Digital tax administration cannot be built on borrowed intelligence. If countries want to modernize their tax systems, they must turn their own laws, data, and enforcement logic into trusted national infrastructure. Otherwise, they risk outsourcing not only technology, but also part of their fiscal sovereignty.

The new frontier is not a chatbot

The artificial intelligence debate has been captured by the visible part of the technology. Executives admire systems that can draft emails, summarize legal PDFs and generate polished text in seconds. Governments, too, are under pressure to show that they are not missing the AI moment. Yet the most important use of AI in public administration will not be measured by the fluency of a chatbot. It will be measured by whether a country can turn its own legal and administrative knowledge into systems that are accurate, explainable, auditable and under national control.

Tax compliance is the place where this shift becomes impossible to ignore. Tax is not a normal business process. It is one of the core mechanisms through which the state finances itself, allocates incentives, observes the economy and builds trust with citizens. When a country digitizes tax, it is not only modernizing forms and portals. It is creating a nervous system for the economy.

For developing countries, this nervous system cannot be outsourced without consequences. A tax administration may own the law on paper, but if the interpretation, classification, risk scoring, reporting logic and transaction evidence are processed through external AI platforms, then part of the state’s regulatory intelligence has effectively moved outside its control. That is why AI sovereignty and tax compliance now belong in the same conversation.

The illusion of borrowed intelligence

Borrowed intelligence can be useful for general productivity, translation, research support and education. It becomes dangerous when it enters the administrative core of a country without a clear governance model. A large commercial language model can summarize a tax law, but it cannot become the law. It can assist an auditor, but it cannot be allowed to create untraceable reasoning behind enforcement. It can help a software vendor understand a compliance obligation, but it cannot replace deterministic requirements that a retailer or POS provider must implement.

The problem is not only accuracy; it is jurisdiction, institutional control and legal exposure. Sensitive fiscal data includes invoices, receipts, transaction logs, business identifiers, payment patterns, refund behavior, audit trails and risk signals. When those data streams are used in AI workflows, the key question is not simply where the data is stored. The deeper question is who controls the infrastructure, who can access the logs, who governs model training, who owns the embeddings, who can change the rules and which legal system applies when conflicts arise.

The legal reality is already complicated. Under the U.S. CLOUD Act, a provider of electronic communication or remote computing services must comply with disclosure obligations for data within its possession, custody or control, regardless of whether that data is located inside or outside the United States. The European Union, by contrast, has built a transfer regime under which personal data leaving the European Economic Area must remain protected through adequacy decisions, contractual clauses, binding corporate rules or other safeguards. These legal architectures were not designed specifically for AI tax systems, but they show why data sovereignty is no longer an abstract policy slogan.

Developing countries should not read this as a reason to reject international technology. They should read it as a reason to separate convenience from sovereignty. General AI services may support drafting, education and non-sensitive analysis. National compliance intelligence, however, must be governed as critical infrastructure.

Tax Administration 3.0 is already pointing in this direction

The direction is visible in the work of international tax institutions. The OECD’s Tax Administration 3.0 vision describes a future in which taxation becomes more seamless and frictionless over time, increasingly embedded in the natural systems that taxpayers and businesses already use. That is a profound change. It means tax is moving upstream, closer to the transaction, closer to accounting systems, closer to e-invoicing networks, closer to POS systems and closer to the moment where economic activity actually happens.

The OECD has also noted that tax administrations have used AI across their operating models for years, particularly because collecting and analyzing data is already central to tax administration. The main areas include fraud detection, decision-support and better taxpayer services. This confirms what many practitioners already see in the field: the tax authority of the future will not be a passive receiver of declarations. It will be a data-driven institution that continuously compares reported activity with expected behavior, known risks and regulatory requirements.

For developing countries, the potential is significant. AI can help detect fraud rings, identify duplicate filings, flag identity mismatches, prioritize audits and improve taxpayer service. Armenia offers a concrete example. Supported by the World Bank, the country has been piloting an AI-powered tax administration tool, with the American University of Armenia involved in capacity building. The first use case focuses on reading invoices, detecting fraud rings and identifying anomalies, and the World Bank blog reporting on the project cites conservative expectations of a 10% to 15% increase in voluntary and enforced compliance through AI.

This is not science fiction; it is the beginning of smart tax administration. The more important issue is whether such systems are built as sovereign, explainable and locally governed capacity, or whether they become another layer of dependency on external platforms.

The missing layer is compliance intelligence

The next step is not to add a chatbot to a tax portal. The missing layer is Compliance Intelligence. This layer connects official legal sources, administrative guidance, technical schemas, process models, certification rules, audit requirements and implementation logic into a source-grounded knowledge system. It does not merely answer what a regulation says. It explains which process is affected, which data fields are required, which evidence must be retained, which timing rules apply and how a business system should behave.

In retail, this distinction matters. Fiscalization rules are not only legal text. They affect receipt creation, fiscal signatures, transaction sequencing, offline scenarios, refunds, self-checkout, e-commerce flows, cash management, middleware, archiving, reporting and audit exports. A retailer operating across countries cannot rely on a generic answer that sounds plausible. A POS provider cannot implement an obligation that has no version, no source, no validation date and no testable interpretation.

Compliance Intelligence should therefore combine several layers. The first is a trusted source layer with laws, decrees, technical specifications, tax authority guidance and certification documents. The second is a semantic layer that maps legal concepts to business processes, data objects and system responsibilities. The third is a deterministic layer with rules, validations, schemas, test cases and executable requirements. The fourth is an AI interface that allows humans to search, compare, explain and reason across the knowledge base without losing the audit trail.

In this architecture, a language model is not the authority. It is an interface to authority. The authority remains the source, the versioned interpretation, the expert validation and the deterministic compliance model that can be tested.

Why this matters most for developing countries

Many developing countries face the same structural challenge. They need to increase domestic revenue, reduce informality, modernize administration and make compliance less painful for honest businesses. At the same time, they often have limited technical capacity, fragmented registries, uneven digital infrastructure, scarce AI talent and high dependency on foreign vendors. This combination makes them both ideal candidates for AI-enabled tax modernization and highly vulnerable to dependency if modernization is designed poorly.

The global development ecosystem increasingly understands this risk. The United Nations Global Dialogue on AI Governance explicitly includes bridging AI divides, capacity-building, access, digital foundations, high-performance computing skills, open-source software, open data and open AI models among its thematic priorities. UNDP describes digital public infrastructure as foundational systems that enable secure interactions between people, businesses and governments, including identity, payments and data exchange. The Digital Public Goods Alliance defines digital public goods as open-source software, open data, open AI models, open standards and open content that comply with privacy and other legal safeguards. IDRC’s AI for Development program supports responsible AI ecosystems where local experts can solve their own development challenges and participate in governance debates.

The pattern across these initiatives is clear. The world is moving toward the idea that AI capacity is not only a matter of buying tools. It requires local expertise, representative data, public infrastructure, governance and institutional maturity. Lacuna Fund makes the same point from the data side by supporting datasets for underserved communities, low-resource languages and locally relevant AI applications. Without local data and local context, AI systems do not merely become less useful. They can become structurally blind.

Tax compliance is one of the areas where that blindness would be expensive. A model trained on general legal text may not understand local invoice practices, informal trading patterns, language variations, retail workflows, tax authority procedures or the practical difference between what the law says and how systems are certified. A sovereign compliance intelligence layer can preserve that local context and make it usable.

Sovereignty does not mean isolation

The argument for AI sovereignty is sometimes misunderstood as an argument for technological isolation. That would be the wrong conclusion. Few countries can or should build every model, every chip, every cloud platform and every application alone. The smarter strategy is selective sovereignty. Countries should know which layers can be sourced from global markets, which layers should be based on open standards and digital public goods, and which layers must remain under domestic control.

In tax compliance, the most sensitive layers are not always the largest or most expensive. They are the legal knowledge base, the transaction evidence, the compliance rules, the risk models, the audit history and the governance process around automated decisions. A country may use open-source models, regional cloud providers, international standards and external expertise. It should still retain control over the authoritative compliance knowledge, the training and evaluation data that reflects its own economy, and the decision frameworks that affect taxpayers.

This is also where universities and research institutions become essential. Stanford’s RegLab has shown how machine learning and data science can be used to study government decision-making and improve public administration. The American University of Armenia’s role in the Armenia tax AI project shows how local academic capacity can support government modernization. Oxford’s work on international AI governance shows that AI is now inseparable from international politics, benefit-sharing and risk mitigation. For developing countries, partnerships with universities are not symbolic. They are a way to build national capability instead of only procuring software.

From regulation to executable compliance

The final destination is executable compliance. In this model, regulation is not left as a static PDF, and compliance guidance is not buried in long documents that only a few experts can interpret. Legal and technical requirements are translated into structured, versioned and testable models that systems can use. Human experts remain responsible for interpretation, but the interpretation becomes reusable by software.

For a retailer or POS provider, this changes the practical meaning of compliance. Instead of asking a consultant to interpret a country requirement from the beginning each time, the company could rely on a compliance intelligence layer that connects the official source to business processes and implementation tasks. It could identify whether a refund requires a reference to the original receipt, whether an offline transaction must be chained later, whether a QR code is mandatory, which invoice status must be monitored and which audit export must be available.

For developing countries, such a capability could become a strategic asset. It would reduce dependency on foreign implementation knowledge, help local software vendors build compliant systems, make tax administration more transparent, and create a common language between regulators, businesses and technology providers. It would also allow countries to publish compliance expectations in a form that is easier to implement, easier to test and easier to update when the law changes.

The governance burden

There is one unavoidable warning that should stay at the center of every AI tax strategy. A smart tax system without governance can become a dangerous system. Automated risk scoring, AI-assisted audits and predictive enforcement must be explainable, contestable and subject to human oversight. Taxpayers need rights, appeal mechanisms and transparency about how evidence is used. Governments need audit logs, model evaluation, data minimization, cybersecurity controls and independent review. The more powerful the compliance engine becomes, the more important its governance becomes.

This is the reason why AI in tax should not be left only to technology departments. It should be equally managed by tax authorities, ministries of finance, data protection regulators, courts, universities, civil society and the local software industry. Trust is not built by automation alone. Trust is built when automation is bounded by law, evidence, accountability and competent institutions.

The strategic choice

Developing countries now face a strategic choice. They can adopt AI as a layer of rented productivity, using external systems to accelerate tasks while leaving the deeper structures of regulation unchanged. Or they can use AI as an opportunity to build national compliance intelligence, turning their own fiscal law, economic data and administrative experience into a sovereign digital capability.

The first path is faster and easier to justify in a budget cycle, but the second path is the one that creates lasting capacity.

AI should be treated as an asset of humanity, not as a privilege controlled by a small number of companies or countries. Yet access alone is not enough. A country that can use AI but cannot govern the intelligence layer behind its own tax system has not achieved sovereignty. It has only upgraded its dependency.

For developing countries, the prize is larger than administrative efficiency. The prize is the ability to finance public goods, support local innovation, protect citizens, strengthen businesses and participate in the AI economy as makers rather than passive consumers.

The countries that own their compliance intelligence will own a meaningful part of their economic future.