Artificial intelligence is rapidly becoming part of the tax landscape. VAT is no exception. AI tools can analyse large volumes of transactions, identify unusual patterns, review invoices, compare VAT treatments and assist with researching technical questions. Used properly, these capabilities can make VAT reviews faster and potentially more effective.

However, there is an important distinction between using AI to assist with a VAT review and allowing AI to determine the correct VAT treatment.

VAT is not simply a matter of applying a fixed set of rules to a set of numbers. It frequently involves questions of fact, interpretation, evidence, commercial reality and professional judgement. Those characteristics create significant risks when an AI-generated answer is accepted without appropriate human review.

HMRC’s current guidance recognises this distinction. Its VAT compliance guidance specifically identifies AI as an area requiring appropriate controls and states that tax decisions relying on algorithms should be explainable, supported by an audit trail and monitored for accuracy. HMRC also expects human oversight of generative AI used in tax-related software. 

The attraction of AI in VAT reviews

There are obvious reasons why businesses and advisers are turning to AI.

A VAT review can involve thousands of transactions, multiple VAT codes, invoices, contracts, correspondence and accounting records. AI can process information at a speed that would be impractical for an individual reviewer.

For example, an AI system may be able to:

• identify transactions that appear to have an unusual VAT treatment;
• compare large numbers of invoices against expected VAT coding;
• identify potentially inconsistent treatment of similar transactions;
• summarise contracts and supporting documentation;
• identify transactions requiring further investigation;
• assist with research into VAT legislation, HMRC guidance and case law;
• analyse historical VAT returns and identify recurring anomalies; and
• help prepare questions for management or the finance team.

These are potentially valuable uses of technology. Indeed, HMRC itself uses a VAT Return Analysis Tool to identify VAT returns that may warrant further scrutiny. Importantly, however, HMRC states that the tool does not itself make decisions: the VAT officer remains responsible for deciding whether compliance action should be taken. 

That distinction is critical.

AI can produce a confident answer that is nevertheless wrong

One of the greatest risks in using generative AI for VAT is that an answer can appear authoritative without actually being correct.

AI systems can produce what are commonly described as “hallucinations”: information that appears plausible but is factually incorrect or unsupported. HMRC’s guidance expressly recognises this risk and says users should understand the limitations of the model and the possibility of inaccuracies. 

In a VAT context, the consequences can be significant.

An AI system might:

• cite legislation that does not support the conclusion;
• rely on an outdated version of HMRC guidance;
• confuse UK VAT rules with rules from another jurisdiction;
• misunderstand the factual circumstances of a transaction;
• overlook an exception or special rule;
• fail to distinguish between a zero-rated, exempt and outside-the-scope supply;
• incorrectly determine the place of supply;
• overlook the significance of contractual terms;
• treat two superficially similar transactions as having the same VAT treatment; or
• reach a conclusion without identifying the evidential information that is still required.

The danger is not necessarily that the answer will look obviously wrong.

The more serious problem is that the answer may look entirely reasonable.

That can create a false sense of certainty.

VAT is particularly dependent upon facts and interpretation

VAT legislation contains many rules that cannot safely be applied without understanding the underlying facts.

The VAT treatment of a transaction can depend upon precisely what is being supplied, to whom, where the parties are established, how the arrangements are structured and what the contractual and economic reality of the transaction is.

Questions concerning single or multiple supplies provide a good illustration. HMRC’s own guidance acknowledges that whether an arrangement should be treated in a particular way can depend upon the circumstances and economic reality of the transaction. 

This presents a fundamental limitation for AI.

An AI system can only analyse the information that it is given and the information that its underlying sources enable it to access. If the factual information is incomplete, misunderstood or presented without the relevant commercial context, the resulting VAT analysis may also be flawed.

A technically impressive answer to the wrong question is still the wrong answer.

The problem of interpretation

VAT review is not always about finding a rule and applying it.

There are circumstances in which the legislation, case law and HMRC guidance need to be considered together. There may be competing interpretations. There may be uncertainty over how a particular court decision applies to a different set of facts. HMRC’s published position may need to be distinguished from the underlying legislation and case law.

This is where professional judgement becomes particularly important.

AI can assist by identifying potentially relevant legislation, cases and guidance. It can help a practitioner formulate questions and explore alternative interpretations.

But it should not be assumed that the AI has resolved the legal or technical question simply because it has produced a coherent explanation.

The reviewer needs to ask:

What is the authority for this conclusion?

Does that authority actually say what the AI claims it says?

Is the authority current?

Are there subsequent cases, legislative changes or HMRC updates?

Are the facts sufficiently understood to apply the authority?

Is there another reasonable interpretation?

Those are professional questions, not simply technological ones.

The danger of treating AI as an authority

Perhaps the most important principle is that AI is not itself a source of VAT law.

An AI-generated response should not be treated as equivalent to legislation, binding case law or authoritative HMRC material.

Even where an AI tool is specifically designed for tax purposes, its output requires appropriate controls. HMRC’s guidance expects reliable source data, including legislation, established case law and official HMRC publications, together with testing, monitoring and updating. 

A responsible VAT review should therefore work backwards from the AI conclusion to the underlying authority.

If AI says that a transaction is zero-rated, for example, the question should not simply be whether the answer “sounds right”. The reviewer should establish:

1. What provision is being relied upon?
2. What are the conditions for applying it?
3. Are all those conditions satisfied?
4. What evidence demonstrates that they are satisfied?
5. Is there relevant case law?
6. Is there relevant HMRC guidance?
7. Has the law or guidance changed?
8. Are there facts which could produce a different result?

Only after those questions have been addressed should the VAT treatment be accepted.

The importance of qualified human oversight

The answer is not to reject AI.

The answer is to put the technology in the correct place within the VAT review process.

AI can be the first reviewer, an analytical assistant or an additional pair of eyes. It should not automatically become the final decision-maker.

A suitably qualified VAT professional can provide something that an AI system cannot reliably provide: professional accountability for the conclusion reached.

That professional can challenge the AI output, consider the factual background, assess the quality of the evidence, interpret legislation and case law, recognise uncertainty and determine whether specialist advice is required.

This is consistent with current professional guidance. The CIOT has emphasised that AI should not replace professional judgement and that tax professionals must understand the limitations of AI tools and properly review their outputs. 

A sensible model: AI identifies, humans decide

A useful way to approach AI-assisted VAT review is to divide the process into two stages.

AI identifies and assists.

The qualified reviewer evaluates and decides.

For example, AI might identify 250 transactions that appear to require investigation. A VAT professional can then determine which of those transactions genuinely present a VAT risk, obtain additional information, establish the relevant legislation and guidance, and reach a conclusion.

This approach also makes better economic sense.

There is little value in having a highly qualified VAT professional manually examine every routine transaction if technology can identify the transactions that deserve attention.

Conversely, there is considerable risk in allowing an AI system to make complex VAT determinations without appropriate human intervention simply because doing so appears more efficient.

The objective should therefore be better use of professional time, rather than elimination of professional judgement.

An audit trail is essential

Another important consideration is the ability to explain how a VAT conclusion was reached.

If an AI system identifies a transaction as incorrect, the reviewer should be able to establish why.

What information did the system consider?

What rules or sources did it rely upon?

What prompt or methodology was used?

What conclusion did it reach?

Was the conclusion challenged?

What additional evidence was obtained?

Who reviewed the conclusion?

What final decision was made?

HMRC’s current VAT compliance guidance specifically highlights explainability, decision factors and audit trails when tax decisions rely upon algorithms. 

This is not merely an IT issue. It is an important element of tax governance.

Confidentiality and data protection

VAT reviews can contain commercially sensitive information, including customer details, supplier information, contracts, pricing, financial data and potentially personal data.

Businesses therefore need to understand what happens to information entered into an AI system.

Before using an AI tool for VAT review, consideration should be given to:

• whether confidential information is permitted to be uploaded;
• how the information is stored;
• whether it is used for model training;
• who can access it;
• whether data is transferred outside the relevant jurisdiction;
• whether appropriate contractual and data protection safeguards exist; and
• whether the organisation’s internal policies permit its use.

HMRC’s guidance specifically identifies data security, privacy by design and UK GDPR compliance as important considerations for AI-enhanced tax software. 

Guidance should not be replaced by technology

There is an understandable temptation to ask AI a VAT question and regard the resulting response as a substitute for professional advice.

That is particularly risky where the issue is material, unusual or contentious.

The more complex the VAT question, the more important it becomes to involve someone with the appropriate technical knowledge and experience.

A qualified VAT professional can determine not only what the answer might be, but also how confident the business should be in that answer.

That distinction is often overlooked.

Tax advice is not simply the production of an answer. It is also the assessment of uncertainty, risk, evidence and consequences.

AI should be treated as a risk tool, not a replacement for expertise

The most effective approach to AI in VAT is therefore neither to embrace it uncritically nor to reject it altogether.

Used properly, AI can make VAT reviews more efficient. It can identify patterns that a human reviewer might miss, reduce the time spent on repetitive tasks and help direct professional attention towards areas of greater risk.

But the technology has limitations.

It can misunderstand facts. It can produce inaccurate information. It can rely upon inappropriate or outdated material. It can present an uncertain conclusion with excessive confidence. And it cannot replace the professional judgement required to interpret complex VAT rules in their factual and commercial context.

The current direction of travel from HMRC and the tax profession is therefore significant: AI should support human judgement, not remove it. HMRC’s own AI guidance states that systems should be designed with human oversight and control and that complex or nuanced tax matters should be flagged for further investigation or advice from a qualified tax professional. 

The future of VAT review is unlikely to be “AI versus the VAT professional”.

It is much more likely to be AI working alongside the VAT professional.

The businesses that gain the greatest benefit will be those that understand the difference.

AI can find the anomaly.

AI can suggest the answer.

AI can point towards the relevant guidance.

But where the VAT treatment matters, someone suitably qualified still needs to ask the most important question:

“Is this actually correct, and can we demonstrate why?”

That final responsibility cannot safely be outsourced to a machine.