VAT reconciliation with AI — AI as a second pair of eyes

VAT reconciliation with AI means applying AI to signals, cross-checks, and intra-community-supplies control in the VAT cycle — not to file the return itself, but to check the reconciliation between balance sheet and filed return as a never-tired second pair of eyes. For finance teams this lowers the chance of a discrepancy between VAT balance and return, without taking away the team's own responsibility for accuracy.

The VAT return is, for most SME and scale-up controllers, one of the tightest deadlines of the month or quarter. The tax authority is patient about very few things, and a discrepancy between the VAT balance on the balance sheet and the filed return is one of them. At the same time, VAT reconciliation is precisely the kind of work that can be described as "discipline, speed, and meticulousness" — which makes it a natural candidate for AI support. Not to let AI do the return, but to let it check the reconciliation as a second pair of eyes that never gets tired and always sees everything at once.

This piece describes what AI realistically adds today in a VAT cycle, where the line sits, and how a working workflow looks.

Where the time and risk sit

A typical VAT cycle at a 50-FTE trading company with EU suppliers and EU customers looks roughly like this:

• Add up the bookkeeping balance of VAT per balance-sheet account (owed, recoverable, intracommunity).
• Generate the return printout from the accounting system.
• Determine the difference between the two — there is always a difference from timing, corrections, or wrong postings.
• Explain the difference item by item.
• Book correction entries.
• Generate the intracommunity (ICP) return separately and check it against the main return.
• File the return.

The time isn't in filing but in steps 3, 4, and 6: the work of hunting for differences. The risks sit in subtly-wrong rate applications, in EU trade booked under the wrong rubric, in personal use not corrected, and in bad debts where VAT could have been recovered already.

Recurring mistakes seen in practice:

• VAT applied at 21% on a service that falls under 9% (or vice versa).
• ICP deliveries booked as 0%-export but missing from the ICP statement.
• Input VAT on personal use (car, telephony) not corrected.
• VAT on bad debts not recovered, or recovered too early.
• Correction postings from prior periods not picked up in the return.
• VAT reverse-charged where it shouldn't be, or not reverse-charged where it should be.

These are exactly the errors an AI can signal well — not because it "understands" them, but because it can hold patterns side by side that are too many for a human at once.

What AI does and doesn't do in the VAT cycle

What AI does well:

• Cross-check balance vs return. Compare the balances on VAT accounts with the return printout and pinpoint exactly where the difference sits, at invoice or posting level.
• Pattern signaling. For example: "customer X has historically always carried 21% VAT, this invoice carries 9% — intended or wrong?" Or: "supplier Y has always delivered in NL, this invoice from Germany is booked as domestic VAT, possibly ICP."
• ICP check. Walk through booked 0%-EU deliveries, validate the customer's VAT number (via VIES), and check whether the booking lands in the ICP statement.
• Prepare corrections. Draft correction postings on identified differences for controller approval.
• Reconcile with the prior period's VAT return. For example: did corrections from Q1 that should have been processed in Q2 actually get picked up?

What AI doesn't do and shouldn't want to:

• End responsibility for the return. The return is signed (digitally) by a person, with tax-legal force. An AI doesn't sign.
• Materiality judgment. Whether a difference of €127 is worth a correction or can stay as noise is a controller judgment — heavily dependent on pattern and history.
• Tax interpretation on borderline cases. A service in a grey zone between rates or regimes (think conference organization with international speakers) demands tax judgment. AI can describe options, not choose.
• Assess company-specific arrangements. A supplier with a non-standard invoicing arrangement (advance invoicing, bundling), or a customer with a specific rate application based on a ruling — that's company knowledge, not pattern recognition.

Anatomy of a workable workflow

An AI-assisted VAT reconciliation can look step by step like this.

Step 1 — Data pull and balance determination

AI: Pulls balance positions on all VAT accounts, fetches the return printout from the accounting system, and presents a reconciliation table with the difference per rubric. Human: No intervention needed — this is read-only work. HITL: N/A.

Step 2 — Difference analysis and signaling

AI: Walks each unexplained variance, traces it back to one or more postings, and classifies: timing, rate question, missing correction, booked-to-wrong-rubric, or "unknown, controller judgment required." Human: Controller reviews each signal in a queue: timing differences can usually stay, rate questions get verified, corrections are approved or adjusted. HITL: On every signal that leads to a posting.

Step 3 — ICP check

AI: Walks every 0%-EU delivery, validates the customer's VAT number via VIES, checks whether the booking is in the ICP statement, and flags discrepancies. Human: Controller verifies per flag whether it's an error (no valid VAT number, ICP statement needs supplementing) or an explainable exception. HITL: On every correction of a 0%-rate or amendment to the ICP statement.

Step 4 — Draft corrections

AI: Drafts for every approved signal a journal entry, with amount, rubric, contra account, and explanation. Human: Controller (or under four-eyes: second controller) approves per posting. HITL: Standard, because these are formal VAT corrections.

Step 5 — Return preparation and filing

AI: Generates a pre-filing overview: what is going into the return, with source and amount per rubric. Human: Controller or finance manager reads the overview, does the final check, and files the return themselves via the official channel. HITL: Before filing. The return itself is submitted by a human; no agent clicks "submit."

Audit grade — what to keep for review

The tax authority and your external auditor want to establish afterwards that the return holds up. For an AI-assisted cycle that means:

An audit trail per signal and per correction. Which balance was input, which postings underpinned it, which AI classification was given, which controller approved at which moment, which correction was booked. The same discipline as manual bookkeeping — only more automated from now on.

An evidence trail of the return itself. The eventually filed return plus all documents and corrections that led to it. Not only in the accounting system, but in a separate folder or system where the return version and the underlying AI actions are visible together.

A periodic quality check on the AI signaling itself. For example: per quarter pick 20 random signals from the queue and manually check whether the AI's classification holds. If the error rate climbs, you intervene before it lands in a return. This meta-check is itself audit evidence: you've shown you monitor your AI tool's behavior.

What it delivers

Teams that have been working with this for a few months see time savings of typically 40-60% on the time spent on VAT reconciliation — mostly on steps 2 and 3 (difference analysis and ICP check). The quality gain is just as important: errors currently only caught in a supplementary return get signaled earlier. A supplementary return costs not just time and penalty risk but also credibility with the tax authority — and credibility, once damaged, is slow to recover.

Limits — when AI doesn't help here

A few situations where AI-assisted VAT reconciliation contributes less:

• Strongly shifting or new business models. A SaaS company that has just started international licensing or a new product line with non-standard VAT treatment — there the AI lacks the patterns needed for good signaling. Do it yourself first, then automate.
• Unstable booking discipline. If postings get restructured weekly or the chart of accounts changes continuously, the AI can't learn reliable patterns. First put booking discipline in order, then deploy AI.
• One-person finance department without backup. AI raises the speed at which errors can be made. Without a second pair of eyes (no four-eyes, no external auditor watching along), the dependency on AI correctness becomes greater than is sensible for VAT.