What it is
Human-in-the-Loop represents a fundamental role transition in business processes. When an AI agent executes a workflow, it acts as the executor — performing tasks, making low-risk decisions, gathering information. But for consequential actions — approving a transaction, releasing data, making a binding commitment — the process pauses. A human shifts from passive observer to active approver. They review the complete context, make the judgment call, and apply their cryptographic signature. This signature, with full credentials and timestamp, becomes part of the Evidence Pack. It's not a UI pattern — it's an architectural guarantee enforced at the platform level.
Routine runs autonomously. Consequential actions wait for a named signature — appended to the evidence pack.
Why it matters
Regulatory frameworks and liability law require human accountability for consequential decisions. 'The AI did it' is not a defense. Human-in-the-Loop solves this by making approval a platform primitive, not an afterthought. The signing individual sees the complete decision context — not a summary, not a redacted version, but the full evidence the AI considered. Their signature proves they had the information needed to make an informed judgment. This transforms vague 'oversight' into cryptographically verifiable accountability.
Where it lives in AIOP
Integrated into the Attest primitive. Workflows define sign-off chains: which roles must approve which action types. When a consequential action triggers, the workflow pauses, emits an Attestation Request to the designated role, and blocks until the signature arrives. The approval decision, context, and signature are recorded in the Evidence Pack. The platform enforces this at the perimeter — unsigned actions that require signatures cannot execute.
Satisfy regulatory requirements for human accountability without slowing operations.
- Reduce liability exposure by documenting informed approvals.
- Enable AI automation for routine workflow steps while ensuring human judgment on the consequential ones.
- Defend against 'inadequate oversight' claims with cryptographic proof of approval.
Make informed decisions with complete context, not blind sign-offs.
Prove that approvers saw what they approved.
Defend against liability with documented approval chains.
Automate routine work while escalating exceptions to humans.