AI in Tax Administration (What 80% Use, 10% Govern)

I sat in a workshop last week where an IMF presenter asked three questions in a row.

"How many of you use ChatGPT at home?"

Every hand in the room went up.

"How many of you use it for work?"

Every hand went up again.

"How many of you have an official AI policy in your administration?"

Not a single hand.

That moment captured everything you need to know about where tax administrations actually stand with AI in 2026. We are already using it. We are just pretending we are not.

Quick answer

Roughly 80% of advanced-economy tax administrations now use AI in some form. About half of emerging-market administrations do. But fewer than 1 in 10 have an actual AI policy on paper. Most of the AI in use today is informal, unmanaged, and walking around in employees' pockets.

That gap is the real story.

What the IMF Survey Actually Says

The numbers come from ISORA, the International Survey on Revenue Administration. The IMF runs it across its member countries, and the latest results give us the cleanest view of where we are.

A few statistics that frame the conversation:

• About 1 billion taxpayers are registered in tax administration IT systems worldwide.
• Around 20% of global GDP flows through tax administrations every year.
• Roughly 80% of advanced-economy tax administrations report using AI today.
• About 95% have plans to expand its use.
• For emerging markets, AI adoption sits closer to 50%.

If you compare the revenue of the IRS to the revenue of the largest US tech companies, the IRS is bigger. Tax administrations sit on enormous volumes of data. That is exactly why AI matters here, and exactly why we should be careful with it.

The Digitalization Divide

Adoption is not evenly spread. The IMF's data shows a clear pattern. Larger tax administrations tend to be more digitized. Smaller ones lag behind.

Here is the part most people miss. Smaller administrations have a stronger business case for AI, not a weaker one. If you have 800 IT staff in a workforce of 20,000, you can absorb a slow AI rollout. If your IT team is 30 people in a workforce of 1,500, AI is one of the only realistic ways to close the gap.

The data also shows another useful correlation. More digitalization tends to reduce the compliance gap. The link between IT spending, AI capability, and tax revenue is empirical, not just intuitive.

There is a second pattern worth flagging. Emerging-market tax administrations spend roughly 4.5% of their budget on IT, on average. Advanced economies spend significantly more. The smaller the administration, the smaller the slice usually goes to security and AI governance specifically.

Who Is Actually Using AI Right Now

When tax administrations describe their AI use, the answers split sharply by income group.

Advanced economies typically have:

• Around 80% current AI adoption
• About 95% with planned expansions
• In-house data scientists
• Dedicated machine learning use cases like case selection, fraud detection, and segmentation

Emerging markets typically have:

• Around 50% current adoption
• Heavier reliance on commercial vendors or open-source tools
• Fewer or no dedicated data science teams
• AI features bundled inside larger IT contracts, often invisible to leadership

In the workshop, the most common tools I heard mentioned were R, KNIME (drag-and-drop machine learning), SAS, SPSS, and Tableau. None of those are new. What is new is generative AI sitting on every employee's phone.

The Shadow AI Problem

This brings us back to the three questions from the start.

Almost everyone in a modern tax administration is using AI tools daily. They use ChatGPT to draft emails. They use Claude to summarize meeting notes. They use Gemini to translate documents. Some use it to write code.

What almost no tax administration has is a written policy on any of this.

That gap is the single biggest governance issue in tax administration AI right now. It is not the algorithms. It is not the models. It is the fact that 10,000 employees are quietly making 10,000 individual decisions about what to paste into a public AI service.

The IMF's own information security team recently reported on the volume of internal data leaving the organization through external AI services. The numbers were uncomfortable. If that is happening inside the IMF, it is happening inside your administration too.

Where AI Is Actually Being Used

One survey result surprised the room. When tax administrations report what they use AI for, the biggest category is not citizen service. It is not internal productivity. It is enforcement.

Enforcement covers:

• Audit case selection
• Fraud detection
• Risk segmentation
• Anomaly detection in returns and invoices

This is also the most legally sensitive use of AI in tax. It is the area where a court can demand to know exactly why a case was selected. It is the area where the Netherlands childcare benefits scandal happened, where bad training data eventually brought down a government.

Most administrations are starting AI exactly where the risk is highest. That is worth thinking about.

What This Means for Your Administration

If you are a CIO, a head of IT, or a senior auditor reading this, the practical takeaways are short.

1. Your staff are already using AI. The question is whether they are using it safely.
2. You need a written AI policy. Even a one-page document is better than nothing. Spell out what is allowed, what is forbidden, and what data must never leave the network.
3. Start with a use case inventory. You cannot govern what you cannot see. A simple spreadsheet listing every place AI touches your operations is the first step.
4. Risk-assess each use case. The IMF publishes a framework that looks at input, throughput, and output. I will cover it in detail in a later post.
5. Begin with low-risk areas. Internal productivity, summarization, and translation are safer first steps than case selection or fraud scoring.

The Bottom Line

The conversation about AI in tax administration has shifted. Two years ago the question was "should we use AI?" Today the honest question is "do we know how our people are already using it?"

If you cannot answer that second question, the policy gap is your problem. Closing it is cheaper, faster, and lower-risk than any algorithm you might buy.

That is where every serious AI program in a tax administration has to start.

Frequently Asked Questions

What is AI in tax administration?

AI in tax administration is the use of machine learning, generative AI, and other algorithmic tools to support work like audit case selection, fraud detection, taxpayer services, risk assessment, and internal productivity. It includes both purpose-built systems and general tools like ChatGPT used by staff.

How many tax administrations use AI?

According to the IMF's ISORA survey, roughly 80% of advanced-economy tax administrations use AI today, while about 50% of emerging-market administrations do. Almost 95% of advanced economies have planned expansions.

Why do most tax administrations not have an AI policy?

AI adoption has moved faster than governance. Most administrations have general data security policies, but few have AI-specific rules covering acceptable use, data handling, human oversight, and accountability. The result is widespread informal use without documented limits.

What is the biggest AI risk for tax administrations?

The biggest risk is enforcement use without explainability. If a case selected by AI ends up in court, the administration must be able to defend the logic clearly. Black-box algorithms make that very difficult, as the Netherlands childcare benefits scandal showed.

What is the first step to adopting AI responsibly?

Start with a written AI policy and a use case inventory. Both are low-cost, low-risk, and create the foundation for everything else. Once you can see where AI is already in use, you can start risk-assessing it.

What tools are tax administrations using for AI today?

The most common are R and RStudio, KNIME, SAS, SPSS, Tableau for analytics, plus general-purpose generative AI tools like ChatGPT, Claude, and Gemini for productivity. Larger administrations often have custom-built machine learning pipelines as well.

Suggested internal links (future posts in this series)

• The history of AI in tax and customs (1970s to today)
• What counts as AI: a working definition for tax professionals
• When AI goes wrong: the Netherlands childcare benefits scandal explained
• Generative AI for tax administrations: what is actually different this time
• AI agents in tax and customs: the next shift
• How to risk-assess an AI use case in a tax administration

Suggested external authority sources

• IMF data portal and ISORA survey results
• IMF technical notes on AI in revenue administration (search the IMF eLibrary)
• OECD Tax Administration series reports
• David Hadwen's public AI use cases database for tax administrations
• The Dutch parliamentary inquiry on the childcare benefits affair (toeslagenaffaire)
• Brazil's RFB published AI policy (frequently cited as a current reference model)

If this resonated, the next post in the series goes deeper. I unpack what actually counts as AI inside a tax administration, why the autopilot on an airplane is a useful thought experiment, and why the IRS has been running an AI system since the 1970s without anyone calling it that.

Read it next...