not duplicated here as a vendor.
unverified · verified 2026-08-18 community-maintained
selection metrics: Append-only guarantees and who can rotate or delete (including the vendor); anchoring mechanism (qualified timestamp, transparency log, notarisation) and whether verification works without the vendor; retention and export in a readable format at end of contract; throughput and cost at your event volume.
supplies: WORM / Immutable Audit Vault
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| Fact0 | cryptographic evidence ledger | Positions itself as a tamper-evident ledger for AI decision records. Typical: decision records, audit trail. | not checked | supports Art. 12 record-keeping (claimed) |
| Traccia | audit trail & traceability | Positions itself around traceability of AI pipeline steps and artefacts. Typical: traceability, artifact lineage. | not checked | supports Art. 12 record-keeping (claimed) |
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
Runtime Guardrails & Enforcement
Policy enforcement in the request path: input/output validation, injection and exfiltration defence, structured-output constraints and action blocking. Distinct from observability layers because these products are in-line and can refuse. Selection questions: added latency at p95, whether enforcement is fail-open or fail-closed, whether policies are versioned artefacts, and whether the layer can be self-hosted inside your data boundary.
unverified · verified 2026-08-18 community-maintained
selection metrics: Where enforcement sits (inline proxy, sidecar, SDK) and the added latency at your token volumes; whether policy is versioned and testable as code; fail-open vs. fail-closed behaviour under guardrail outage; language and modality coverage; whether every block writes an evidence record you can cite later.
supplies: Deterministic Policy Engine (OPA / Cedar)
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| Guardrails AI | validation framework | Open-source validator framework for structured output and content policies in the request path. Typical: output validation, structured output. | open source | supports Art. 15 robustness measures (claimed) |
| NVIDIA NeMo Guardrails | dialogue policy rails | Programmable dialogue and topic rails placed around an LLM application. Typical: topic control, dialogue policy. | open source | supports Art. 50 interaction disclosure patterns (claimed) |
| Lakera AI | guardrail proxy | Inline prompt-injection and content detection at request time. Typical: injection defence, content filtering. | SaaS (vendor cloud) | SOC 2 (claimed)supports Art. 15 robustness measures (claimed) |
| Credal AI | enterprise access & policy layer | Permission-aware access layer with data-loss controls in front of enterprise assistants. Typical: access control, DLP. | SaaS (vendor cloud) | SOC 2 (claimed) |
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
Agent Orchestration & SDLC Toolkits
Developer middleware for multi-agent networks, tool-use chains, RAG abstraction, state and memory persistence, and model routing. Named products live in marketExamples; prose here describes the class. Regulatory posture: orchestration code is where autonomy tiering, propose-action objects and fallback routing get implemented — the framework choice constrains which controls are cheap and which are retrofits. Selection metrics: see meta.marketLandscape.selectionMetrics.orchestration.
unverified · verified 2026-08-18 community-maintained
selection metrics: broad model-API abstraction, state/memory management, error recovery, fallback routing hooks
supplies: Multi-Model Router & Fallback Abstraction · Cognitive Orchestrator
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| LangChain / LangGraph | agent framework | Graph-structured agent runtime; interrupt/pause nodes support implementing human approval at defined steps. Typical: multi-step agents, approval workflows. | not checked | supports implementing Art. 14 oversight (claimed)supports Art. 12 step logging (claimed) |
| LlamaIndex | RAG framework | Indexing and query abstractions over documents and structured sources. Typical: enterprise RAG, document agents. | open source | retrieval-governance positioning |
| Microsoft AutoGen | multi-agent framework | Conversational multi-agent patterns with pluggable tool executors. Typical: multi-agent research, code agents. | not checked | research/OSS, no vendor certification |
| CrewAI | multi-agent framework | Role-based agent teams with task delegation and process templates. Typical: process automation, role-based agents. | not checked | vendor-stated security posture |
and 6 more in the stack advisor →
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
Regulated Foundation-Model Platforms
Frontier commercial APIs and open-weight models under enterprise controls: zero-data-retention tiers, data isolation, fine-tuning governance, safety alignment documentation, EU-sovereign options. Named products live in marketExamples, where the deployment model is recorded in the hosting field rather than asserted in prose. What the class buys you: a model supply relationship with contractual data handling and documentation you can pass to a customer. GPAI-chapter duties and provider due diligence attach at this layer. Selection metrics: see meta.marketLandscape.selectionMetrics.models.
unverified · verified 2026-08-18 community-maintained
selection metrics: ZDR enterprise tiers, data isolation, EU-sovereign options, fine-tuning controls, safety alignment documentation
supplies: Multi-Model Router & Fallback Abstraction
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| OpenAI (Enterprise / API) | proprietary frontier | Enterprise tiers offer zero-data-retention and no-training commitments over the commercial API. Typical: general copilots, document reasoning. | not checked | SOC 2 (claimed)ISO 27001 (claimed)zero-data-retention tier (claimed)GDPR-positioned |
| Anthropic Claude (Enterprise) | proprietary frontier | Enterprise/ZDR tiers with published safety and model documentation practice. Typical: regulated assistants, long-context analysis. | not checked | SOC 2 (claimed)ISO 27001 (claimed)zero-data-retention tier (claimed)HIPAA-eligible (claimed) |
| Google Gemini Enterprise | proprietary frontier | Vertex-hosted frontier models with regional grounding and customer-managed keys. Typical: enterprise search, multimodal workflows. | not checked | SOC 2 (claimed)ISO 27001 (claimed)HIPAA-eligible (claimed)EU data-boundary positioning |
| Cohere | proprietary frontier | Private-cloud and on-prem deployment of retrieval-oriented models. Typical: private RAG, enterprise search. | self-hostable | SOC 2 (claimed) |
and 7 more in the stack advisor →
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
Grounding, Retrieval & Agent Memory
The grounding layer between raw sources and the model: document parsers, embedding models, vector databases and — new in the agentic era — persistent agent memory stores. Memory is the hard part: once a personal fact is embedded, GDPR Art. 17 erasure has to reach the vector and the memory record, not just the source row, and embeddings are partially reconstructable (see IronCore in the privacy layer). Retrieval quality is also a data-governance question under Art. 10: what got parsed, chunked and indexed is what the system 'knows'.
unverified · verified 2026-08-18 community-maintained
selection metrics: Parsing fidelity on your worst document class; retrieval precision/recall on a labelled set; tenant and ACL isolation model; per-vector encryption and erasure path; memory TTL and record semantics; self-host option.
supplies: Bitemporal Memory (GDPR×Art.12)
Filters to self-hostable, customer-VPC and open-source options when personal or confidential data cannot leave the EU.
| Example | Sub-category | What it does | Hosting | Claimed alignments |
|---|---|---|---|---|
| Docling | document parser | Open-source layout-aware parsing of PDFs and office formats into structured chunks. Typical: RAG ingestion, air-gapped pipelines. | self-hostable | EU sovereignty positioning |
| LlamaParse | document parser | Managed parsing service tuned for tables and complex documents feeding RAG. Typical: RAG ingestion, table extraction. | not checked | SOC 2 (claimed) |
| Amazon Textract | document parser | OCR and form/table extraction with per-page pricing inside AWS. Typical: document intake, claims processing. | not checked | SOC 2 (claimed)HIPAA-eligible (claimed)ISO 27001 (claimed) |
| Diffbot | web/knowledge extraction | Structured extraction and knowledge-graph construction from web sources. Typical: market monitoring, entity resolution. | not checked | vendor-stated security posture |
and 12 more in the stack advisor →
Community-maintained, disputable examples — not an endorsement and not a ranking. Alignments are as claimed by vendors or the source compilation, not verified by RAIN; a certification is shown as a certification only where a certificate or registry reference is recorded.
Disclosure: RAI·N·avigator operates in this category too, so we have a commercial interest in any comparison here. That is why this layer maps product classes to control objectives and lists named products as community-maintained examples — we publish no rankings, no quadrants and no coverage assertions about any vendor, including ourselves.
Procurement rule: Derived from three-lines-of-defense separation: the second-line GRC platform must be procured and deployed independently of any first-line runtime or model vendor — a governance tool that only sees its own vendor's models cannot govern a multi-model estate, and closed third-party SaaS AI can only be governed contractually (intake, attestation, AI-BOM disclosure), never by inline inspection.
Outsourced delivery BPO · SaaS · Service-as-a-Software caveats
Delivery Model — BPO · SaaS · Service-as-a-Software
Spectrum
BPO: input-priced (billable hours/FTEs), linear headcount scaling, human error & attrition as primary risk
SaaS: capability-priced (software access), client operates the workload, implementation/adoption failure as primary risk
Service-as-a-Software: outcome-priced (SLA on completed work), provider-managed AI executes 60–80% of cognitive tasks with specialist supervision, algorithmic bias & non-compliance as primary risk
Caveats in regulated markets
— Outcome SLAs move compliance risk onto the provider — but NOT the buyer's deployer duties: Art. 26 oversight, log retention and FRIA obligations stay with the enterprise even when execution is outsourced.
— Provider role analysis is the central legal question: a productized platform that fine-tunes, re-purposes or chains models can flip into the Art. 25 provider role with full high-risk obligations.
— Certified operations (ISO 42001) function as a procurement moat and shortcut third-party risk assessment — but organizational certificate ≠ product conformity (never conflate, see meta.assuranceEcosystem).
— The buyer's evidence chain must reach into the provider: contractually mandated AI-BOM disclosure, ZDR certificates, bias-audit reports and logging-ledger access are the artifacts that make an outsourced workflow auditable.
Threat Profile
LLM09 Misinformation
Hallucinated or wrong outputs create liability and decision risk.
mitigate with: Output Rails / Groundedness Check, Explainability API (SHAP/LIME/CoT)
Audit-Trail Manipulation (Insider)
Evidence tampering by parties with legitimate admin access: rebuilding the WORM vault, truncating hash chains before export, backdating records, selective deletion between audits. The sharper audit test: not whether logs are immutable in normal operation, but whether someone who administers the store can alter them unnoticed. Defeats every log-derived artifact at once (Art. 12, DORA, NIS2, PLD disclosure defence) if successful.
mitigate with: External Trust Anchor (Qualified Timestamp / Ledger), WORM / Immutable Audit Vault, Agentic Zero Trust, Cyber Event Recovery Plan & Playbooks
LLM06 Excessive Agency
Over-broad rights/functions of autonomous agents lead to uncontrolled actions.
mitigate with: MCP Gateway / Proxy, Agentic Zero Trust, Per-Action Autonomy Tiering, Trinity Defense (TCB + Command Gates + IFC), Deterministic Policy Engine (OPA / Cedar), Guardian Agents (Runtime Policy Enforcement), Non-Human Identity Credential Broker, Tool-Use Boundary Proxy