AI in Tax: Compliance, Indirect Tax and Research

By Jane Ha · Updated Oct 11, 2026

AI in tax is the use of machine learning and language models to prepare compliance data, determine indirect tax on transactions, and search and summarize tax law, while a qualified professional reviews the output and remains responsible for every position taken and every return filed. Tax is a good fit for AI because the work is rule-heavy, document-heavy and repetitive. It's also a risky fit, because a confident wrong answer becomes a penalty, and "the software said so" isn't a defense.

This guide is for tax directors, CFOs, controllers and finance managers deciding where AI belongs in the tax function. It's general information, not tax advice. For the wider picture, start with AI in finance and accounting, and for the language-model side of the work, read generative AI in accounting.

Key Facts

• In its 2018 South Dakota v. Wayfair decision, the U.S. Supreme Court overruled the physical presence rule for sales tax. South Dakota's law, cited by the Court, applies to sellers delivering over $100,000 of goods or services into the state or making 200 or more separate transactions there (Cornell LII, Wayfair opinion).
• Under Treasury Circular 230, a practitioner must exercise due diligence in preparing and filing returns and in checking statements made to clients about IRS matters (31 CFR 10.22).
• A practitioner may rely in good faith on client information without verification, but may not ignore the implications of information that looks incorrect, inconsistent or incomplete (31 CFR 10.34).
• A preprint study of large language models answering tax law questions found accuracy improved with each model generation and rose further with few-shot prompting and the correct legal texts, yet the models were "not yet at expert tax lawyer levels" (Nay et al., arXiv).
• The EU's VAT in the Digital Age package introduces digital reporting for cross-border business-to-business supplies from 1 July 2030, built on e-invoicing, and the Commission estimates it could cut VAT fraud by up to EUR 11 billion a year (European Commission).

What AI Does in Tax

Tax work breaks into three streams, and AI helps differently in each.

Compliance is the recurring production of returns, filings and supporting schedules. AI reads source documents, maps ledger accounts to return lines, reconciles differences between book and tax numbers, and flags items that don't tie out.

Indirect tax (VAT, GST and sales and use tax) is a transaction-level problem. Every invoice needs the right tax code, rate, place of supply and exemption treatment. AI helps classify products and services, extract data from invoices and certificates, and spot lines where the tax charged looks off.

Research and advisory is reading law, rulings and guidance and applying them to facts. A language model can find relevant passages, summarize them and draft a memo outline. It can also invent a citation, which is why the research stream needs the tightest controls.

What ties the three together: AI speeds up preparation and review, and it shouldn't be the final authority on any position.

Where AI Helps by Tax Function

Compliance: Preparing Returns Faster

The biggest compliance gain comes from the front of the process, getting data into shape. Preparers spend much of their time collecting trial balances, matching accounts to return lines and chasing explanations for variances. AI can read the trial balance, propose the mapping, carry forward last year's mapping with changes highlighted, and list the accounts where the numbers moved enough to deserve a look.

An illustrative case: a mid-sized company with five legal entities closes the year and needs entity-level returns. A mapping model proposes how each account feeds each return line, using last year's final mapping as the starting point. The preparer reviews only the mappings where confidence is low or the balance changed sharply, rather than every account. The model doesn't decide the tax treatment of a new item. It flags that the item is new and asks.

The same logic applies to month-end close: if the books are clean, tax preparation is faster. If they aren't, AI just finds the mess sooner, which is still useful.

Indirect Tax: Determination at Transaction Scale

Indirect tax is where volume makes AI attractive. A retailer with thousands of products selling into many jurisdictions can't have a person look at every line. The pressure grew after Wayfair, which let states require collection based on economic activity alone. A seller with no physical presence can still have registration and collection duties once it crosses a state's thresholds, so the question became "where do we have to collect?" instead of "where are we located?"

AI helps in four ways:

1. Product and service classification. A model reads a description and proposes a tax category. Ambiguous items (software with services bundled in, for instance) go to a person.
2. Document extraction. Vision and language models read invoices, exemption certificates and registration numbers, so the data isn't retyped. The mechanics resemble AI in accounts payable, where invoice capture is already common.
3. Anomaly detection. A model compares tax charged against history and peers and flags the invoice that carried 0% when similar ones carried a rate.
4. Threshold tracking. Sales are summed by jurisdiction and compared with economic nexus thresholds, so someone reviews registration before the exposure grows. Thresholds differ by jurisdiction and change, so pull them from the authority's own page rather than from a model's memory.

Outside the U.S., the direction of travel is toward structured data. The EU's digital reporting rules for cross-border supplies start on 1 July 2030, and they rest on e-invoicing, with Member States already allowed to mandate e-invoicing domestically. Companies that already extract clean, structured invoice data will find that work reusable.

Tax Research: Fast, and Easy to Get Wrong

Research is where language models feel most impressive and fail most visibly. A model can summarize a code section in seconds. It can also cite a ruling that doesn't exist, quote a rule that's been superseded, or apply a general principle to a fact pattern where an exception controls.

The academic evidence points the same way. A study testing language models on tax law questions found that results improved with each new model generation, and improved further when the model was given the correct legal texts and shown worked examples. Even then the authors concluded the models were not yet at expert tax lawyer levels. The practical lesson: give the model the source (retrieval from a library you control), and treat its output as a lead, not an authority.

A safe research workflow looks like this:

• The model searches your approved library of statutes, regulations, rulings and prior memos.
• It returns passages with exact citations and links.
• A person opens each cited passage and confirms it says what the summary claims, and that it's still in force.
• The memo carries the human's name, not the tool's.

That's a retrieval-based assistant pattern, and it's the difference between a research aid and a liability.

Professional Responsibility: Who Owns the Answer

The tools change. The duty doesn't. In the U.S., Treasury Circular 230 requires practitioners to exercise due diligence in preparing and filing returns, and in checking the correctness of statements they make to clients. A practitioner who relies on someone else's work is presumed to have been diligent only if they used reasonable care in engaging, supervising and evaluating that person. Few tax authorities would treat an unreviewed AI output as better than an unreviewed junior's.

Another rule matters here. A practitioner can generally rely in good faith on client information without verification, but can't ignore information that looks wrong, inconsistent or incomplete. An AI system that flags an inconsistency creates a record that someone saw it. If the team then ignores the flag, that's harder to defend than never having had the flag at all.

Three practical consequences:

• Review isn't optional. A named professional approves every filing and every position memo, regardless of who or what drafted it.
• Confidentiality. Client and taxpayer data shouldn't go into unapproved public chatbots. Use tools with contractual data protections and check your engagement terms and local privacy law.
• Disclosure and records. Keep a record of which tool was used on what, what it produced and what the reviewer changed. It helps when a tax authority asks how a number was reached. Teams that already run internal controls for AI can extend the same log.

Check your own jurisdiction's rules and your professional body's guidance. This article doesn't replace them.

How to Roll Out AI in Tax, Step by Step

1. Pick one stream. Start with indirect tax data extraction or return-line mapping, not with positions on uncertain law.
2. Fix the data first. Clean product master data, customer registration numbers and chart of accounts mapping. Finance data readiness lists what to check.
3. Baseline your error rate. Count how often returns are amended, how often notices arrive and how many tax code corrections happen after posting. Without a baseline you can't prove improvement.
4. Run in parallel. For two or three cycles, let the tool prepare alongside your normal process and compare line by line.
5. Set confidence thresholds. Auto-accept only high-confidence suggestions in low-risk categories. Route everything else to a person.
6. Define review levels. Who reviews what, how deep, and what evidence they sign. Have internal audit read the design; AI in internal audit covers how audit teams evaluate these tools.
7. Lock down research. Point the research assistant at an approved library, require citations and require a human to open each one.
8. Review quarterly. Rates, thresholds and rules change, and so does model behavior. Retest on a sample of known answers.

Common Pitfalls

• Trusting rates and thresholds from memory. A model's idea of a rate or threshold may be years old. Pull current figures from the authority.
• Fabricated citations. Language models can produce references that look right and don't exist. Open every one.
• Treating classification as settled. One product code applied to thousands of lines multiplies a single mistake thousands of times. Sample and review.
• No override trail. If staff can change a tax code without a record, you can't learn from errors or defend decisions.
• Automating a bad process. AI over messy product data and inconsistent customer records just produces wrong answers faster. A model trained on bad master data learns the mess, so clean the data before you automate.
• Ignoring multi-jurisdiction drift. Mandates such as e-invoicing roll out country by country and on different dates. A setup that works in one market may fail in the next.
• Skipping the human sign-off. Efficiency gains vanish the first time an unreviewed output reaches a tax authority.

What to Measure

• Accuracy: tax code correction rate after posting, share of suggestions accepted unchanged, and error rate found in sample review.
• Compliance outcomes: notices received, amended returns and late filings.
• Speed: days from period end to filing-ready workpapers and hours per return.
• Control health: overrides logged, exceptions closed on time and research memos with every citation verified.

Where to Go Next

If you do one thing this quarter, baseline your tax code correction rate and test an extraction or mapping tool in parallel with your current process. Then widen to threshold monitoring. Keep research behind an approved source library with a human reading every citation. Pair this guide with AI in finance for governance and generative AI in accounting for language-model basics.