A human review item is a structured action included in an Evidence Pack that indicates a person must make a judgment call. It is distinct from a risk flag, which is a data-level discrepancy that an agent might route or log automatically. A human review item explicitly signals that the decision cannot be delegated to the agent or resolved by querying another source.
What triggers a human review item
Human review items arise when:
- A risk flag is present and the correct response depends on business context the agent does not have (for example, a known prior security interest that has been verbally discharged but not yet removed from the PPSR)
- Source data is inconclusive or temporarily unavailable, and the workflow cannot continue without a verification result
- The transaction type has a regulatory requirement for human sign-off regardless of source check results
- Conflicting source data cannot be resolved by querying additional sources
What a human review item contains
- Item type: the category of action required (for example, confirm identity, verify discharge, review discrepancy)
- Description: what the reviewer needs to check or decide
- Related risk flags: which risk flags, if any, prompted the item
- Source references: which source checks are relevant to the review
- Priority: whether the item blocks the workflow or is advisory
Why agents need this
Agents need a clear signal that a workflow should pause for human input, rather than attempting to continue or making an autonomous decision on a matter that requires judgment. A human review item provides that signal in a structured form that an agent can use to: pause a workflow, generate a notification for a human reviewer, include relevant context in an escalation, and record that human review was required before proceeding.
Hoist does not tell agents to block or approve transactions. It tells agents what source data says and whether a human needs to weigh in. That boundary is intentional: source checks are evidence, not decisions.
How Hoist uses this
Human review items appear in the humanReviewItems field of an Evidence Pack. At S1-gate, items are included in fixture-backed Evidence Packs so agent workflows can be tested against realistic escalation patterns. The structure is consistent across Evidence Pack types so an agent handling a PPSR check and an agent handling an ABN check can use the same escalation logic.
Hoist provides source-backed evidence. It does not guarantee that a transaction is safe, legal, compliant, or financeable. Human review items are part of how that boundary is made explicit in the Evidence Pack.
What Hoist does not infer
A human review item is not a decline recommendation. It does not mean the transaction cannot proceed; it means a person must decide. Hoist does not know the operator's risk appetite, credit policy, or regulatory obligations. It does not know whether a risk flag is material to a specific transaction. The human reviewer uses the Evidence Pack, including the source checks and risk flags, to make that call.
Related terms
- Evidence Pack: the bundle that contains human review items alongside source checks and risk flags.
- Risk flag: a data-level discrepancy that may trigger a human review item.
- Source check: the individual verification step that provides the evidence a reviewer uses to make a judgment.
- AI agent: the agent runtime that receives human review items and routes them to reviewers.
- PPSR: source whose encumbrance data may require human review if a security interest is found.
- NPII: AFSA insolvency index whose results may require human review before a counterparty transaction proceeds.