How does the tourist tax refund AI platform handle PCI-DSS compliance?

Card data is extracted once using PCI-compliant AI on Azure Custom Vision and AWS Bedrock, then immediately vaulted. All downstream systems operate on tokens only, reducing PCI scope by 90%.

What OCR accuracy does the Vertekx tourist tax refund system achieve?

The platform achieves 94% OCR accuracy across receipt formats, languages, and image quality, and 92% document classification accuracy, without retraining for new merchants or document formats.

All Success Stories

Case Study · FinTech · AI

Tourist Tax Refund, Reimagined with Secure AI

An AI-driven tax-refund platform for tourists shopping in Mexico, designed and orchestrated at Vertekx. Machine learning and generative AI are embedded across the entire refund lifecycle, wrapped in a security-first architecture built for PCI and PII compliance.

Machine LearningGenerative AIPCI-DSSPII ComplianceHuman-in-the-Loop

94%OCR accuracy across receipt formats

92%Document classification accuracy

90%PCI scope reduction via vaulting

The Challenge

Thousands of receipts. Sensitive card data. Zero room for error.

Every refund claim arrives as a bundle of photographed receipts, passports, and boarding passes in varying formats, languages, and quality. Each one must be classified, read, and validated before a refund can be issued to the tourist's credit card.

Traditional OCR made this worse, not better: every document type demanded extensive training, and even a small format change meant re-training the model on hundreds of samples before results were usable again. Every new merchant became a bottleneck.

And doing it with AI raises a harder question still: how do you let models process documents and payments without ever exposing cardholder data?

Unstructured documents at volumeReceipts and travel documents arrive in every format and quality imaginable, and all must be machine-readable.

Cardholder data in the flowRefunds pay out to credit cards, putting the platform squarely inside PCI-DSS territory.

High cost of a wrong decisionA bad approval is fraud exposure; a bad rejection is a tourist who never gets their money back.

OCR that couldn't scaleLegacy OCR needed hundreds of training samples per document format. Every new merchant meant weeks of re-training.

The Solution

AI across the entire refund lifecycle

From the moment a tourist submits a claim to the moment the refund lands on their card, every step is automated, with human-in-the-loop review exactly where judgment matters.

Document classificationCustom ML models sort every incoming document (receipt, passport, boarding pass) before a human ever touches the case.

Zero-training OCRGenerative AI reads documents by contextual understanding, not templates: new invoice, POS, or payment-slip formats work from day one, with 94% accuracy and no re-training. Unlimited merchants, unlimited scale.

Automated validationExtracted fields are cross-checked against eligibility rules and each other. Inconsistencies are flagged, not guessed.

Mathematical validationAI outputs are verified with hard math wherever possible: Luhn and BIN checks on card numbers, line items summing to the invoice total, tax amounts matching the actual calculation.

Human-in-the-loopFlagged and high-risk cases route to trained reviewers. AI handles the volume; people handle the judgment calls.

Case-success forecastingBefore a claim is submitted, models score its likelihood of approval, so weak cases are fixed early, not rejected late.

Inside the Platform

The system, at work

Every uploaded document is classified and OCRed by AI the moment it arrives. The signed refund form goes further: after pre-processing normalizes scan resolution and image size, AI locates the exact field coordinates on the scanned PDF and fills it with the customer's verified data, no template, no manual retyping.

Passports, receipts and card slips classified and OCRed by AI on arrival

Scanned and signed form filled dynamically by AI

A signed application filled dynamically: AI matches exact coordinates on the normalized scan and writes each verified field

Operations dashboard with load per status

Live operations view: load per status across operations and data analysis

Intake analytics: portal vs QR entries, top airlines, cruises and locations

Security First

Card data is extracted once, then vaulted forever

Tourists submit a card image with the CVV redacted; every other detail stays intact because the refund needs it. PCI-compliant AI/OCR, built on Azure Custom Vision and AWS Bedrock, confidently extracts the card information, and the moment extraction completes, everything is vaulted. From then on, no user can ever see a raw card image; every workflow downstream operates on tokens only.

Redacted at the sourceThe tourist's card image is submitted with the CVV redacted; only the details the refund actually needs are captured.

PCI-compliant AI extractionAzure Custom Vision and AWS Bedrock models running inside the PCI boundary extract card details with confidence scoring.

Vaulted immediatelyThe instant extraction completes, card data is encrypted and vaulted. No raw card image remains visible to any user.

Tokenization everywhereEvery downstream system, including all AI pipelines, references cards by irreversible token, not PAN.

Least-privilege accessDetokenization is possible only at the payment edge, under audit logging and multi-factor controls.

On the roadmap: vaulting at the edgePlanned next: vaulting and tokenizing card data at the point of capture, so it enters the platform already tokenized.

The Results

Measured, not promised

94%

OCR AccuracyFields extracted correctly across receipt formats, languages, and image quality.

92%

Classification AccuracyDocuments routed to the right pipeline automatically, without manual sorting.

90%

PCI Scope ReductionSystems removed from audit scope through vaulting and tokenization.

Every number on this page was earned in the lab first.

These figures come from our proprietary AI benchmarking platform, the same tool we use across all of our AI work. Before any model, prompt, or pipeline touches production, it runs against thousands of real-world test cases spanning every receipt format, language, and image quality we've encountered.

Accuracy, cost, and failure modes are measured side by side across candidate models and architectures. Nothing is marked production-ready on intuition: every claim is evidence, and every decision is data-backed. When the numbers move, we know exactly why; when they don't hold up, it never ships.

From the Operations Team

“

This platform transformed how our team operates. Manual effort per claim is down roughly 80%, and since the OCR needs zero training, we bring new document types online in a day instead of months, letting us scale to volumes that simply weren't possible before. Just as critical: the security-first design keeps sensitive card data out of our systems entirely, cutting our PCI scope by about 90%. We move faster, and we sleep better.

Operations ManagerTourist Tax Refund Program

Building something with sensitive data and high volume?

We design AI systems that regulators, auditors, and customers can all trust. Let's talk about yours.

Get a Free Consultation