← Back to catalogue
Published

AI system, agent and execution

EM-AI-01 · em-ai-01

Support accountable operation, permission review and investigation of system outputs by linking each execution to its configuration, dependencies and responsible parties.

Enterprise profiles Enterprise subject EnterpriseAI

Bundle → Layer → Finding → Questions Filled

2 bundles · 3 layers · 4 findings · 12 questions

System identity Describe the deployed system and its responsible parties.

System, model and product

Separate the deployed system from its components and service endpoints.

System boundary

Establish the system boundary and allocation of responsibility.

  1. Which deployed system and context of use does this system record describe?
  2. Which model versions are associated with this system record?
  3. How does this system record distinguish the accountable operator from other parties with assigned responsibilities?

Endpoint

Identify the serving endpoint and its change policy.

  1. Which endpoint and operating party does this endpoint record identify?
  2. What recording requirement applies when the model served by this endpoint changes?
Configuration and runs Connect delegated permissions and execution evidence to versioned configurations.

Agent configuration

Describe the tools and delegated authority assigned to a configuration.

Tool bindings and permissions

Record tool access within the operator's authority.

  1. Which tool bindings and permissions does this configuration authorize?
  2. What evidence shows that this configuration's delegated rights remain within the operator's authority?

Execution

Record execution evidence and assess whether controlled replay is feasible.

Run record

Link a run to its dependencies, actions and replay limitations.

  1. Which configuration and model versions did this run use?
  2. What evidence verifies the actions recorded for this run?
  3. What availability evidence does this run record retain for its inputs, artifacts, execution state and tool dependencies?
  4. Which controls would prevent replay of this run from repeating external effects?
  5. What replay feasibility assessment does this run record contain, including nondeterminism and expected result fidelity?

Classifiers Filled

Family
Enterprise profiles
Category
Enterprise subject
Entry kind
subject
Domain
EnterpriseAI
Industry
Cross-industry
Tags
EM-AI-01W2subjectAISystemAgentConfigurationToolBindingAgentRunInferenceEndpoint

What it is Filled

An AI system record describes its context of use, agent configuration, tools, executions and endpoints, distinguishing the accountable operator from providers, deployers and other parties with assigned responsibilities.

Why it exists Filled

Support accountable operation, permission review and investigation of system outputs by linking each execution to its configuration, dependencies and responsible parties.

Distinguishing features Filled

  • The system is the deployed whole in a context of use; the model is one component and the product is what is sold.
  • A configuration and a run are separate records with different identifiers.
  • An agent acts with rights delegated by a defined operator and never gains its own legal personality.
  • Distinct from the AI model registry entry, which describes an artifact rather than its use.

What robots and AI may and may not do Filled

Must not

  • Grant an agent authority because a field calls it an agent; rights come only from an operator.
  • Delegate to another agent or tool any right the operator does not hold.
  • Present the AI system as a person or a legal subject.
  • Overwrite a configuration in place so that past runs can no longer be traced to it.
  • Hide from affected people that a decision about them was made by an AI system.

Only with a human decision

  • Granting an agent new tools or wider permissions.
  • Deploying a changed model or configuration to production use.
  • Accepting a high-risk use of the system under applicable AI regulation.

May

  • Record runs with the configuration, model versions, tools and permissions they used.
  • Report which system, endpoint and model version produced a given output.
  • Flag a run whose actions exceeded the permissions in its configuration.

Moral aspects Filled

  • Accountability must stay with an identifiable operator, not dissolve into the system.
  • People affected by automated decisions are entitled to know and to contest them.
  • Unbounded delegation lets errors and misuse spread faster than people can stop them.

Who is affected

  • People about whom the system decides
  • Operators and their staff
  • Owners of the data and tools the agent touches

Owners Filled

Steward

The head of AI or research who answers for the system on behalf of the operator.

Master systems

  • Model registry
  • Experiment tracker
  • Evaluation store

Links to other meta-models Filled

neighbor

  • EM-AI-02
  • EM-AI-03
  • EM-FIN-05

references

  • WM-AI-001 - conceptual-candidate
  • WM-AI-002 - conceptual-candidate
  • WM-AI-005 - conceptual-candidate
  • WM-AI-004 - conceptual-candidate

What else AI and robots need to interact with it Filled

Identity and identifiers required Filled

  • A system is identified by the operator's system identifier together with its context of use.
  • Each configuration version and each run carries its own identifier.

Direct properties not applicable Not applicable

Not applicable

Enterprise record contour: physical properties belong to referenced world models.

Recognition required Filled

  • A system record names an operator, a context of use and associated model versions.
  • The deployed system is distinguished from its model artifacts and commercial product.

Capabilities and actions required Filled

  • Configurations can be versioned, compared and rolled back.
  • Replay feasibility depends on available inputs, artifacts, execution state, tool dependencies and controls against repeating external effects.
  • Replay assessments state the achievable result fidelity and limitations from nondeterminism.

Hazards and failure modes required Filled

  • Chained delegation can escalate privileges across agents and tools.
  • Silent model changes behind an endpoint can invalidate earlier evaluations.
  • Actions can become untraceable when runs are not recorded separately from configurations.
  • Replay can repeat external effects unless effective controls prevent them.

Standards and interfaces optional Filled

  • PROV-O from the World Wide Web Consortium supports provenance of runs and their outputs.
  • ISO/IEC 42001 from the International Organization for Standardization and the International Electrotechnical Commission addresses AI management systems.
  • The AI Risk Management Framework from the National Institute of Standards and Technology supports context and risk records.

Context of use required Filled

  • The record supports organizations deploying AI agents with tools in operational processes.
  • Regulatory obligations depend on jurisdiction, applicable scope, use classification and the role of each responsible actor.

Sources Filled

  1. NIST AI RMF: context, evaluation and risk management
  2. PROV/SPDX: provenance of data, artifacts and actions
  3. Model registry/evaluation tooling practice; map system, artifact, run and endpoint
  4. NIST AI Risk Management Framework - Checked live (HTTP 200) on 2026-10-07.
  5. W3C PROV-O: The PROV Ontology - Checked live (HTTP 200) on 2026-10-07.
  6. SPDX Specification - Checked live (HTTP 200) on 2026-10-07.

Open questions

  • How does a system differ from a model and a product?
  • How are agent actions and delegation restricted?
  • How to reproduce an execution when the model and tools change?
  • Установить границу и решение reuse/extend/new по действующим спецификациям.
  • Подтвердить semantic crosswalk, права и source mastership.
  • Выбрать immutable refs; провести проверки fixtures до заявления о публикационной готовности.

Machine files

Provenance

enterprise research programme · published-partial

Built from: enterprise/models/em-ai-01/brief.json, ver-cy/world-models/card-supplements/em-ai-01.json

Still in Russian: suggested owner, blocking decisions, vercy candidates. Translate in research/enterprise/i18n/units.en.json.