Skip to main content
Framework: EU AI Act
Article: 14.4(b) Official Requirement
Human oversight measures shall enable the individuals to whom human oversight is assigned to override, or not to use, the output produced by the high-risk AI system. How Katyar Addresses This Requirement Katyar delivers a traceable, auditable override mechanism through its Human-in-the-Loop (HITL) approval workflow. The control is considered satisfied when real override behavior has been demonstrated in production or testing data. Evaluation Criteria
Katyar considers the control satisfied when:
  • At least one human-in-the-loop approval request has been explicitly denied (overridden) in the last 30 days.
Evidence Collected (Quantitative)
  • Number of HITL events with status = ‘denied’
  • Associated policy ID(s) that triggered the approval request
  • Reason/comment provided by the human approver (mandatory field)
  • Approver identity and exact timestamp of the denial
  • Original AI-suggested action vs. final overridden outcome
Katyar Features That Enable Override Capability
  • Clear Deny Action — Prominent “Deny” button in Slack, Teams, dashboard, or custom approval UI
  • Mandatory Justification — Approvers must provide a reason/comment for every denial (stored permanently in audit trail)
  • Immediate Interruption — Denial instantly stops the agent workflow and prevents tool execution
  • Custom Error Feedback — Agent receives a structured error message explaining the denial (configurable)
  • Audit-Ready Logging — Every override is cryptographically signed with:
    • Original prompt/tool call
    • AI output that was overridden
    • Human decision (deny + comment)
    • Approver identity
    • Timestamp (millisecond precision)
  • Visibility in Dashboard — Denials appear in real-time event stream, Approvals tab, and Compliance activity feed
Remediation Steps to Strengthen This Control
  1. Ensure at least one approval policy is active that can realistically be denied (e.g. high-value refund, destructive database query, bulk sensitive action).
  2. Run test scenarios or live agent usage that triggers the approval flow.
  3. Have a human reviewer explicitly deny at least one request (do not auto-approve everything).
  4. Verify the denial appears in:
    • Approvals tab (with comment)
    • Audit logs (search for status=‘denied’)
    • Compliance dashboard → EU-14.2 card
  5. Optionally simulate 3–5 denials to build stronger evidence for auditors.
Auditor Expectations
Regulators will look for:
  • Proof that overrides have actually occurred (not just theoretical capability)
  • Clear documentation of why each override was made (justification comments)
  • Evidence that the override was effective (AI action was stopped, no tool call executed)
  • Traceability linking the denied AI output to the human decision
  • Reasonable frequency of overrides in high-risk scenarios (demonstrates the mechanism is used in practice)
Katyar’s override system is designed to be friction-low for humans while being maximally auditable — combining real-time interruption, mandatory justification, and permanent signed logs to meet both operational needs and strict regulatory scrutiny. Official Reference
Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)
→ Article 14 – Human oversight