Intelligent taxation and AI: Promises and pitfalls

This post is part of a new series on intelligent taxation and AI, which brings together evidence, policy lessons, and challenges of using AI to strengthen tax administration. Explore the Tax for Growth initiative’s new policy toolkit: Harnessing AI and data for tax administration, which synthesises lessons for governments and Commonwealth tax administrators.

Tax administrations constitute a pillar of government – they are tasked with collecting taxes from broad segments of society, limiting evasion and avoidance, acting as an intermediary between citizens and the state, and integrating domestic and international policies. 

Many tax authorities operate in structurally challenging environments, with limited staff and resources, and their work is based on uneven underlying data. Artificial intelligence (AI) has the potential to assist with these tasks, but implementing AI for taxation must be done intelligently – as part of a wider reform agenda that aims to create better data, clearer processes, and stronger accountability.

AI for taxation is not some future possibility; it is already here. A survey conducted as part of a collaboration between the International Growth Centre’s Tax for Growth (T4G) initiative and the Commonwealth Association of Tax Administrators (CATA) found that by the end of 2025, 80% of tax authorities had already deployed AI-based processes or were close to doing so. Similarly, in the OECD’s tax technology inventory, 48.5% of administrations report using AI.

Where do the major promises for AI in taxation lie? 

• Smarter enforcement with less red tape: A high-return use case of AI is likely to be better targeting – choosing which cases to audit, which invoices to scrutinise, or which import shipments to inspect. If targeting improves, agencies can raise revenue while reducing unnecessary inspections and compliance costs. In Italy, recent research found that replacing the 10% least productive status quo audits with cases selected by a trained algorithm raised detected evasion by 38%.
• Taxpayer service at scale: AI can help taxpayers navigate complex systems through better triage, clearer guidance, and faster resolution of routine queries. Even incremental improvements matter when millions of interactions are involved. For example, in the United Kingdom, HM Revenue & Customs’ “Ask HMRC online” digital assistant recorded 5.48 million interactions in 2024-25¹.
• Data plumbing’ that makes risk management possible: Many administrations struggle with scanned forms, inconsistent identifiers, and fragmented databases. AI methods for document processing, entity matching, and classification can help convert paperwork into structured data, which is often a prerequisite for any serious risk‑based enforcement.

What works when using AI in tax administration? 

1. When the signal is clear, enforcement nudges can unlock large gains: In Ecuador, the tax authority identified clients of “ghost firms” (fake suppliers selling fraudulent deductions). Within three months, identified firms made back payments totalling USD 20.6 million – an 81% increase. This work highlights how the returns to compliance and enforcement may be substantial if AI is deployed for better identification.
2. Accuracy of AI signals does not guarantee field impact: At the same time, evidence from India’s tax administration highlights a gap between prediction and policy. A dataset of tax returns in an Indian state reveals that while machine-learning‑driven inspections can accurately identify ghost firms, they do not necessarily increase enforcement outcomes. This is because the models are trained on proxies that imperfectly map to legally actionable violations.
3. Algorithmic gains depend on implementation: In Senegal, a comparison of algorithmically selected tax audits with inspector‑selected audits at scale finds that the main bottleneck is implementation. Algorithm-selected full audits are 18 percentage points less likely to be executed by officials than inspector-selected ones. Moreover, inspector-selected audits uncover larger evasion amounts, and firms rate algorithm-selected audits less positively.
4. AI can boost hit rates by concentrating scarce attention: In Paraguay, a risk algorithm has been designed to focus inspections on a smaller set of high-risk shipments. Results suggest that fraud detection among inspected shipments rises from 18% under the status quo to 73% when using the algorithm. This cuts the number of full inspections by around 70%, thereby improving the speed of customs clearance for importers.

What are the operational and governance challenges of using AI for taxation?

In many administrations, the binding constraint may not be sophisticated modelling but rather bottlenecks along the enforcement pipeline: linking data across silos, turning flags into cases, executing audits, and collecting assessed liabilities. The work from India shows that realised collections did not significantly increase when using AI, because auditor teams faced workflow frictions and resource constraints that limited follow‑through.
The notion of ‘fully automated’ enforcement may not be desirable, as AI outputs can trigger legal consequences. Administrations need to implement governance processes that alert taxpayers when they are interacting with AI or are subject to AI‑influenced decisions, and ensure that processes are transparent and include grievance redress mechanisms.

Before fully committing to AI, tax administrations can – and perhaps should – reap high returns by first focusing on fundamental steps, including in data governance and integration (resolving identifier mismatches, basic data processing pipelines, and quality controls), and design processes that mitigate bias from selective labels. 

How can policymakers manage risks in an AI-based tax system?

• Selective labels and feedback loops: Models trained on historically audited cases can inherit past biases. Random audits and periodic reverification help minimise this risk.
• Optimising the wrong metric: Accurately detecting evasion does not automatically translate into revenue, and revenue is not the same as welfare. Agencies need to define explicit objectives and make it clear how their performance metrics map onto those objectives.
• Gaming and leakage: Predictable rules or leakage can invite adaptation.
• Cybersecurity risk: AI raises the stakes for data security and insider threats. Generative AI can also tempt staff to paste sensitive data into external services.
• Fairness, appeals, and trust: Risk models can concentrate enforcement on certain sectors or regions. Publishing a register of AI use cases, transparent reviews, and accessible appeals can protect legitimacy.
• Vendor lock‑in and capability gaps: Tax administrations need in‑house capacity to monitor drift and update systems as laws and behaviours change.

A robust roadmap for a tax authority that has the intention of implementing AI may include the following:

1. Start by building the foundations: Digitise core workflows, improve identifiers and data quality, link core datasets (value-added tax (VAT), income tax, customs, third‑party reporting), and establish protocols for access and review.
2. Focus initially on high-signal use cases: Pilot risk targeting where outcomes can be measured quickly and the operational pipeline is under agency control (case assignment, follow‑through, and collections).
3. Evaluate, then scale: Use phased roll‑outs (and randomisation where feasible). Track revenue, delays, compliance costs, and corruption risks. Invest in the routine functions: logging, dashboards, and case management.

Hear more from Anders Jensen and others about intelligent taxation and the use of AI in the Harnessing AI and Data for Tax Administration Webinar, organised by the International Growth Centre in collaboration with the Commonwealth Association of Tax Administrators.