From Paperwork to Proof: Can AI Make Tax Compliance Affordable for Small Firms?

A functioning tax system rests on a simple bargain: firms meet their obligations at a cost proportionate to what they owe. For small and medium enterprises (SMEs) in developing economies, that bargain rarely holds. Before a small firm can pay its tax, it must record transactions, classify them correctly, compute its liabilities and keep documentation that will survive an audit. Each step demands accounting capacity that many small firms simply do not have. For many owner-managers, that capacity is their own evenings, spent reconciling receipts rather than growing the business.

The tax small firms pay twice

Compliance cost works like a second tax, and it falls hardest on the smallest firms. Survey work across developing countries has found that compliance costs are regressive, weighing most heavily on small businesses relative to their size. South Africa's baseline survey shows how the burden accumulates: small businesses spent an average of 255 hours a year on tax compliance, with recording information alone taking more than half of that internal time.

When compliance is this expensive, firms respond predictably. They under-comply, file inaccurately or stay informal altogether. Joel Slemrod's review of the compliance literature shows that compliance depends on the costs and opportunities firms face, not only on their attitudes towards paying. The revenue authority therefore loses twice: it forgoes revenue, and it loses sight of firms it cannot affordably bring into the system.

Digitisation helps, but not automatically

Revenue authorities have already turned to technology, with real results. In Tajikistan, an experimental rollout of electronic filing among small and medium firms cut the time firms spent on taxes by 40 percent. Yet the gains are conditional. A World Bank study of South Africa, Ukraine and Nepal found that requiring paper submissions alongside e-filing roughly cancelled out the savings, and that small firms face a steep learning curve.

Zambia offers a live example. The Zambia Revenue Authority's Smart Invoice system validates invoices in near real time, and documents issued outside it cannot support input VAT claims. Systems like this digitise the authority's side of compliance. On their own, they do not give a small firm the accounting capacity to classify, compute and substantiate its obligations correctly.

What AI could add

This is where AI-driven tax-accounting automation enters. In principle, it could capture and classify transactions, compute obligations and assemble the documentation an authority requires, performing the accounting work that compliance demands without the in-house expertise a small firm lacks. It would also fit the direction tax administrations are already travelling. The OECD's Tax Administration 3.0 vision describes taxation becoming increasingly seamless, with compliance built into the systems businesses already use, and more than 80 percent of tax administrations in the OECD's Forum on Tax Administration are developing application programming interfaces to connect with third-party software. An AI layer sitting between a firm's transactions and the authority's systems is a plausible next step.

Fluent is not the same as correct

The promise, however, rests on an untested assumption: that the automation is right. Generative AI systems are known to produce fluent output that is unfaithful to their source or simply unfounded, a failure commonly called hallucination. In tax, that failure is costly. When researchers tested GPT-3 on statutory reasoning, they identified several types of clear errors and found poor performance even on simple synthetic statutes. Newer models may perform better, but tax rules in developing economies are complex, jurisdiction-specific and frequently amended, which is precisely where errors are hardest to detect.

An erroneous classification or computation is not a harmless glitch. A small firm that relied in good faith on automated output could face penalties, while a revenue authority may decline to accept computations or documentation it cannot audit. Unvalidated automation could therefore add risk to the very firms it is meant to help.

Three conditions to test

The research question is therefore not whether AI can automate tax accounting, but whether it can do so safely for SMEs in developing economies. Three conditions deserve rigorous study. The first is accuracy: how reliably automated systems classify transactions and compute liabilities under local rules. The second is trust: when small-business owners understand, rely on and appropriately check AI-generated tax outputs. The third is authority acceptance: whether, and on what terms, revenue authorities will recognise automated computations and supporting documentation.

Evidence on these three conditions would give revenue authorities, policymakers and technology providers a sound basis for tools that make compliance affordable, widen the tax base and draw informal firms into the system. Without it, AI in tax risks being adopted on enthusiasm rather than proof. For the small firm that already pays twice, getting this right matters. The same test applies to emerging technologies in any sector: judge them by evidence of what works, for whom and at what risk.

Sources

Primary sources

•      OECD (2020). Tax Administration 3.0: The Digital Transformation of Tax Administration. Paris: OECD Publishing

•      OECD (2025). Tax Administration Digitalisation and Digital Transformation Initiatives. Paris: OECD Publishing

•      VATupdate (2026). Zambia: E-Invoicing and E-Reporting Country Booklet (industry summary of the ZRA Smart Invoice regime)

Research

•      Blair-Stanek, A., Holzenberger, N. and Van Durme, B. (2023). Can GPT-3 perform statutory reasoning?. Proceedings of the Nineteenth International Conference on Artificial Intelligence and Law (ICAIL 2023), pp. 22–31

•      Coolidge, J. (2012). Findings of tax compliance cost surveys in developing countries. eJournal of Tax Research, 10(2), pp. 250–287

•      Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y.J., Madotto, A. and Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), Article 248, pp. 1–38

•      Okunogbe, O. and Pouliquen, V. (2022). Technology, taxation, and corruption: evidence from the introduction of electronic tax filing. American Economic Journal: Economic Policy, 14(1), pp. 341–372

•      Slemrod, J. (2019). Tax compliance and enforcement. Journal of Economic Literature, 57(4), pp. 904–954

•      Smulders, S., Stiglingh, M., Franzsen, R. and Fletcher, L. (2012). Tax compliance costs for the small business sector in South Africa: establishing a baseline. eJournal of Tax Research, 10(2), pp. 184–226

•      Yilmaz, F. and Coolidge, J. (2013). Can e-filing reduce tax compliance costs in developing countries?. Policy Research Working Paper 6647. Washington, DC: World Bank