> ## Documentation Index
> Fetch the complete documentation index at: https://docs.katyar.com/llms.txt
> Use this file to discover all available pages before exploring further.

# EU-14.2 – Ability to Override AI Decisions

> EU AI Act Article 14.4(b) requirement to allow humans to override high-risk AI system outputs

**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](https://artificialintelligenceact.eu/article/14/)