Understand the terms. Ask a better question.

A protocol for communication and task exchange between agent systems.

In practice: A research agent exchanges a task with another agent; the receiving organisation still controls access.

Decision boundary: An advertised capability or valid message does not establish trust or permission for a business action.

Software that uses a model to choose steps and tools while pursuing a task.

In practice: An assistant proposes a refund; a separate authority check determines whether it may execute it.

Decision boundary: Tool access does not establish permission to act for your organisation.

A structured examination of an AI system or intended use against stated criteria.

In practice: Assess a customer-service assistant against the actual request types and escalation rules.

Decision boundary: A questionnaire or workshop is not automatically certification or legal approval.

The AI systems, models, providers, data and tools an organisation needs to understand and manage.

In practice: Record which model and cloud a service assistant actually uses, with the responsible owner.

Decision boundary: An inventory records what is known; it does not itself enforce controls. The review Workspace is a demonstration.

The responsibilities, rules and practices used to direct and oversee AI use.

In practice: Name the system owner, required evidence and review trigger when a model changes.

Decision boundary: A management-system standard does not grant an agent operational permission.

Preparing to identify applicable AI Act obligations and establish supporting responsibilities, controls and evidence.

In practice: Map the intended use and responsible owner before deciding which obligations apply.

Decision boundary: Readiness work is separate from a legal opinion or conformity assessment.

A person examines a result or proposed action under a defined responsibility.

In practice: A reviewer sees the customer evidence before approving a disputed refund.

Decision boundary: A nominal approval step does not prove an effective review. Include review effort in cost assumptions.

Running a trained model on an input to produce an output.

In practice: Include rejected answers, retrieval and human review when costing a support response.

Decision boundary: A low inference price is not the full cost per accepted business outcome.

A protocol through which AI applications can connect to tools, resources and prompts.

In practice: Allow an assistant to read an approved knowledge source without granting write access to the CRM.

Decision boundary: A protocol connection or token does not decide whether a business action is authorised.

Observing systems and relevant changes over time.

In practice: A changed supplier retention policy triggers a documented review.

Decision boundary: An observation does not itself change identity, verification, publication or assessment outcomes.

A description of what a source or system showed at a particular time.

In practice: A supplier page states a retention period; record the date and applicable plan.

Decision boundary: An observation is separate from verification, endorsement and publication.

AI made available under terms and with materials that permit the freedoms set out in an explicit open source definition.

In practice: Check the model licence and required materials before using an open source label in procurement.

Decision boundary: Downloadable weights alone do not establish that definition or unrestricted commercial use.

Model parameters are made available for download under specified terms.

In practice: A downloadable model may still restrict certain uses or require a commercial agreement.

Decision boundary: Inspect the actual licence, restrictions and available code/data information.

A buyer’s confidence that a specific offering can meet the intended need under the agreed conditions.

In practice: Compare two suppliers on the same workload, data conditions and accepted outcome.

Decision boundary: This buyer judgement is not an independently calculated certification or a new trust score.

Information used to support a particular claim, within its stated scope.

In practice: A regional hosting option is evidence about that option, not every plan or account.

Decision boundary: One documented fact does not establish every related claim or certification.

A published indicator about a specific recorded claim, review or status.

In practice: Follow a stated certification to the issuer’s record and check the covered product.

Decision boundary: A directory entry or logo alone is not a certification.

ETHORITY’s authority for publication of records under its defined rules.

In practice: Inspect a sourced market profile without assuming a certification or Registry admission.

Decision boundary: A record in this research review does not automatically have published Registry or verified status.

Retrieving relevant material and supplying it as context for a model-generated answer.

In practice: Answer a service question from the applicable contract edition and show the supporting passage.

Decision boundary: Retrieved text can be wrong, outdated or malicious; retrieval does not guarantee a correct answer.

The explicit permission an actor has for a particular action under current conditions.

In practice: Permission to prepare a payment is separate from permission to execute it.

Decision boundary: This is an architectural concept; the glossary does not demonstrate an operational enforcement service.

A model-specific unit used to represent text or other input/output.

In practice: Estimate input and output separately using the tokenizer and billing unit for the selected model.

Decision boundary: Do not assume one token equals one word or transfer a count unchanged between models.

A neural-network architecture that uses attention to relate parts of an input sequence.

In practice: Evaluate a document assistant on your actual document types rather than its architecture label.

Decision boundary: The architecture name does not establish business accuracy, privacy or suitability.

A history of trust-related evidence and events.

In practice: Keep the earlier policy source when recording a corrected supplier fact.

Decision boundary: A log entry does not independently verify the recorded claim.

How recorded evidence, assessments, verification and publication states change under their respective rules.

In practice: A changed model version prompts reassessment while the earlier decision remains traceable.

Decision boundary: A change observation does not itself authorise a new verification or publication state.

Finding items whose numerical representations are similar under a chosen measure.

In practice: Retrieve possible policy passages, then check access rights and the actual answer context.

Decision boundary: Numerical similarity does not establish factual truth or permission to read a document.

Checking a stated claim against specified evidence and criteria within a defined scope.

In practice: Read what a certificate actually covers before applying it to a particular product.

Decision boundary: A provider’s claim, a source link or directory inclusion alone is not independent verification.

A defined sequence or arrangement of steps used to complete work.

In practice: Classify a request, prepare a response, obtain review and only then use an authorised sending step.

Decision boundary: Adding an AI step does not make every workflow an autonomous agent.