AI in revenue administration: help and risk

Why this matters for revenue officers

Revenue work is information work. Returns, declarations, registrations, notices, objections, audit files and correspondence all pass through your hands, often in large volumes and under deadlines. AI tools are good at some of this: drafting, summarising, translating and spotting patterns in large amounts of data. That makes them attractive to revenue departments everywhere.

Revenue work also carries legal power. A notice can create a liability. An audit selection can cost a business time and money. A mistake in a taxpayer's record can follow them for years. When AI is involved, these consequences do not change, but the ways things can go wrong do. This module sets out where AI can help, where it adds risk and the principles to hold on to throughout this pathway.

Where AI can help

Most of the safe, useful work is the same as in any office, applied to revenue tasks:

• Drafting and plain language. Turning technical guidance into clear notices, FAQs and replies that taxpayers can understand, in English and Urdu (module 2).
• Summarising. Getting through long objections, audit reports or correspondence quickly, then checking the points that matter against the original.
• Organising information. Pulling figures or headings out of public documents into tables, with careful checking.
• Internal knowledge. Helping staff find the right section of a long internal manual, as long as the answer is checked against the manual itself.

A second category is more sensitive: tools that analyse taxpayer data to find patterns, predict non-compliance or flag cases for attention. This is often called predictive analytics. It can help a department use limited audit capacity more sensibly, but it raises serious questions about fairness and explanation (module 3).

Where AI creates risk

• Invented law. AI assistants produce hallucinations: sections, schedules, rates, deadlines or rulings that do not exist. In revenue work, a confident but wrong reference can mislead a taxpayer or an officer.
• Out-of-date law. Tax law changes often. An AI tool may describe an earlier position or another country's rules as if they were current here.
• Confidentiality. Taxpayer information is protected. Pasting returns, names or tax numbers into a public AI tool can amount to data leakage (module 4).
• Unfair targeting. A scoring tool trained on past cases can repeat past patterns, including bias against particular regions, sectors or types of taxpayer.
• Over-trust. Under pressure, people tend to accept machine output too readily. This is automation bias and it is a particular danger when a tool produces a score that looks precise.

Principles for this pathway

1. The law decides, people apply it. AI can help explain or organise, but liability, penalties and appeals are decided by authorised officers under the law.
2. Check every legal statement. Any reference to a provision, rate or deadline is checked against the official text before it is used.
3. Taxpayer data stays in approved systems. If you are not sure a tool is approved for confidential information, assume it is not.
4. A flag is not a finding. A tool that marks a case as high risk has given you a reason to look, not a conclusion.
5. Taxpayers can understand and challenge. Decisions should be explainable in plain words and open to redress.

Questions to ask before using any AI tool

• Has my department approved this tool for this kind of work?
• What information will I put into it and where does that information go?
• Who checks the output and against what source?
• If a taxpayer asked how a decision was reached, could we explain it?

The checking AI output module in the operational officers pathway gives a general routine that applies here too.

Try it

Pick a publicly available guidance note or FAQ from a revenue authority's website. Ask an AI assistant to explain one provision it covers and to cite the relevant section. Then compare the answer with the official text. Note anything that was invented, out of date or simplified so far that it became wrong.

Check yourself

1. Name two kinds of revenue task where AI can safely help.
2. Why is out-of-date law a particular risk with AI tools in revenue work?
3. What does "a flag is not a finding" mean in practice?

Answers: (1) Any two of: drafting plain-language material, summarising long documents, organising public information, finding sections of internal manuals (with checking). (2) Tax law changes often and AI tools may describe an earlier position or another country's rules as if current. (3) A tool's high-risk flag is a reason for an officer to look at the case, not a conclusion about the taxpayer.