Home / Board Briefings
BOARD BRIEFING REPORTEuropean Banking & Financial MarketsManagement Board Question · For Discussion

Where Is AI Allowed to Act Without Human Approval?

TopicAgentic AI & Decision Rights
AudienceManagement Board · Supervisory Board · Executive Committee
Read Time7 minutes
Target StatusGREEN — when AI action boundaries are defined, evidenced and reversible

At a glance

The issue is no longer only model output.

The issue is the action boundary.

Board AskWhere is AI allowed to act?

Jump to answer calibrationDownload PDF2.0 MB

Situation

AI is moving from recommendation to action.

The issue is no longer only model output.

The issue is the action boundary.

An AI-supported workflow is no longer only producing text, summaries or recommendations. It can trigger a workflow step, update a record, escalate an alert, draft a client communication, prioritise a case, recommend a control response or prepare an action for execution.

Management Summary

Agentic AI changes the governance question. Traditional AI governance asks whether a model is accurate, explainable, monitored and controlled. Agentic AI adds another question: what is the AI allowed to do?

Capability is not permission. A model can be capable of acting autonomously. That does not decide where it is allowed to act. Permission must be determined by materiality, reversibility, customer impact, legal effect and risk appetite.

Human oversight must be designed. A strong answer is not "there is a human in the loop". A strong answer shows where AI proposes, where AI executes only after approval, where bounded autonomy is allowed, and where autonomous action is not permitted.

Management Report Panel

Situation in One Sentence
The issue is the action boundary.
Key Issue
Traditional AI governance asks whether a model is accurate, explainable, monitored and controlled. Agentic AI adds another question: what is the AI allowed to do?
Primary Risk
A model can be capable of acting autonomously. That does not decide where it is allowed to act. Permission must be determined by materiality, reversibility, customer impact, legal effect and risk appetite.
Board Ask
Where is AI allowed to act?
Target RAG
GREEN is an evidenced control boundary.

Answer Quality Calibration

GREEN Decision-Grade
AMBER Partially Evidenced
RED Not Decision-Grade
InventoryAI action inventory exists and is current.
InventoryAI use cases are inventoried, but action rights are not fully mapped.
Inventory“It only recommends.”
Autonomy LevelEach AI-supported workflow has a defined autonomy level.
Autonomy LevelHuman review exists, but is inconsistently evidenced.
Autonomy Level“A human is somewhere in the loop.”
ApprovalHuman approval points are mapped by role and authority.
ApprovalTechnical permissions exist, but are not linked to business decision rights.
Approval“The workflow is internal.”
PermissionsTool access, API access and workflow permissions are controlled.
PermissionsException logs exist, but are not reviewed at governance level.
Permissions“The user can override it.”
LoggingActions are logged with actor, authorisation, timestamp and outcome.
LoggingBusiness ownership exists, but not for autonomy boundaries.
Logging“The model has no final decision right.”
ExceptionsException, override and stop mechanisms are monitored.
ExceptionsStop mechanisms exist in design, but have not been tested.
Exceptions“The vendor controls the permissions.”
Board ReportingBoard reporting distinguishes assistive, advisory and action-triggering AI.
Board Reporting—
Board Reporting“We will define human approval during rollout.”
Ownership—
Ownership—
Ownership“This is only a productivity tool.”

GREEN is not a human-in-the-loop label. GREEN is an evidenced control boundary.

The Autonomy Ladder

AI autonomy is not a binary switch. It should be calibrated by materiality and reversibility.

01
SUGGEST
AI proposes. A human decides and acts. Default for material, irreversible or legally sensitive actions.
02
EXECUTE WITH APPROVAL
AI prepares or executes an action only after explicit human approval. Suitable where automation creates efficiency, but the decision remains human.
03
BOUNDED AUTONOMY
AI acts independently within a narrow, predefined action space. Suitable for reversible, high-volume, tightly monitored activities.
04
MONITORED AUTONOMY
AI acts autonomously while monitoring detects exceptions, anomalies and boundary breaches. Only suitable for low-risk, reversible and fully logged contexts with tested stop mechanisms.

The autonomy level follows the materiality and reversibility of the action. It does not follow the capability of the model.

Who Should Answer

CIO / CTOBusiness OwnerCRO / Operational RiskCompliance / Legal / DPOCISO / Technology RiskCOOInternal Audit

Accountability note:
Technology May Enable Autonomy.
The Business Owns the Action Impact.
The Board Needs to See the Boundary.

Evidence the Board Should Request

  1. 01AI Action Inventory
  2. 02Autonomy Level Classification
  3. 03Action Boundary Map
  4. 04Human Approval Matrix
  5. 05Tool / API Permission Register
  6. 06Decision Rights Map
  7. 07Exception and Override Log
  8. 08Client / Conduct Impact View
  9. 09Testing and Simulation Evidence
  10. 10Stop / Rollback Protocol

Warning Signals

  • “It only recommends.”
  • “A human is in the loop.”
  • “The workflow is internal.”
  • “The user can override it.”
  • “The model has no final decision right.”
  • “The vendor controls the permissions.”
  • “We will define human approval during rollout.”
  • “This is only a productivity tool.”

None of these statements is necessarily wrong. They are simply not sufficient evidence for board-level reliance.

Path to Green

  • Action Rights Visible
  • Autonomy Classified
  • Human Approval Meaningful
  • Permissions Controlled
  • Actions Reconstructable
  • Stop Mechanism Tested

GREEN is not a human-in-the-loop label.
GREEN is an evidenced control boundary.

Suggested Next Question

Which AI-supported workflow would create the greatest accountability issue if it acted tomorrow without explicit human approval — and can we evidence the boundary that prevents that?

Selected Source Base

  • EU AI Act, especially Articles 12, 14, 15, 19, 26 and Annex III.
  • GDPR, especially Article 22 on automated decision-making.
  • DORA, especially ICT governance, logging, detection and continuity requirements.
  • ESMA, Public Statement on AI and Investment Services, May 2024.
  • IOSCO, Supervisory Toolkit for AI Use in Capital Markets, 2026.
  • NIST AI Risk Management Framework and Generative AI Profile.

Related insights

← All briefings

Reza Monshizadeh
Reza Monshizadeh
Senior Technology Executive · Frankfurt

Two decades leading transformation, governance and technology strategy across ECB-supervised European financial services. Head of Governance & Strategy Europe at J.P. Morgan SE, with prior leadership at UBS Europe SE, Merz Pharma and Finanz Informatik.

View experience →LinkedIn →Get in touch →