Inland Revenue Office, New Road (Ministry of Finance, Government of Nepal)

The first Global Dialogue succeeds when Nepal's tax officers and Geneva policymakers use identical AI safety checklists for agricultural models serving 450M rice farmers. Concrete, time-bound deliverables: LDC-adapted AI safety tiers (like EU AI Act) for agriculture/tax systems - adopted by July 2027 $500M capacity fund - train 10,000 Global South officials within 24 months "Sovereign AI stack" standards - national models running on local infrastructure by 2028 Global explainability protocol - mandatory audit trails for tax/welfare AI decisions Open government AI models (non-sensitive) - 100+ countries commit to replicate solutions like my KAIST methane mitigation research As Inland Revenue Officer deploying AI daily, I know LDCs need toolkits, not declarations. Kathmandu cannot wait 10 years for Silicon Valley standards. Success = implementable frameworks by July 2027 New York meeting. My IRO role manages 8M taxpayers with 60% digital compliance gaps. AI could close this tomorrow with proper governance standards. Nepal coordinates GCF/World Bank climate finance daily - we implement global frameworks when they're practical. The Dialogue fails if developing nations leave with photo ops but no deployment guides. Success = Kathmandu tax officers using Geneva-approved AI safety checklists for rice methane models by December 2027. That's how 450M South Asian farmers benefit from global cooperation.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• AI capacity-building
• Transparency, accountability, and human oversight

Safe AI: My IRO tax systems handle 8M taxpayers. One algorithmic error = millions lost. KAIST research proved 27% methane abatement requires trustworthy models first. Capacity-building: Nepal has policy talent, zero AI governance experts. My IMF SARTTAC training showed standardized frameworks work. Need 10,000 officials trained in 24 months. Transparency: Taxpayers demand "why AI flagged my return." Global explainability standards = public trust. My daily reality: explain algorithms to non-technical citizens. Social implications: My methane model transforms 450M rice farmers but fails without Nepali/Maithili interfaces, smallholder economics ($2/day incomes), cultural adaptation. GCF/World Bank coordination proves this. These aren't academic - they're my daily gaps closing 60% digital compliance shortfall. Other priorities matter after we can actually execute AI governance.

Four critical gaps threaten LDC AI sovereignty: 1. Sovereign AI Infrastructure: Nepal cannot outsource tax collection/disaster response to US/China cloud providers. Need standards for national AI stacks using open models on local servers - revenue sovereignty demands it. 2. Public Sector AI Governance: Governments deploy highest-risk AI (taxation, welfare, justice). My IRO needs election-cycle audit trails, legacy system integration for 30M citizens, explainability across government handovers. 3. Climate Adaptation AI: 24% of Nepal's emissions from agriculture. My KAIST methane research (27% abatement potential) needs governance frameworks for national-scale deployment across rice farming, water management, disaster prediction. 4. Linguistic Data Justice: 7,000+ global languages ignored. Nepali/Maithili datasets cannot train methane models without standards protecting cultural sovereignty while enabling regional fine-tuning. These create "AI governance colonialism" - LDCs become testing grounds for foreign models without capacity to govern them. My tax officer reality: 8M taxpayers, 60% compliance gaps. We cannot risk revenue sovereignty on unaccountable algorithms any more than monetary policy. Current themes assume high-income preconditions (cloud infra, technical experts, English datasets). LDCs face opposite reality. Kathmandu tax officers need sovereign capability first, then safety standards. Without these cross-cutting issues, Global Dialogue produces Geneva declarations unusable in New Road tax office. Success requires LDC-specific governance enabling actual deployment, not theoretical frameworks requiring infrastructure Nepal builds over decades. My IRO + KAIST perspective reveals daily execution gaps UN frameworks miss. These aren't nice-to-haves - they're deployment prerequisites for 1.4B South Asians relying on AI governance reaching beyond current thematic silos.

12. How governance gaps affect Nepal's tax sector (294 words) CHALLENGES - Nepal's 60% digital tax compliance gap widens daily: Safe AI Gap: No national standards for tax AI deployment. My IRO systems risk algorithmic bias across 8M taxpayers - one error cascades millions in revenue loss. Nepal's National AI Policy 2025 exists but lacks LDC-specific implementation. Capacity Void: Zero AI governance experts among 45,000 civil servants. My IMF SARTTAC training is exception, not norm. Kathmandu tax officers cannot audit complex models, blocking deployment of my KAIST methane solution regionally. Transparency Crisis: Taxpayers demand "why AI flagged my return" but get no explainability. Legacy systems + black-box algorithms erode trust. 70% small businesses avoid digital filing due to opacity fears. Cultural Mismatch: Western AI models fail Nepali/Maithili-speaking farmers ($2/day incomes). My 27% methane abatement research cannot scale without localized interfaces. OPPORTUNITIES - $2B revenue + 20% emission cuts possible: Safe standards = deploy my KAIST model nationally, serving 5M rice farmers Capacity training = upskill 10,000 tax officers, close 60% compliance gap Transparency protocols = 30% compliance boost via taxpayer trust Social adaptation = AI interfaces in 10+ Nepali dialects = agricultural revolution South Asia loses $50B/year in tax leakage + 1.5GtCO2e agricultural emissions due to these gaps. Nepal coordinates GCF/World Bank daily - we implement when frameworks work. Current governance vacuum = AI deployment paralysis. My IRO cannot risk revenue sovereignty on untested models. Global Dialogue standards would unlock $1.2B revenue-neutral green tax shift across rice economies. Nepal's reality: policy-rich, implementation-poor. These gaps aren't theoretical - they're my daily barrier preventing AI from serving 30M citizens effectively.

The AI Dialogue must become the "AI governance OSHA" for developing nations - delivering standardized, deployable frameworks that Kathmandu tax officers can implement tomorrow. Five critical roles: LDC Testing Ground: Pilot safety standards in real-world deployments (Nepal IRO, Indian agriculture). Refine through South-South cooperation before global adoption. Capacity Multiplier: Coordinate $500M training fund across ITU/IMF/World Bank. Standardize 10-week "AI Governance Officer" certification - scale my IMF SARTTAC model globally. Sovereign AI Blueprint: Develop "national AI stack" reference architecture (open models + local infra). Nepal cannot outsource tax sovereignty to foreign clouds. Explainability Codex: Create universal "AI decision audit" templates for tax/welfare systems. Translate into 100+ languages - my taxpayers need Nepali/Maithili explanations. Climate AI Accelerator: Establish "AI for Adaptation" working group. My KAIST methane research (27% abatement) needs cross-border standards for 450M rice farmers. Current gap: High-income standards fail LDC reality. EU AI Act assumes cloud infra, English datasets, technical experts - Nepal has none. Dialogue succeeds by reverse-engineering: start with my IRO deployment constraints, build upward. Nepal coordinates GCF/World Bank daily - we implement when frameworks work. AI Dialogue becomes coordination hub linking Geneva theory to Kathmandu execution. Without this practitioner focus, Dialogue produces declarations gathering dust in New Road tax office. Success = standardized AI safety checklists deployed across 100 LDCs by 2028, closing $50B tax leakage + 1.5GtCO2e emissions. My IRO + KAIST lens reveals the stakes: revenue sovereignty, climate survival, digital trust. AI Dialogue succeeds by governing AI deployment, not just discussion.

AI Dialogue must integrate these practitioner-tested mechanisms: IMF SARTTAC (my tax training) → Scale to "AI Governance Officer" certification for 45,000 Nepal civil servants ITU Asia-Pacific AI Standards → LDC-adapt standards for tax/agriculture AI deployment GCF/World Bank Climate Finance (my coordination role) → "AI for Adaptation" facility with sovereign stack requirements South-South AI Collaboration (IndiaAI, LatAm-GPT) → Kathmandu-Delhi methane model sharing protocols OECD AI Tax Governance → Translate explainability frameworks into Nepali/Maithili for 8M taxpayers Current gap: These produce reports, not deployment toolkits. AI Dialogue's unique added value: 1. LDC Implementation Playbook: Convert OECD/ITU standards into "Day 1" checklists for Kathmandu tax officers - not Geneva consultants. My IRO needs plug-and-play AI audit templates tomorrow. 2. Sovereign AI Reference Stack: Open-source "Nepal AI Stack" (local servers + open models) tested across 10 LDCs. Revenue sovereignty demands we control our tax algorithms. 3. Capacity Funding Coordination: Pool $500M from ITU/IMF/World Bank into single "Global South AI Governance Fellowship" - 10,000 officials trained in 24 months. 4. Regional Testing Networks: Kathmandu-New Delhi-Bangkok "Rice AI Safety Network" - deploy my KAIST methane model across 450M farmers with shared governance protocols. 5. Multilingual Explainability Codex: 100+ language AI decision templates. My taxpayers need "why AI flagged my return" in Nepali, not English. Without AI Dialogue orchestration, these initiatives remain siloed. Nepal coordinates GCF/World Bank daily - we implement when frameworks connect. AI Dialogue succeeds by being the "missing middleware" between Geneva standards and Kathmandu deployment. My reality: 60% compliance gap, $1.2B green tax potential blocked by governance fragmentation.

Different stakeholders contribute uniquely to AI Dialogue success: Governments (my IRO role): Deploy real-world test cases. Nepal tests safety standards across 8M taxpayers before global scaling. Tax officers provide "deployment reality checks." Private sector: Build sovereign AI stacks. Indian startups could open-source Nepali/Maithili methane models serving 450M farmers. Need binding commitments, not marketing. Civil society: Test explainability in local languages. Nepali farmer cooperatives validate AI interfaces before national rollout. Academia (KAIST): Stress-test models. My 27% methane abatement research needs multi-country replication with governance guardrails. Technical community: Translate standards. Kathmandu tax officers cannot implement English-only protocols. AI Dialogue format/structure recommendations: 70% practitioner sessions: Tax officers, not consultants. "Kathmandu vs Geneva" deployment gap workshops. LDC regional hubs: Kathmandu hosts South Asia hub (July 2027). Test standards in rice economies first. "Fail-fast" pilots: Deploy AI safety standards in 10 LDCs within 6 months. My IRO volunteers first. Multilingual working groups: Nepali/Maithili technical papers mandatory. No English-only governance. Capacity track parallel: Train 1,000 officials per meeting using IMF SARTTAC model I completed. South-South speed dating: Nepal-India-Bangladesh "methane model governance" protocols in 15 minutes. Current format risk: Geneva academics debating while Kathmandu taxpayers lose $1.2B to ungoverned AI. Success = Kathmandu tax officer co-chairs next session. My reality: 60% compliance gap, zero AI governance experts among 45,000 civil servants. Dialogue must train us while theorizing. Recommendation: 50% agenda controlled by LDC practitioners. We implement global frameworks daily (GCF/World Bank). Let tax officers from New Road design the safety checklists, not Silicon Valley consultants. Inclusive ≠ token representation. Inclusive = LDC practitioners set 50% agenda.

Tax officers, Nepali rice farmers, Maithili-speaking smallholders - the 1.4B South Asians deploying AI daily, not discussing it. Who's missing: Public sector practitioners - IRO officers managing 8M taxpayers cannot attend Geneva. We face real deployment risks, not theoretical ones. Non-English civil servants - 45,000 Nepal government staff need Nepali/Maithili AI governance training, not English whitepapers. Smallholder farmers ($2/day incomes) - My KAIST methane model's 450M end-users have zero voice in safety standards affecting their livelihoods. LDC data custodians - Tax officers control national datasets but lack governance protocols to share methane/agricultural data regionally. How to include them: Regional practitioner hubs - Kathmandu hosts South Asia (July 2027). Tax officers co-chair, not observe. Multilingual live translation - Real-time Nepali/Maithili sessions mandatory. No English-only technical tracks. "Deployment reality" panels - 70% speakers from LDC public sector. My IRO volunteers first. Paid civil servant fellowships - $5K stipends bring 500 tax officers/farm ministry staff to Geneva. Citizen juries - Nepali farmer cooperatives test AI safety standards before global adoption. South-South video links - Kathmandu-New Delhi-Bangkok live practitioner sessions, no travel required. Current discussions = Silicon Valley + Geneva academics. Missing: the Kathmandu tax officer explaining AI decisions to non-technical taxpayers daily. True inclusion = LDC practitioners control 50% agenda. My reality: 60% digital compliance gap, zero AI governance capacity among 45,000 civil servants. We implement GCF/World Bank frameworks daily - let us design AI standards too. Without these voices, AI governance produces English-only frameworks gathering dust in New Road tax office. Farmers cannot eat whitepapers.

Replace Geneva monologues with Kathmandu deployment simulations: 1. "Live AI Audit Theater": Tax officers (like me) + farmers live-test safety standards on methane models. Real-time: "This algorithm flagged your return - explain in Nepali." Audience votes fixes. 2. LDC "Deployment Sprints": 48-hour hackathons. Kathmandu team builds sovereign AI stack for rice tax incentives. Delhi validates. Geneva documents. Winners present Day 2. 3. "Taxpayer Tribunal": Farmer cooperatives + citizens jury AI governance proposals. My IRO colleagues face real smallholders: "Will you trust this methane model?" Public verdict shapes outcomes. 4. South-South "Speed Governing": 10-minute rotations - Nepal tax officer + Indian AI startup + Bangladesh farmer co-design standards. 20 tables running parallel. 5. "Legacy Hell" Workshops: Simulate my IRO reality - deploy AI safety standards on 1990s COBOL tax systems. Practitioners document "Day 1 barriers" for global playbook. 6. Multilingual "Simultaneous Failure": Deploy identical AI model in English/Nepali/Maithili. Live-stream compliance gaps across languages. Real-time governance fixes. 7. "Revenue Roulette": Blind test - which AI governance framework closes my 60% compliance gap fastest? $1.2B revenue at stake. Practitioners score, not academics. Current format = PowerPoints nobody implements. My IRO needs plug-and-play checklists, not 100-page reports. Why these work: - 70% practitioner-led (tax officers > consultants) - Real stakes (my 8M taxpayers cannot afford bad standards) - Immediate output (standards tested before delegates leave Geneva) - Multilingual from start (Nepali/Maithili interfaces mandatory) - Success metric: Kathmandu tax officer deploys Geneva standards December 2027. AI Dialogue fails delivering English-only theory to Nepali/Maithili-speaking farmers. These formats ensure LDC practitioners control agenda, test solutions live, leave with implementable toolkits. My reality: 45,000 untrained civil servants, 60% compliance crisis. Let tax officers from New Road run the room.

Practical AI governance works best when it is simple, auditable, and adapted to local realities. In public administration, the most useful approaches are those that improve transparency, reduce discretion, and can be implemented in institutions with limited technical capacity. One relevant example is faceless assessment and digital tax administration, which shows how digital systems can strengthen transparency, accountability, and consistency when properly designed. Another useful practice is checklist-based governance, which can be adapted for AI risk screening, human oversight, and audit trails. Model cards and decision logs are also important because they help public institutions understand what an AI system does, what data it uses, and when human review is needed. This is especially valuable in high-impact sectors such as taxation, welfare, and climate-related services. I also support multilingual and low-bandwidth governance platforms so that AI tools can work in local languages and in settings with limited connectivity or technical infrastructure. For countries like Nepal, governance tools must be practical, not only sophisticated. Finally, capacity-building partnerships are essential. Training in tax administration, climate-related financial risk, and emerging policy leadership has shown me that practical, context-specific learning is more useful than abstract principles alone. In short, effective AI governance should combine transparency, human oversight, multilingual access, and implementation tools that public institutions in developing countries can actually use