| name | ai |
| description | Use when invoking any AI inference in erpax — invoice OCR, bank-transaction matching, sanctions screening, anomaly detection, tax classification, HS-code suggestion, document classification, vector embedding, semantic search, or audit summarisation. Every AI call goes through the single `callWorkersAi` gate (entitlement, audit row, tenant scope, risk class, metering); the AI services barrel. |
| atomPath | ai |
| coordinate | ai · 1/base · 2eb77a91 |
| contentUuid | 248101b2-2ba9-5cd8-8b8f-64a87ed83487 |
| diamondUuid | 93dbdd87-4bb2-88e6-a274-2dc488be70b5 |
| uuid | 2eb77a91-e2e6-8782-8acc-4abbb8b6a97b |
| horo | 1 |
| typography | {"partition":"ai","bondDegree":25} |
| standards | ["COSO-ERM-2017","EN-16931","EU AI Act 2024 transparency-and-risk-classification","EU-AI-Act","EU-CSDDD-2024/1760","EU-Intrastat-Reg-2019/2152","GDPR Art.22(3) right-to-human-intervention","ISO-20022","ISO-27001","ISO-27037","ISO-8601-1","ISO/IEC 42001:2023 ai-management-system","ISO/IEC 42001:2023 ai-management-system`","ISO/IEC-23894","ISO/IEC-25010","ISO/IEC-27001:2022","ISO/IEC-42001","NIST-AI-RMF","NIST-FIPS-180-4","OWASP-ASVS","RFC-4122","RFC-8785","RFC-9562","UN-CEFACT","WCAG-2.1","WCO-HS","— the instrument reads SKILL.md) -->"] |
| bindings | ["ai/AI","vectorize/VECTORIZE_DOCS"] |
| signatures | {"computationUuid":"60aa5991-3535-8654-8705-b2e25a0d0c15","stages":[{"stage":"path","stageUuid":"8e6da7ff-0947-8c12-b941-623c69161362"},{"stage":"trinity","stageUuid":"bd0fca7d-3bac-88ca-9531-f0948731efc9"},{"stage":"boundary","stageUuid":"4456c1a7-d28a-88cc-9348-470a0c7845fb"},{"stage":"links","stageUuid":"f3242731-2be8-8b63-9426-5f5881a9813e"},{"stage":"horo","stageUuid":"6b3a7879-bb93-8d5d-b3f8-14967a81e2e3"},{"stage":"seal","stageUuid":"fc61a7fe-c9ef-81cd-8704-4946877a6865"},{"stage":"uuid","stageUuid":"49e0ad26-e698-860e-9c23-d437d7b9fb70"}]} |
| version | 2 |
ai — the AI services barrel
Every AI inference in erpax flows through one canonical entry point: callWorkersAi in cloudflare-ai.ts. That gate enforces five checks on every call: entitlement, audit row, tenant scope, risk class, and metering. The nine thin per-feature wrappers (invoice OCR, bank matching, sanctions screening, anomaly detection, tax classification, HS-code suggestion, document classification, vector embedding/upsert, semantic search, audit summarisation) all call through it — never bypass it.
GDPR Art. 22(3) requires human intervention for automated decisions with legal/significant effect; sanctions screening hard-codes aiRiskClass: 'high' and the wrapper refuses any auto-accept path.
EU AI Act 2024 transparency and risk classification governs the high-risk AI uses (sanctions, anomaly, fraud); every model invocation carries a risk class that gates the decision path.
ISO/IEC 42001:2023 is the AI management system standard the callWorkersAi audit row fulfils: every inference is logged with tenant, model, risk class, and timing.
Law — [[law]]: every AI inference flows through the one callWorkersAi gate (entitlement · audit row · tenant scope · risk class · metering) — no wrapper bypasses it, so each call is uuid-accounted and high-risk decisions are refused an auto-accept path.
Industry failure modes (hallucination · audit fragmentation · prompt injection · multi-agent collision · cost runaway) map to erpax diamond remedies in [[ai/industry]] — pure fns + tests, not hand-listed mitigations.
Standards
-
@standard ISO/IEC 42001:2023 ai-management-system
-
ISO/IEC 42001:2023 — ai-management-system. Every callWorkersAi invocation produces an audit row.
-
GDPR Art. 22(3) — right-to-human-intervention. High-risk AI decisions require a human gate; auto-accept is refused.
-
EU AI Act 2024 — transparency-and-risk-classification. Each model call carries a risk class that drives the decision path.
