Finance · Process breakdown

Can AI prepare a VAT return?

Step by step: which parts of preparing a VAT return AI can do, which are plain automation, where a person must review and how to run a pilot.

7 stepsTypical mix: Automation firstIllustrative analysisUpdated

Partly. VAT return software calculates the figures with fixed rules, AI can read and code invoices and flag odd items, and a person has to review the exceptions and file the return. Think of it as a chain: someone sets the VAT rules once, AI turns documents into coded transactions, accounting software adds them up, and a finance owner approves what goes to the tax authority. VAT is used across the EU, the UK and many other countries (the US uses sales tax instead), and the details differ by country, so treat this as a map of the work rather than tax advice.

What each step needs

01 Keep human

Set the VAT rules once: rates, schemes and treatments

Before any return is prepared, someone has to decide how your business is treated for VAT: which rates apply to what you sell, whether you use a special scheme, which costs you can reclaim in full or only in part, and how you handle sales to other countries. These are judgment calls with legal weight, and AI should not make them. It can summarize official guidance and draft a plain-language policy, but a finance owner or adviser decides and writes the rules down. Every later step is only as good as this one.

Needed at setup and whenever the business changes: a new product line, a new country, a new scheme threshold or a change in the law. Review at least once a year.
02 AI candidate

Read invoices and receipts and suggest a VAT code

This is where AI helps most. Supplier invoices and receipts arrive as PDFs, photos and emails with different layouts. A model can read the supplier, date, net amount, VAT amount and rate, and suggest the VAT code and ledger account, drawing on how similar items were coded before. It saves the typing and the lookup. It does not know your intent, so it will sometimes code a mixed-use item or an unusual supplier wrongly, and it should show its confidence so a person can check the doubtful ones first.

Works when documents are legible and past coding is consistent. Weak for handwritten receipts, foreign-language invoices with unusual layouts and one-off purchases.
03 Standard automation

Apply the fixed VAT rules to coded transactions

Once a transaction has a VAT code, the rest is arithmetic. Standard and reduced rates, zero-rated and exempt sales, and the split between output VAT you owe and input VAT you can reclaim follow fixed rules that accounting software applies the same way every time. This is ordinary automation, not AI, and it should stay that way: a rule gives the same answer for the same input, and you can audit it. Do not ask a model to calculate totals that a ledger can compute exactly.

Needs an accounting system with VAT codes set up correctly and a closed period, so the figures do not move while you prepare the return.
04 Human review

Review exceptions: reverse charge, cross-border and mixed use

Some items do not fit the rules. Reverse charge purchases, sales to customers in other countries, goods that are partly private or partly exempt, bad debt relief, credit notes across periods and corrections to earlier returns all need someone who understands the facts. AI can flag the transactions that look unusual and pull the supporting documents next to each other, which speeds up the review. The decision on treatment stays with a person who can explain it to the tax authority if asked.

Needed whenever a transaction is flagged, is above a value you set, or involves a new supplier, customer or country.
05 AI candidate

Reconcile the VAT figures against the ledger and prior periods

Before filing, the VAT liability should agree with the ledger and look sensible against earlier periods. AI can compare this period with previous ones, point to lines that moved a lot, spot duplicates, missing invoices and rates that differ from the usual, and write a short note explaining the differences for the reviewer. The check itself is only a pointer, not proof. Totals must still tie to the ledger, and the person reviewing decides whether a swing is real or a coding error.

Works best with at least a few clean periods to compare against. Less useful for a new business or a period with an unusual one-off sale.
06 Human review

Approve the return and file it with the tax authority

A person approves the final figures and files the return, or confirms the filing sent by compatible software. The return is a statement made to the tax authority, and the business is responsible for it even if software prepared it. Filing deadlines are firm and late or wrong returns can bring penalties, so the approver should see the reconciliation, the flagged exceptions and the final boxes together. AI can prepare the summary, but it should not submit on its own authority.

Required for every return. The approver needs authorization with the tax authority and access to the supporting records.
07 Standard automation

Schedule the payment and archive the evidence

After filing, the payment or refund request is scheduled by the due date, the return is stored with its reconciliation, and the invoices and receipts behind it are kept for the required period. This is a fixed routine that rules handle well: a reminder before the deadline, a payment instruction for approval, and an archive folder per period. A model adds little here. A missed payment date is costly, so make the reminder impossible to ignore and name an owner.

Needs a calendar of deadlines, an approval step on payments and an archive that meets your local record-keeping rules.

A sensible first experiment

Pick one VAT period and run it in parallel. Let AI code the supplier invoices and flag exceptions, let your accounting software calculate the return, and have a person review everything flagged and prepare the return as usual. Compare the two versions line by line. Measure the minutes spent on coding and review, the share of AI codes you accepted without change, the number of errors found only at review and the number of adjustments needed after the first draft. Only move to the next period with AI coding if accuracy holds and review time falls.

The trap to avoid

The biggest pitfall is trusting AI-coded transactions without a review. A wrong VAT code looks small on one invoice but repeats across a quarter, and the business, not the software, answers for the return. Keep a person on every flagged item, sample the unflagged ones, and never let a model file a return or choose a VAT treatment by itself.

Questions teams ask

Can AI file a VAT return on its own?

It should not. Software can prepare the figures and some systems submit directly to the tax authority, but a person should approve the return first. The business is responsible for the content, and a wrong or late return can lead to penalties. Use AI to prepare and check, and keep the final approval and submission with someone accountable.

Is VAT return software the same as AI?

No. VAT return software applies fixed rules to coded transactions and produces the boxes or fields the tax authority expects. AI is a separate layer that can read documents, suggest codes and explain differences. Many products bundle both, but you can use VAT return software without AI, and the calculation itself should stay rule-based.

What part of a VAT return is safest to give to AI?

Reading invoices and suggesting VAT codes is the safest and most useful part, as long as a person reviews the doubtful ones. Comparing this period with earlier ones to spot duplicates and swings is also a good fit. Choosing the VAT treatment of unusual transactions is not, and nor is filing.

Does the US have VAT returns?

Not at federal level. The US uses sales and use tax, which is filed per state. This guide applies to businesses that deal with VAT, for example when selling into the EU or the UK, or operating there. The same split between rules, AI and human review also applies to sales tax returns.

Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.

Now, what about your process?

Get an assessment based on your own steps, systems, and constraints.

Audit my process See an example report ↗