Initial positive trials on low-risk VAT refunds facilitate full automation
Italy’s Agenzia delle Entrate tax authority is deploying artificial intelligence to speed up VAT refund processing and improve fraud detection, while keeping final decision-making under human supervision.
This follows the recent one Introducing Italy’s AI VAT chatbot. This follows the others Tax authorities deploy AI.
Tax authorities report to Congress
At a Congressional hearing in October, officials explained that AI is being incorporated into the agency’s risk analysis and compliance systems to improve efficiency and accuracy. This technology is applicable only in the following locations: Preliminary evaluation This is the stage where refund claims are reviewed for discrepancies and does not affect audit or enforcement decisions.
The new system will enable the rapid collection, organization and evaluation of large amounts of taxpayer and transaction data, allowing auditors to more effectively identify potential anomalies. The goal is not to replace human staff, but to enhance their ability to focus on complex or high-risk cases.
Machine learning using various data sources
To reduce time to refund, VAT-related data flows into the Revenue’s central tax data warehouse, including:
-
Electronic bill ( interscambio system),
-
cross-border transaction report,
-
Customs import/export declaration, and
-
Banking and withholding tax data.
First, AI and machine learning models Normalization and cross-references We clean these datasets for duplicates, tag trading partners, and adjust the data structure for further analysis.
VAT anomaly and risk scoring (pre-assessment stage)
For refund requests, automatically graded Check risks before human review.
The algorithm uses a combination of:
-
pattern recognition (e.g. sudden spike in refund claims on sales, chain transaction patterns, mismatched VAT ID activity),
-
network analysis (to flag known high-risk traders or connections to carousel chains),
-
historical profiling (based on historical reimbursement reliability, sector averages, and bill density);
-
text AI To categorize invoice descriptions or identify unusual tax code usage.
Rather than determining outcomes, the model ranks claims by: Possibility of error or frauda feed dashboard used by auditors.
Reduced VAT collection processing time
By 2026, low-risk VAT refunds will be processed automatically, reducing waiting times and providing faster liquidity for compliant businesses. Time saved through automation can be redirected to deeper investigation. Early data is already showing progress, with the average refund period reduced to 67 days by September 2025, exceeding Treasury’s 2025 target of 70 days.
Officials emphasized that Italy’s approach to AI remains cautious. Unlike some other EU countries that use automated scraping of online and social media data, the Italian Revenue Agency avoids such practices in order to adhere to strict privacy and data protection standards. These safeguards need to be maintained to align with Europe’s broader digital tax framework in the future, they said.
