AI-assisted VAT and GST return review

What AI-assisted VAT and GST return review means

AI-assisted VAT and GST return review is the use of a language model or a rules engine to check a VAT or GST return against the ledger and source documents before a person files it, testing tax rates, reverse charge treatment and reconciliation to the general ledger for anomalies. It matters because the software finds mismatches; a qualified preparer still investigates, corrects and signs what actually gets filed.

The tool never files anything. It reads a period's sales, purchases and expense records, recomputes the tax it expects to see, compares that with what was actually charged or claimed, and hands a reviewer a short list of exceptions instead of a full ledger to eyeball. This guide covers three regimes with real differences in mechanics: European Union VAT and its reverse charge rules, the United Kingdom's Making Tax Digital for VAT, and Canadian GST and HST across provinces with different rates, then a worked example, a reviewer checklist and the data-protection practices that keep tax data safe.

Where AI fits in the review cycle

A VAT or GST review has the same shape everywhere: extract, recompute, compare, explain. AI is useful at each step but decides nothing on its own.

• Extraction: reading invoices, receipts and bank lines for the tax rate, tax amount, country or province, and whether the line is a supply of goods or services
• Recomputation: applying the rate that should have applied, based on the customer's location, the reverse charge rules and the type of supply
• Comparison: matching the return's box totals to the general ledger's VAT or GST accounts and to the underlying transaction list
• Anomaly flagging: rate mismatches, missing reverse charge entries, duplicate credit notes, and transactions dated outside the return period
• Drafting: a short, plain-language explanation of each flagged item for the reviewer to accept, correct or dismiss

European Union: reverse charge checks

Under the EU VAT Directive, the general rule for business-to-business services is that the customer, not the supplier, accounts for VAT: article 196 of Council Directive 2006/112/EC makes the recipient liable when the supplier is not established in the customer's member state. An AI review reads each purchase invoice from a supplier in another member state and checks that no local VAT was charged and that a matching self-assessed entry exists on both the input and output side, since a correctly applied reverse charge nets to zero.

Beyond the general business-to-business rule, member states may apply a domestic reverse charge to specific high-fraud sectors, such as mobile phones, integrated circuits, gas, electricity and certain metals, under articles 199 and 199a of the Directive. As of September 2026, the European Commission confirms that this optional mechanism, together with the Quick Reaction Mechanism for sudden fraud, runs until 31 December 2026, so a review for a business in an affected sector should flag any invoice in one of those categories that was not reverse-charged.

• No VAT shown on the supplier's invoice, VAT self-assessed by the customer: check the entry is present on both the input and output side
• Sector on the domestic reverse charge list: check the supply is genuinely one of the listed categories, not just the supplier's usual trade
• Same rate applied to the notional output and input lines: a mismatch usually means one side was keyed incorrectly

United Kingdom: Making Tax Digital for VAT

Making Tax Digital for VAT has applied to every VAT-registered business in the UK since 1 April 2022, regardless of turnover, under HMRC's VAT Notice 700/22. It requires records to be kept digitally in functional compatible software and a digital link between every piece of software that touches the return's figures, meaning no manual retyping once a number is part of the digital VAT account; copying a total by hand from one spreadsheet into another breaks that chain.

An AI review adds value by testing what the digital-links rule assumes is already true: that the return's boxes tie to the underlying sales and purchase ledgers without a manual step in between. It checks that the standard, reduced and zero rates, 20%, 5% and 0% respectively, were applied to the right line items, flags a rate that looks wrong for the product description, and lists any manual adjustment that was not carried through the same digital chain as the rest of the return.

• Every VAT-registered business must file this way; there is no turnover exemption left
• A rate applied to a line does not match past invoices for the same product or the same customer
• A manual journal adjusts the VAT account outside the digital links used to build the return
• An exempt or zero-rated supply is coded at the standard rate, or the reverse

Canada: GST and HST across provinces

Canada charges a federal Goods and Services Tax of 5% in provinces without a harmonised tax, and a combined Harmonized Sales Tax in the provinces that harmonised their own tax with the GST: 13% in Ontario, 14% in Nova Scotia since the province cut its portion on 1 April 2025, and 15% in New Brunswick, Newfoundland and Labrador, and Prince Edward Island. The rate that applies to a sale depends on the place of supply, normally the customer's province, not the seller's.

An AI review recomputes the rate that each invoice's ship-to or place-of-supply data implies and compares it with the rate actually charged. A stale customer record, a rate table that was not updated after a provincial change, or a service coded to the wrong province are the usual causes of a mismatch. It also checks that input tax credits claimed on purchases have supporting invoices carrying the supplier's registration details, since a credit without a valid invoice is a common rejection at filing.

Worked example: a rate change the ledger missed

A wholesaler's ledger for the quarter shows net sales of CAD 220,000 to Ontario customers, CAD 90,000 to Nova Scotia customers and CAD 140,000 to Alberta customers, taxed at 13%, 14% and 5%. Ontario tax is 220,000 times 13%, or CAD 28,600, and Alberta tax is 140,000 times 5%, or CAD 7,000. Correctly taxed, Nova Scotia should be 90,000 times 14%, or CAD 12,600, giving total tax due of 28,600 plus 12,600 plus 7,000, which is CAD 48,200.

The AI review recomputes each invoice from its recorded ship-to province and finds that fifteen invoices to Nova Scotia customers, totalling CAD 20,000, were charged at the old 15% rate instead of the 14% rate that has applied since 1 April 2025: 20,000 times 15% is CAD 3,000, so the ledger actually holds 70,000 times 14%, CAD 9,800, plus that CAD 3,000, or CAD 12,800 for Nova Scotia, CAD 200 more than the CAD 12,600 that was due. Total recorded tax across the three provinces is 28,600 plus 12,800 plus 7,000, which is CAD 48,400, CAD 200 above the correct CAD 48,200.

The reviewer confirms the fifteen ship-to addresses really are in Nova Scotia and the invoice dates are after the rate change, then approves a correction: credit notes for CAD 200 in total, posted Dr HST payable 200 / Cr Trade receivables 200, which reduces the tax collected without touching revenue. The reviewer rejects the AI's first suggestion to leave the extra 200 as immaterial, because it is customers' money, not the business's, and the return must show only the tax the business was entitled to collect.

What the reviewer checks before filing

An AI-flagged exception list is a starting point, not a finished review. Before anything is corrected, filed or dismissed, a competent reviewer works through a short list of checks so a plausible-looking pattern is not accepted just because software found it.

• Re-derive at least one flagged item by hand from the source document, not from the AI's summary of it
• Confirm the population is complete: does the sum of every line the tool worked from equal the trial balance for the VAT or GST accounts
• Check the direction of any correction: overcollected tax belongs to the customer or the tax authority, not to the business
• Look for exceptions the tool did not flag; a rules engine or model only checks what it was told to check
• Record who reviewed what, and what was accepted, corrected or overruled, before the return is filed

Reconciling the return to the ledger

A VAT or GST return that no AI tool has touched still has to reconcile to the ledger, and that reconciliation is the real safeguard, not the AI. The return's boxes should equal the balances on the VAT or GST control accounts for the period, adjusted only for known timing differences such as invoices raised after the period end but dated within it. A review that cannot show its numbers tying back to those control accounts has not actually reviewed the return; it has produced an opinion about it.

This is also where most of the value shows up before a tax authority ever asks a question. A business that keeps this reconciliation as a matter of routine, AI-assisted or not, answers an enquiry in minutes instead of weeks.

Data protection when reviewing tax data

Sales and purchase ledgers carry customer names, addresses and sometimes personal data, so sending them to an AI tool is a personal-data processing activity under the GDPR that needs a lawful basis, a processor agreement with the provider, and a valid basis for any transfer outside the EEA; this is general information, not legal advice.

Treat every uploaded invoice as untrusted input, not just as data: a PDF can carry hidden text instructing an assistant to change its answer, a form of prompt injection, so a review tool should extract figures and never act on instructions found inside a document. Give the tool read access to ledger exports only, never write access to the return or the general ledger, and log what was reviewed, by which tool, and what a person did with the result.

• As of September 2026, Anthropic states that by default it does not use inputs or outputs from its commercial products, including the API, to train its models, unless the customer chooses to share them
• Limit the data sent to what the check needs: tax rate, amount, date, customer country or province, not full customer files
• Keep filing itself as a separate, authorised step a person takes, never an automated action of the review tool

Doing this in Skyline Nexus ERP

Skyline Nexus ERP already gives a reviewer the reconciliation that any VAT or GST check depends on. Tax rates and tax groups are set once in Settings, including a cascading option for excise-type taxes and a field for a ZATCA exemption code, and each sale computes VAT per line the same way the VAT return does, so the prefilled VAT Return form and the VAT Analysis report in Fiscal Authority read from the same VAT input and output GL accounts.

That shared basis is what any automated or AI-assisted review checks against: if the return and the ledger are built from the same figures, a mismatch between them means one of the two was changed after the fact, which is exactly what a reviewer wants to be told. Skyline Nexus ERP does not run an AI check over a VAT or GST return itself; that kind of AI-assisted anomaly review is on the Skyline Nexus ERP roadmap and being rolled out, and the in-app assistant already answers how-do-I questions about running the VAT Return and VAT Analysis reports, citing the help page it used.