Global AI Assurance · Regulatory assurance
Global AI Control Baseline
International prudential standards never tried to make every country's banking law identical. They set common principles, and each jurisdiction implemented them through its own regime. AI assurance needs the same shape.
One baseline of control objectives, then per-jurisdiction overlays naming the additional obligations, controls, evidence and supervisory expectations that apply locally.
Published as reference architecture. Control weights, grading rubrics and the full control set are set per engagement.
Principle
An international baseline does not make every jurisdiction's law identical. It states the common control objectives, and each jurisdiction implements through its own regime. The overlay is where the extra obligations, controls, evidence and supervisory expectations are named.
Fourteen baseline objectives
Each written as something that can observably fail, with the artefact an outsider may inspect. Weights and grading rubrics are set per engagement.
Every AI system in the estate is discovered, registered and owned.
Register with owner, purpose, classification and approval state; discovery coverage stated.
Each system carries a classification proportionate to its consequence.
Classification record with the test applied and the reviewer.
Capability is assessed against a defined domain set, not a benchmark.
Domain-level capability read with evidence sources and uncertainty named.
Agents hold an identity, an owner and an explicit authority envelope.
Identity record plus the envelope, including what the agent may not do.
Authority is not transitive across agents.
Delegation decisions and refusals, with the chain evaluated.
Actions are evaluated before they take effect where consequence warrants.
Runtime decision records: allow, conditions, human review, block, escalate.
A person can interrupt an action within a stated latency.
Intervention capability tested, with measured latency.
Independent evaluation covers the material capability domains.
Evaluation records with disposition and date, by an assessor named.
Applicable requirements are resolved from the full activity chain.
Jurisdiction intersection record per activity, with the disposition.
Sovereignty requirements are set by the workload and verified against the configuration.
Sovereignty classification with the configuration contracted and tested.
Incidents are detected, classified and reported to each regime that requires it.
Incident record with the fan-out of filings and clocks.
Concentration and substitutability are assessed across providers and infrastructure.
Dependency map with exit position stated.
Evidence is complete, current and held where it can be compelled.
Freshness and completeness measures, with residency.
The board receives a movement in the institution's resilience position.
Attested board record tying AI assurance to the resilience read.
Jurisdictional overlays
The overlay is where local law lives. The baseline does not move; the overlay adds.
EU
US
Canada
UK
Japan
Singapore
Gulf
Australia
A baseline that cannot be monitored across a population is a document, not a control.
See the supervisory view