The new World Bank World Development Report argues that tax administration is one of the more plausible early use cases for effective use of AI by governments. Yet AI for development has been largely focussed on domains such as health, agriculture and education, with far less attention paid to the potential transformative impact on governance.
Nevertheless, revenue authorities in low- and middle-income countries are pioneering innovative uses of AI. Some applications are operational, others still at pilot or design stage. However, the conversation tends to start with the tool itself. What can AI do? Where could it be introduced next? Which model should an administration use? It starts less often with what those tools need to work, or how this affects taxpayers’ rights.
At the Future of Tax Reform Forum hosted in July by the International Centre for Tax and Development (ICTD) in partnership with the Zambia Revenue Authority in Lusaka, we convened a working group of experts to explore what systems are needed for an effective and responsible usage of AI in tax administrations. The group comprised representatives from African tax administrations, academic institutions, civil society organisations, and development partners. They reviewed what data, digital infrastructure, capacity, and governance revenue authorities need to use AI effectively and responsibly. Four messages stood out:
Lesson 1: AI depends on good data and the systems already in place
A working group member’s comment captured a crucial issue:
“If your AI output isn’t good, blame your data.”
AI works with the data available to it. In tax administration, that means taxpayer registration, returns, payments, customs declarations, invoices, audits, or third-party data. When records are incomplete, duplicated or outdated, the results reflect those weaknesses. The same problem arises when filing, payment, and customs systems cannot communicate with each other, or when different agencies record the same taxpayer under different identifiers. The work required to address these problems often happens long before an AI tool is introduced. It includes cleaning taxpayer registers, connecting databases, improving the quality of historical records and agreeing on common standards for exchanging information.
However, digital platforms are not enough to address the data problem. Tax administrations may offer e-filing and e-payment while its taxpayer records remain fragmented, or while access to information held elsewhere in government remains limited. Digital Public Infrastructure (DPI) therefore is a key part of the foundations needed for AI implementation. Digital identity helps link people and businesses across public systems; digital payments create records of transactions and payment behaviour; and secure data exchange allows revenue authorities to connect their information with business registration, customs, social security, procurement, and other government systems.
Furthermore, AI and DPI require foundational enablers of digital connectivity, literacy, and skills, both for governments and citizens. Digital exclusion persists, and nine in ten people in sub-Saharan Africa are not covered by a 5G network. The gap is also gendered. About 280 million more men than women were using the internet in 2025. This is not only an access problem. Taxpayers who remain outside digital systems are thinly represented in the records that AI tools learn from and are the least likely to benefit from AI-supported services.
Lesson 2: Identifying administrative problems should come prior to AI adoption
The cases presented in Lusaka showed how varied AI use in tax administration has become:
- Kenya: a chatbot delivered through WhatsApp helps taxpayers complete routine tasks and find information through a platform they already use. It reflects a phrase heard repeatedly during the working group: “meet taxpayers where they are”.
- Zambia: a nudging system draws on taxpayers’ previous filing behaviour and sends different messages to on-time filers, late filers, and non-filers, replacing broad reminders with communication that better reflects each taxpayer’s circumstances.
- Ghana: Publican AI supports risk profiling of import declarations, helping customs officers identify shipments requiring closer attention. It responds to long-standing problems such as undervaluation, misclassification and false declarations.
- Pakistan: Digital Eye uses cameras and AI to estimate production in manufacturing, developed in response to under-reporting and the limits of continuous physical inspection.
These examples address different administrative problems and rely on different combinations of data, digital systems, staff capacity, and oversight. They also show that AI should not be the starting point. Revenue authorities first need to identify the problem they want to solve, then assess whether AI is the most appropriate tool.
As one participant said:
“Instead of asking if a tax administration is using AI, let’s ask what problem the administration is trying to solve?”
The answer determines what information is needed and whether AI adds anything beyond existing tools. Some problems can be addressed through clearer procedures, better data or relatively simple automation. Other tasks may benefit from machine learning, image recognition or language-based systems. The choice should follow the administrative need.
There is also no single route towards AI adoption. Some revenue authorities are improving basic data systems and interoperability, others have started with chatbots or targeted pilots, and Zambia has placed an AI Unit within its Innovation and Project Management Department. Each path raises different questions about cost, capacity, maintenance and ownership.
Lesson 3: Limitations of AI use especially around complex, non-routine decision-making
The discussion became more cautious when it moved from routine services to decisions affecting individual taxpayers.
One participant asked:
“Is AI the first interaction we want with our clients?”
For simple requests, an automated first interaction may be useful. A taxpayer looking for a filing deadline, a payment reference or guidance on registration may value a quick response at any hour. The picture changes once a system contributes to an audit, a refund decision or an enforcement action. A taxpayer selected for audit should be able to understand the basis for that decision. Someone whose refund is delayed should receive an explanation. Where a system flags a business as high risk, officers need enough information to judge whether the result makes sense. Taxpayers also need a route to question or challenge decisions affecting them, and the bodies that hear those challenges need the capacity to interrogate decisions an algorithm helped shape.
This creates a limitation of AI usage. Some advanced models produce accurate predictions without providing a clear explanation of how they reached each result. That might be acceptable for a low-stakes internal task, but it becomes far more problematic when the output affects a taxpayer’s rights, money or legal obligations.
The level of explanation and human involvement should therefore vary by task. A chatbot answering standard questions does not carry the same risks as a model that could lead to a refund being blocked. Some tools may operate with limited day-to-day human review once tested; others should remain decision-support only, with an officer taking responsibility for the final call. AI can support human judgement, but legal responsibility stays with the revenue authority.
Lesson 4: Governance and accountability must be clear from the start when using AI tools
Governance was discussed through practical questions that administrations tend to face later than they should. Who owns an AI application once it is in use? Who checks whether the data remain reliable? Who responds when a taxpayer challenges a decision? These become harder when responsibilities are split between business departments, IT teams, data scientists, vendors, and senior management. The sessions on institutionalising AI stressed the need for clear ownership: operational departments define the problem and remain accountable for how outputs are used, technical teams build and maintain the systems, and senior management keeps projects aligned with wider objectives.
Ownership does not end at the design stage. A strong model can still fail if officers do not trust it or if the output does not fit existing workflows, and a pilot may become unusable because nobody budgeted for maintenance. Monitoring therefore needs to continue after deployment, with processes for correcting errors and withdrawing systems that no longer perform well. Dependency on external providers becomes a concern when the administration cannot examine, adapt, or maintain the system itself, so contracts need to address data access, ownership of models and knowledge transfer.
Governance also covers the everyday use of generative AI by staff, who may already be drafting letters or summarising documents with public tools. Waiting for a major AI project before setting rules on what can be entered into them leaves an immediate gap.
The 2026 edition of the Global Index on Responsible AI shows that only 26 per cent of countries have frameworks for fair, accountable public procurement of AI, and only 18 per cent require public disclosure of government algorithmic systems. Every safeguard discussed here, from contesting a decision to auditing a model for bias, presupposes that someone can first know an AI system was involved, and in sub-Saharan Africa that precondition is unmet almost everywhere.
Disclosure is not only a matter of taxpayer rights, but also of the administration’s own interest. Voluntary compliance depends on perceived fairness, so any erosion of trust in AI-influenced enforcement has second-order effects on revenue.
Six practical priorities follow, for revenue authorities, governments, and development partners:
- Build AI readiness before procuring AI tools, by strengthening data quality, interoperability, legal frameworks, and internal capacity.
- Invest in reliable taxpayer identification and lawful data linkage across tax, customs, business registration, payment, and third-party systems.
- Develop secure and accountable data exchange arrangements, including clear data-sharing agreements, audit trails. and privacy safeguards.
- Keep humans in the loop, so that AI supports professional judgement rather than replacing it.
- Start with narrow, high-value use cases and evaluate AI tools before scaling them.
- Be transparent about AI use. Transparency has three components: openness, or what information on AI is shared; intelligibility, or whether the receiver can understand it; and explainability, or why an AI-influenced decision was taken.
Reliable administrative data, digital public infrastructure, staff capacity, and governance arrangements all shape what AI can achieve. Adoption is early enough in most African revenue authorities to build these foundations alongside the tools.