The most dangerous AI decision may not be made by an algorithm. It may be the risk decision your Board never made.
Imagine this.
Your organization has approved 50 AI initiatives.
The Sales team wants autonomous customer engagement.
Finance wants AI-powered transaction approvals.
HR wants AI-assisted recruitment.
Cybersecurity wants AI Agents that can automatically respond to threats.
Everyone has a business case.
Everyone promises productivity.
Everyone wants faster deployment.
But one question remains unanswered:
"How much AI risk is our organization actually willing to accept?"
The CEO assumes Security has decided.
Security assumes Risk Management has decided.
Risk Management assumes the Board has approved the strategy.
And the Board assumes someone has established the boundaries.
That is how organizations accumulate AI risk without consciously accepting it.
And this is not just a corporate problem. Government ministries, regulatory authorities, public service agencies, provincial governments and local governments face the same challenge.
Nepal is moving towards an AI-driven future. But are its institutions equally prepared to accept responsibility for the risks that come with it?
1. AI Adoption: Everyone Wants the Benefits
Artificial Intelligence is no longer an experimental technology reserved for research laboratories and technology companies.
Organizations are integrating AI into everyday operations.
Banks are exploring AI-powered fraud detection, customer analytics and credit assessment. Government agencies are exploring intelligent public service delivery. Hospitals are investigating AI-assisted diagnostics. Universities are introducing AI into teaching, research and administration.
The motivation is understandable.
AI promises:
- Faster decision-making.
- Reduced operational costs.
- Improved productivity.
- Better customer experiences.
- Automated repetitive tasks.
- Enhanced data analysis.
- More efficient public service delivery.
For Nepal, these opportunities could be particularly valuable.
AI could help government agencies improve administrative efficiency, support agricultural forecasting, strengthen disaster preparedness and make public information more accessible.
Similarly, Nepalese financial institutions could use AI to improve fraud detection, identify suspicious activities and support operational risk management.
However, there is an important distinction that organizations must understand.
Approving AI adoption does not automatically mean approving the risks associated with AI.
An organization may approve an AI initiative because its business case looks attractive. But that does not mean its leadership has evaluated the potential consequences of inaccurate outputs, data exposure, algorithmic bias, operational disruption or autonomous decision-making.
This is where the real governance challenge begins.
2. The Missing Decision: AI Risk Acceptance
AI Risk Appetite refers to the amount and type of AI-related risk an organization is willing to accept while pursuing its strategic and operational objectives.
It establishes the boundaries within which AI technologies can be developed, purchased, deployed and operated.
Consider two management decisions.
Decision A: AI Adoption
"We approve the implementation of an AI-powered system to automate customer services and improve operational efficiency."
Decision B: AI Risk Acceptance
"We understand that this system may generate incorrect responses, expose sensitive information or produce unintended outcomes. We have assessed these risks, established acceptable boundaries, assigned accountability and approved appropriate controls."
These are two fundamentally different decisions.
The first approves innovation.
The second establishes organizational responsibility for the risks associated with that innovation.
An organization that has approved AI adoption without defining its AI risk appetite may unknowingly permit individual departments to make decisions that exceed the organization's actual risk tolerance.
For example:
- Sales may introduce an AI agent that accesses customer information.
- Finance may automate transaction approvals.
- HR may use AI to shortlist candidates.
- IT may integrate third-party AI services into internal systems.
- Cybersecurity may authorize AI agents to execute incident response actions.
Each initiative may appear reasonable when evaluated independently.
But collectively, they can create an interconnected risk environment that nobody has properly assessed.
The organization may have adopted AI successfully from a business perspective while failing to govern its overall AI exposure.
3. Why AI Risk Appetite Must Start at the Top
One common mistake is treating AI risk as exclusively an IT or cybersecurity responsibility.
It is not.
AI risk extends across financial, operational, legal, ethical, regulatory, reputational and cybersecurity domains.
A cybersecurity team can identify vulnerabilities in an AI system, but it cannot independently decide how much financial or reputational risk the organization should accept.
Similarly, Risk Management can assess exposure, but it cannot automatically authorize a high-impact business decision outside its delegated authority.
These responsibilities must be clearly established.
An effective AI governance structure should include:
Board of Directors or Governing Body
- Approves the organization's overall AI risk appetite.
- Establishes strategic boundaries for AI adoption.
- Ensures appropriate governance and accountability.
Senior Management
- Ensures AI initiatives remain within approved risk boundaries.
- Allocates resources for risk management.
- Ensures appropriate oversight of AI systems.
Risk Management
- Identifies and evaluates AI-related risks.
- Maintains AI risk registers and key risk indicators.
- Monitors exposure against approved risk appetite.
Information Security and Cybersecurity
- Assesses AI-related attack surfaces.
- Implements access controls, monitoring and security safeguards.
- Evaluates risks associated with AI agents, third-party models and sensitive information.
Legal, Compliance and Data Governance
- Evaluates applicable regulatory obligations.
- Oversees privacy, data protection and contractual requirements.
- Supports transparency and accountability.
Internal Audit
- Independently evaluates AI governance, risk management and control effectiveness.
The principle is straightforward:
The people responsible for implementing AI should not automatically have unrestricted authority to accept the risks associated with it.
4. Nepal's AI Governance Challenge
Nepal has already taken an important step towards establishing a national direction for AI adoption.
The Government of Nepal introduced the National Artificial Intelligence Policy, 2082, recognizing the importance of AI in national development, public service delivery, research, innovation and economic growth.
The policy also addresses responsible AI development, ethical considerations, privacy, security, transparency and risk management.
It envisages institutional arrangements and coordination mechanisms to support AI governance across different sectors and levels of government.
However, national policy direction and organizational risk acceptance are two different things.
A national AI policy can establish strategic objectives and governance expectations.
But every organization still needs to translate those expectations into operational decisions.
Consider a government agency introducing an AI-powered system to process citizen applications.
The system promises faster service delivery and reduced administrative workloads.
But what happens when the AI system incorrectly rejects an application?
Who reviews the decision?
Can citizens challenge the outcome?
Who is accountable for the error?
What happens if sensitive citizen information is exposed?
Similarly, a regulatory authority using AI to identify suspicious financial activities must determine whether AI-generated findings are sufficient for enforcement decisions or whether additional investigation and human validation are required.
The same challenge applies to provincial and local governments introducing AI into public administration.
AI governance cannot stop at the national policy level. It must become part of institutional governance and accountability.
The financial sector: A practical example
Nepal Rastra Bank has taken a particularly relevant step through its December 2025 Artificial Intelligence Guidelines.
The guidelines apply to NRB-licensed institutions, including commercial banks, development banks, finance companies, microfinance institutions, Nepal Infrastructure Bank, payment system operators and payment service providers.
They address AI applications such as credit scoring, fraud detection, customer service, risk management and compliance monitoring.
Importantly, the guidelines establish that the Board of Directors must define the institution's AI-related risk tolerance within its overall risk management framework.
They also emphasize senior management accountability, AI governance structures, human oversight, auditability and risk management.
This is an important distinction.
AI governance is no longer simply a matter of selecting a model, deploying an application or purchasing an AI-powered solution.
It is an institutional governance responsibility.
For Nepalese financial institutions, the message is clear: AI adoption must remain aligned with the institution's approved risk appetite and applicable regulatory requirements.
Reference: Nepal Rastra Bank – Artificial Intelligence Guidelines, December 2025.
5. Three AI Risk Boundaries Every Organization Should Establish
Organizations need a practical way to distinguish acceptable AI initiatives from those requiring stronger controls or restrictions.
A useful starting point is a three-level AI risk appetite model.
GREEN – Acceptable AI Risk
These are generally lower-impact applications that can be introduced under established organizational controls.
Examples:
- Internal document summarization.
- Administrative workflow automation.
- AI-assisted research.
- Internal knowledge management.
- Routine report generation.
Even these applications require appropriate data classification, access control and output verification.
AMBER – Restricted AI Risk
These applications may create significant consequences if they operate incorrectly or without adequate oversight.
Examples:
- AI-assisted recruitment.
- Customer profiling.
- Fraud detection.
- Credit assessment.
- Predictive public service decisions.
- AI agents interacting with enterprise applications.
These initiatives should require formal risk assessment, documented authorization, testing, continuous monitoring and appropriate human oversight.
RED – Unacceptable AI Risk
These are activities that violate legal requirements or exceed the organization's approved risk boundaries.
Examples:
- Uploading confidential customer or citizen information into unauthorized public AI platforms.
- Allowing unrestricted AI agent access to critical production systems.
- Deploying discriminatory decision-making practices.
- Permitting unauthorized autonomous financial transactions.
- Operating high-impact AI systems without required governance and safeguards.
The organization should explicitly define which activities are prohibited and which exceptional activities, if legally permissible, require additional authorization.
The purpose is not to stop AI innovation.
It is to ensure that innovation happens within consciously established boundaries.
6. The AI Agent Problem: When Adoption Becomes Operational Risk
The discussion becomes even more important as organizations move from AI assistants to autonomous AI agents.
Traditional AI applications often provide recommendations, summaries or predictions.
AI agents can go further.
Depending on their permissions and integrations, they may:
- Access organizational databases.
- Retrieve confidential information.
- Execute scripts.
- Modify configurations.
- Interact with enterprise applications.
- Initiate workflows.
- Trigger security responses.
- Perform actions through connected APIs.
Now imagine a cybersecurity department deploying an AI agent to automatically respond to suspicious activities.
The agent identifies a potentially compromised account and disables it.
But the activity was legitimate.
A critical employee loses access to essential systems during an important business operation.
Who approved the agent's authority to disable accounts?
Was its decision-making threshold tested?
Could the action be reversed?
Was human approval required?
What happens if an attacker manipulates the information the agent uses to make decisions?
The organization may have approved the AI deployment because it wanted faster incident response.
But did it consciously accept the operational risk of allowing an AI agent to take independent action?
This is why AI agent permissions, autonomy, monitoring and human intervention must be included in the organization's AI risk appetite framework.
The more authority an AI system receives, the more carefully its risk boundaries must be defined.
7. From AI Adoption to AI Risk Acceptance: A Practical Framework
Before approving an AI initiative, organizations should establish a consistent decision-making process.
The following framework provides a practical starting point.
| Stage | Key question | Required outcome |
|---|---|---|
| 1. AI identification | What AI system are we introducing? | AI inventory |
| 2. Business justification | What organizational objective will it achieve? | Business case |
| 3. Risk classification | What could go wrong, and who could be affected? | AI risk assessment |
| 4. Appetite evaluation | Are these risks within approved boundaries? | Risk appetite assessment |
| 5. Authorization | Who has authority to approve residual risk? | Documented approval |
| 6. Deployment | Are required controls implemented? | Controlled implementation |
| 7. Monitoring | Is the system operating within acceptable limits? | Continuous assurance |
A particularly important step is distinguishing between inherent risk and residual risk.
Inherent risk is the exposure before additional controls are applied.
Residual risk is the exposure remaining after the organization implements its controls.
For example, an AI-powered transaction monitoring system may initially introduce significant risks involving false positives, incorrect classifications and data exposure.
After implementing encryption, access restrictions, model validation, monitoring and human investigation, the remaining exposure may become acceptable.
However, the decision to accept that residual risk must be made by an authorized person or governing body, according to the organization's delegated authority.
A risk assessment alone does not constitute risk acceptance.
8. What Nepalese Organizations Should Do Now
Whether an organization is a commercial bank, government agency, university, insurance company or private enterprise, it should consider the following priorities.
First: Establish an AI inventory.
Organizations cannot govern what they cannot see.
Identify AI applications, models, autonomous agents, third-party services and employee use of external AI platforms.
Second: Establish an AI risk appetite statement.
Define which AI risks the organization is willing to accept, which require additional controls and which are prohibited.
Third: Assign ownership and accountability.
Every AI system should have an accountable business owner, a designated risk owner and clearly defined approval authority.
Fourth: Integrate AI risk into existing governance structures.
AI risk should be integrated into Enterprise Risk Management (ERM), Information Security Management Systems (ISMS), privacy governance, operational risk and business continuity arrangements.
Fifth: Monitor AI risk continuously.
Approval should not be treated as permanent permission.
Changes to models, data sources, permissions, vendors, integrations and operating environments can introduce new risks.
Organizations should periodically reassess whether AI systems remain within their approved risk boundaries.
International frameworks such as NIST AI Risk Management Framework, ISO/IEC 42001 and ISO/IEC 23894 can help organizations establish and operationalize these practices.
9. The Question Nepal Must Ask
The discussion about AI in Nepal should not stop at how quickly organizations can adopt artificial intelligence.
We should also ask whether our institutions are prepared to govern it.
Can government agencies demonstrate accountability for AI-assisted public decisions?
Can regulatory authorities establish appropriate oversight within their respective sectors?
Can banks demonstrate that their AI systems operate within formally approved risk boundaries?
Can organizations identify who authorized the risks associated with their autonomous AI agents?
Can citizens and customers challenge decisions influenced by AI?
And perhaps most importantly:
If an AI system makes a harmful decision tomorrow, will the responsible institution be able to identify who authorized the risk, who monitored it, and who is accountable?
AI adoption requires investment, infrastructure, skills and ambition.
But responsible AI adoption requires something equally important: conscious risk acceptance.
The question is no longer whether Nepalese organizations should adopt AI.
The question is whether they are prepared to accept responsibility for the risks that AI introduces.
AI adoption without risk acceptance is innovation without accountability.
And in an increasingly AI-driven world, the most important decision may not be which AI system an organization chooses, but which risks its leadership is willing to accept.
AI adoption is a business and public-service decision. AI risk acceptance is a governance decision. And accountability must remain human.
References and further reading
- National Artificial Intelligence Policy, 2082 – Government of Nepal
- Artificial Intelligence Guidelines, December 2025 – Nepal Rastra Bank
- NIST AI Risk Management Framework
- ISO/IEC 42001 – Artificial Intelligence Management Systems
- ISO/IEC 23894 – AI Risk Management
- https://www.linkedin.com/pulse/ai-risk-appetite-dr-deep-pandey-ry0yf/
