AI Automation ·
Which records should Korean finance teams reconcile before using AI on VAT prefills?
In November EY published a survey of 1,600 finance and tax leaders across 30 jurisdictions. It reported that 64% named insufficient AI-ready data as the main barrier to advancing AI initiatives [1].
Problem definition and research question
A Korean retailer needs to resolve discrepancies between ledger card sales and Hometax prefills before filing VAT. If an agent silently defaults to one number, it can hide reversals or timing differences. Where should the review boundary be when two sources disagree?
Evidence and causal analysis
EY measured respondents’ assessments, rather than verified error rates or Korean performance [1]. The Korean tax authority asks filers to reconcile prefilled invoices and card amounts with actual transactions [2]. Treating one feed as authoritative could confuse reversal and authorization dates. That causal mechanism is a pilot hypothesis that needs transaction evidence.
| Dimension | International evidence | Korean evidence and application |
|---|---|---|
| Evidence | Self-reports from 1,600 respondents across 30 jurisdictions [1] | NTS guidance to check prefills against actual transactions [2] |
| Local design | No causal estimate of AI accuracy gains | Pilot explanations that link original transactions, timestamps and reviewers |
Comparison of alternatives
Option A matches identifiers and amounts using rules and sends exceptions to reviewers. It is traceable, but split and aggregated transactions require more staff effort. Option B adds proposed explanations for the same exceptions, such as reversals or reporting dates, and a reviewer confirms them from associated evidence. It could save sorting time while producing misleading classifications. The tax lead retains responsibility for evidence and filing.
Implementation and evaluation
Deidentify closed-quarter samples with reversal and correction cases. Align snapshots, then give both options identical exceptions. Measure review time, correct explanations, missing evidence and ungrounded proposed changes in allocation periods. Disable classification that produces unsupported changes to tax or filing data and revisit reliable identifiers before restarting.
Conclusion and open questions
AI can be tested as a source of candidate explanations. What identifiers and timestamps would distinguish an incorrect underlying transaction from a genuine timing difference between aggregates?