Which decisions may AI make, recommend, or never touch?
Every organisation adopting AI agents is drawing this map, whether it means to or not. We help leaders draw it deliberately, with the controls each category needs.
Three categories of decision authority
Every decision handed to AI passes through the same five steps and lands in one of three zones.
- DecisionName the decision and its owner.
- Risk classificationRate consequence, reversibility and regulatory effect.
- Delegation authorityPlace it in category A, B or C.
- MonitoringLog, sample and watch for drift.
- AccountabilityA named person answers for the outcome.
Category A · Agent-executed
Within explicitly authorised boundaries
AI agents execute predefined activities inside hard limits on value, volume and data. People monitor and can stop them.
Enterprise examples- Invoice matching under a value cap
- Routine service requests
- Data quality checks and reconciliations
- Scheduling and first drafts
Category B · Human approval required
Agent analyses and recommends; a person authorises
AI prepares the analysis or the action. A named person approves, amends or rejects it, with reasons recorded.
Enterprise examples- Credit memos and claims above a threshold
- Supplier changes and purchase orders
- Pricing proposals
- Customer remediation offers
Category C · Human-reserved
Decisions remain under direct human authority
AI may inform, never act. Reserved for decisions with high consequence, low reversibility, legal effect on individuals or a need for legitimacy.
Enterprise examples- Lending decisions about individuals
- Clinical judgement
- Hiring, dismissal and discipline
- Safety-critical and emergency operations
Category placement supports governance decisions. It does not by itself determine legal or regulatory compliance.
What each category needs
The more autonomy an agent has, the stronger the controls around it, not the weaker.
| Control | A · Agent-executed | B · Human approval required | C · Human-reserved |
|---|---|---|---|
| Decision ownership | A named process owner; the agent has its own identity and an accountable owner | A named approver for each decision type, recorded with the decision | A named decision-maker; AI outputs are inputs only |
| Approval thresholds | Hard limits on value, volume, counterparties and data the agent may act on | Thresholds above which a person must approve, with dual approval for the highest tier | Not applicable: every decision needs a person |
| Human override | Anyone accountable can pause or switch off the agent; a tested kill switch | The approver can reject or amend any recommendation, with reasons recorded | Full human discretion; AI advice can be ignored without justification |
| Monitoring and audit trails | Every action logged with inputs, tools used and outcome; sampled review | Recommendation, approval and rationale logged; acceptance rates monitored for rubber-stamping | AI inputs retained where used, so the basis of the decision can be reconstructed |
| Escalation rules | Automatic escalation on exceptions, low confidence, limit breaches or unusual patterns | Escalation to a senior approver when the recommendation is outside policy | Escalation follows existing delegated authority |
| Risk classification | Low materiality, reviewed at least annually | Medium to high materiality, reviewed at least half-yearly | High materiality or legal effect; AI use disclosed where required |
| Model and vendor oversight | Testing before deployment and after changes; vendor model changes notified and re-tested | Validation of the model and its data; performance tracked by segment | Vendor tools assessed for data handling, security and suitability before use |
How decisions move between categories
- Decisions move outward, towards more autonomy, only when evidence shows the agent performs within limits over a defined period.
- Decisions move inward when incidents, drift or regulatory change require it. Moving a decision inward should never need a project.
- The map is reviewed by its owner on a fixed cycle and reported to the board with the Agent Value Ledger.
Where might a decision sit?
Think of one decision you are considering handing to AI, and answer five questions. The result is a starting point for discussion, not a determination.
Answer all five questions to see a suggested category.
How the map aligns with published guidance
The map reflects themes common to Singapore's published guidance, including risk-based human involvement, accountability and life-cycle controls.
- Model AI Governance Framework for Agentic AI ↗22 January 2026IMDA (developed); announced by the Ministry of Digital Development and Information
Guidance on deploying AI agents responsibly: assess and bound risks upfront, keep humans meaningfully accountable, apply technical controls across the agent lifecycle, and enable end-user responsibility.
Announcement ↗ - Guidelines on Artificial Intelligence Risk Management (financial institutions) ↗Consultation 13 November 2025; response to feedback published 7 October 2026Monetary Authority of Singapore
Expectations for board and senior management oversight, AI inventories and risk materiality assessment, life-cycle controls and third-party AI. Published summaries indicate phased application from October 2027.
- IMDA and the AI Verify Foundation
Nine dimensions for a trusted generative AI ecosystem, including accountability, data, trusted development and deployment, incident reporting, and testing and assurance.
- Model AI Governance Framework, second edition ↗21 January 2020Personal Data Protection Commission (PDPC) and IMDA
The foundation of Singapore's approach: internal governance, determining the level of human involvement in AI-augmented decisions, operations management, and stakeholder communication.
- AI Verify testing framework and toolkit ↗International pilot May 2022IMDA; now under the AI Verify Foundation
A testing framework and software toolkit for demonstrating responsible AI through standardised tests and process checks.
References are provided to show how the AI Delegation Map aligns with published guidance. Kiros Intelligence Group is not certified, accredited or endorsed by any of these bodies, and using our frameworks does not by itself demonstrate compliance. Check the source documents for current requirements. Dates checked October 2026.
Draw your delegation map
Our governance work starts with an AI inventory and delegation maps for your material agents.