Architecture Overview &
Tax Determination Methodology

Executive Summary

Cross-border VAT/GST compliance is one of the most complex problems in financial operations. With over 100 jurisdictions, each enforcing distinct rules for how goods and services are classified and taxed, businesses face an impossible matrix of tax codes, place-of-supply rules, reverse-charge mechanisms, and filing obligations. Nowhere is this sharper than for US-based SaaS companies selling into the EU, where a single digital service sale to a French consumer can trigger an immediate VAT registration obligation and quarterly filing cycle.

DeterminedAI is purpose-built for US SaaS businesses operating across borders. We guide EU VAT OSS Non-Union registration with Irish Revenue end-to-end, determine the correct VAT treatment on every transaction, and once registered, take over quarterly OSS return filing via the customer's ROS Digital Certificate.

Legacy tax engines — Avalara, Vertex, Thomson Reuters — solve this with static rule tables maintained by teams of tax analysts. Every product must be manually mapped to a proprietary tax code. New jurisdictions require additional rule authoring. Edge cases often require vendor involvement to resolve.

DeterminedAI takes a fundamentally different approach. By placing Anthropic's Claude at the center of the classification pipeline, we replace manual tax code mapping with AI-powered transaction characterization. A natural-language transaction description goes in; an auditable, ERP-ready tax determination comes out.

The result: faster onboarding (transaction classification in a fraction of the time compared to weeks of manual tax code mapping), broader jurisdiction coverage (103 countries at launch), and a classification system that understands context — not just keywords. Beyond determination, DeterminedAI now provides end-to-end EU VAT OSS compliance for non-EU businesses — guided Non-Union OSS registration with Irish Revenue and automated quarterly return filing once registered.

The Problem with Legacy Tax Engines

Manual classification at scale is unsustainable

Traditional engines require every product or service to be mapped to a vendor-specific tax code before any determination can occur. Avalara's system uses thousands of AvaTax tax codes. Vertex has its own taxonomy. These codes are proprietary, non-portable, and require expert knowledge to assign correctly.

For a company selling 500 digital products across 25 EU countries, this means a significant upfront mapping effort: understanding each vendor's taxonomy, assigning the right codes for each product, and maintaining those assignments as both your product line and the regulations evolve.

Rule engines don't understand what you sell

A rule engine matches inputs to outputs. It doesn't understand that "a downloadable PDF training course on AWS architecture" is an education service in the EU (reduced VAT) but an electronically supplied service in other contexts. It only knows what tax code a human assigned. If the assignment is wrong, every determination for that product is wrong — silently.

The maintenance burden compounds

VAT rules change constantly. The EU's ViDA (VAT in the Digital Age) initiative, OECD guidelines on digital services, and country-specific reforms (e.g., Saudi Arabia's e-invoicing mandate, Nigeria's VAT scope expansion) all require rule updates. While large vendors like Avalara and Vertex maintain dedicated tax content teams, the burden of mapping your products to their codes remains on you — and that mapping must be revisited whenever rules or your products change.

System Architecture

DeterminedAI is built as a modular pipeline where each stage has a single responsibility. The system is designed so that AI handles what AI does best (understanding transaction descriptions) while deterministic engines handle what must be exact (rate lookup, place-of-supply rules, ERP code mapping).

This separation means the AI is never making tax rate decisions. It classifies what something is. Deterministic engines decide how it's taxed. This is a critical architectural choice: tax rates must be exact and auditable, while transaction understanding requires intelligence and context.

AI-Powered Transaction Characterization

The core innovation of DeterminedAI is using Anthropic's Claude to classify transactions into VAT/GST supply types based on their description and full context (countries, B2B/B2C, direction). This is not a generic text classifier — it is a purpose-built tax characterization system with domain-specific prompting, structured output, and human-in-the-loop review.

Why Claude?

Tax classification requires genuine understanding of what a product or service is and how it is delivered. A keyword matcher might see "software" and assign a digital services code. But Claude understands the difference between:

• SaaS platform access — electronically supplied service (standard digital VAT rules)
• Perpetual software license shipped on USB — goods (physical delivery rules apply)
• Custom software development project — consulting service (general B2B reverse charge rules)
• Software training course (live instructor) — education service (may qualify for reduced rate)

These distinctions directly affect which VAT rate applies, where tax is due, and whether reverse charge applies. Getting them wrong means either overcharging customers or underreporting tax — both carry real financial and legal consequences.

The 25 supply types

Claude classifies every transaction into one of 25 supply types organized across seven categories. This taxonomy maps directly to how VAT/GST legislation categorizes economic activity, ensuring precise treatment for every type of supply:

Structured output with confidence scoring

Claude doesn't return free-form text. Every classification response is structured JSON containing the supply type, a confidence score (0.0 to 1.0), and reasoning. Configurable thresholds enable automated workflows:

• High confidence — automatically classified, human review optional
• Medium confidence — classified but flagged for review
• Low confidence — requires manual review before production use

The Determination Pipeline

Once a transaction's supply type is established (via approved classification or real-time AI), the determination pipeline applies jurisdictional tax rules through a series of deterministic stages.

Stage 1: Place of Supply

Place of supply determines which country's tax rules apply. This is not simply "where the customer is." The rules depend on the supply type, the customer type (B2B vs B2C), and specific legislative provisions:

• TBE services (ESS, Telecom, Broadcasting) — B2C: taxed where the customer is located (EU VAT Directive Art. 58). Note: since January 2019, EU-based suppliers whose cross-border B2C TBE turnover stays below EUR 10,000 per year may apply their home member state's VAT instead of the customer's.
• TBE services — B2B: taxed where the customer is established (reverse charge applies)
• General services — B2B: taxed where the customer is established (Art. 44, reverse charge)
• General services — B2C: taxed where the seller is established (Art. 45)
• Goods: taxed at the place of delivery / destination

DeterminedAI implements these rules as a deterministic engine — no AI involved. The place-of-supply module takes (seller_country, customer_country, customer_type, supply_type) and returns the taxing jurisdiction with full rule traceability.

For businesses selling B2C cross-border in the EU, the One Stop Shop (OSS) scheme is the primary compliance simplification. Rather than registering for VAT in every member state where customers are located, suppliers can register for OSS in a single member state and report all EU B2C supplies through one quarterly return. The Union OSS is available to EU-established businesses, while the Non-Union OSS serves suppliers established outside the EU. Together, these schemes dramatically reduce the number of VAT registrations a SaaS company needs to maintain. DeterminedAI's determination output includes the taxing jurisdiction and applicable rate per transaction, which maps directly to the data required for OSS reporting.

Stage 2: Rate Lookup

With the taxing jurisdiction and supply type established, the rate engine looks up the applicable tax rate from a curated database of 5,100+ rate entries across 103 jurisdictions. Rates are organized by:

• Country — the taxing jurisdiction
• Supply type — different rates may apply (e.g., education services at reduced rate)
• Tax treatment — standard, reduced, zero-rated, exempt, or reverse charge

The rate database is maintained as version-controlled YAML files, enabling audit trails for every rate change and easy review by tax professionals.

Stage 3: ERP Tax Code Mapping

The final stage translates the determination result into a tax code that the customer's ERP system understands. Not every ERP exposes a fixed tax code vocabulary — some vendors deliver pre-defined codes per country, while others (QuickBooks, Dynamics 365, SAP) use customer-defined strings that only exist inside each company's installation. DeterminedAI reflects this reality through a two-tier integration model, and never fabricates codes we cannot verify against vendor documentation.

Tier 1 ERPs ship with verified static defaults that DeterminedAI generates from vendor-published sources (Oracle NetSuite SuiteTax documentation, the open-source XeroAPI/Xero-NetStandard TaxType enum, and Sage Intacct API example payloads). Customers can start calculating tax the moment their API key is issued, with no additional configuration for the covered jurisdictions.

Tier 2 ERPs have no static defaults because their tax codes only exist inside each customer's own installation. During onboarding, DeterminedAI either calls the vendor's discovery endpoint directly (where available) or collects the customer's code list via POST /v1/setup/erp-mapping. Until that configuration is in place, our adapter APIs explicitly return null for the ERP tax code and emit a warning instructing the caller not to write that line back to the ERP — we will not invent a code that could be silently rejected by the ERP's write API.

Every customer — Tier 1 or Tier 2 — can also override or extend mappings for their specific configuration through the dashboard or the mapping API, and our Raw API endpoint returns the underlying determination without any ERP-specific formatting for systems without a dedicated adapter.

DeterminedAI vs. Legacy Tax Engines

Security & Compliance

Data handling

• Encryption in transit: TLS 1.2+ for all API communication
• Encryption at rest: AES-256 for stored data
• API authentication: JWT bearer tokens (Supabase Auth) or SHA-256 hashed API keys
• No training on customer data: Transaction descriptions sent to Claude are not used to train AI models (Anthropic's data retention policy)

Audit trail

Every tax determination is logged with full input/output detail: transaction ID, transaction classification, place-of-supply reasoning, rate applied, ERP tax code returned, and timestamp. Determination history is stored for your records, with configurable retention periods to support tax authority inquiries.

GDPR compliance

• DeterminedAI acts as a data processor; the customer is the data controller
• Standard Contractual Clauses (Module 2) for international transfers
• Data Processing Agreement available on request
• Right to access, correct, delete, and export data within 30 days

Infrastructure

Integration Guide

DeterminedAI integrates with your existing systems through a REST API that mirrors the request/response patterns ERP developers already understand.

Onboarding workflow

1. Upload your transactions — CSV with descriptions, countries, and amounts. AI classifies each transaction into a supply type with a confidence score.
2. Review and approve classifications — Dashboard shows AI recommendations with confidence indicators. Approve individually or in bulk. Reject and reclassify as needed.
3. Select your ERP and configure mappings — Tier 1 ERPs (NetSuite SuiteTax for 51 verified countries, Xero UK/AU/NZ/SG, Sage Intacct Australia) ship with verified static defaults you can use out-of-the-box. Tier 2 ERPs (QuickBooks Online, Dynamics 365, SAP S/4HANA) and Tier 1 jurisdictions outside the verified set (e.g., Xero Global edition, Sage Intacct UK/ZA) supply their tax codes during onboarding via POST /v1/setup/erp-mapping. The Raw API endpoint returns determinations without ERP-specific formatting.
4. Add your VAT registrations — Register the jurisdictions where you collect VAT/GST. This ensures determinations respect your nexus.
5. Integrate the API — Point your ERP or billing system at the determination endpoint. Send transaction data, receive tax calculations with ERP-ready tax codes.

API request example

API response example

Jurisdiction Coverage

DeterminedAI launches with rate data for 103 jurisdictions across all major economic regions. Each jurisdiction includes rates across all 25 supply types, covering standard rates, reduced rates, zero-rated supplies, and exempt categories.

Coverage is expanded continuously. New jurisdictions are added by authoring rate entries in the version-controlled YAML database — no code changes required.

Scope & Limitations

DeterminedAI is purpose-built for VAT and GST determination on cross-border transactions. It is important to understand what the system does and does not cover:

In scope

• AI-powered transaction classification into VAT/GST supply types
• Tax rate determination across 103 VAT/GST jurisdictions
• Place-of-supply calculation for B2B and B2C transactions
• ERP tax code mapping for six major ERP systems — Tier 1 verified defaults for NetSuite SuiteTax, Xero, and Sage Intacct; Tier 2 onboarding discovery for QuickBooks Online, Dynamics 365, and SAP S/4HANA — plus a Raw API endpoint
• EU VAT OSS Non-Union registration — guided end-to-end registration with Irish Revenue, credential management, and full status tracking through approval and Digital Certificate retrieval
• Automated quarterly EU OSS returns — once registered, DeterminedAI calculates and submits quarterly VAT returns via ROS Digital Certificate with no manual intervention
• Audit trail with full determination history and configurable retention

Out of scope (current release)

• US state sales tax — US sales/use tax is a fundamentally different regime (origin-based vs. destination-based varies by state, product taxability rules differ from VAT, nexus thresholds are state-specific). This requires a separate product
• Non-OSS VAT return filing — DeterminedAI provides EU OSS quarterly return automation. Filing obligations outside the OSS scheme (e.g. local VAT registrations in individual member states) are not yet automated
• VAT number validation — the system validates EU VAT numbers against the VIES database to verify B2B status; non-EU tax ID validation is not yet supported
• E-invoicing / real-time reporting — mandates like Italy's SDI, Saudi Arabia's FATOORA, or the EU's upcoming ViDA real-time reporting are not yet supported
• Customs duties and import VAT calculation — cross-border goods shipments may trigger import duties that are outside the current determination scope

These limitations represent areas for future development, not architectural constraints. The modular pipeline is designed to accommodate additional tax regimes and compliance features as they are built.

EU VAT OSS Compliance — Registration & Automated Returns

For businesses established outside the EU that make B2C digital service sales to EU consumers, the Non-Union One-Stop-Shop (OSS) scheme provides a single-country registration path. Rather than registering for VAT in each of the 27 EU member states individually, a non-EU supplier registers once — typically with Irish Revenue via the ROS portal — and submits a single quarterly return covering all EU B2C supplies.

The registration journey

DeterminedAI guides users through the complete Non-Union OSS registration lifecycle as a structured, tracked workflow. The process involves four distinct phases:

1. Application preparation — A four-step wizard collects all information Revenue requires: company details and bank account, primary and secondary contacts, trading history, and a user-set Verification Code. The completed form maps directly to the fields on the ROS vatmoss-web registration form.
2. Submission and Revenue review — The user submits the application on ros.ie/vatmoss-web. DeterminedAI tracks the status through each stage: submitted, awaiting confirmation, responded, and approved. Revenue typically responds within a few business days and may request additional information via the Primary Contact email.
3. Post-approval credential retrieval — Once approved, Revenue emails a System Password to the Primary Contact. The user then logs in to ROS using their Primary Contact email, Verification Code, and System Password to retrieve their OSS Registration Number and download their ROS Digital Certificate (.p12 file). All credentials are stored encrypted at rest using AES-128-CBC with HMAC-SHA256 envelope encryption, and are never returned in plaintext except via an explicit reveal endpoint that is audit-logged.
4. Active automated filing — Once the Digital Certificate is uploaded, DeterminedAI takes over quarterly return filing. The system calculates VAT owed per EU member state from your transaction data and submits the OSS return via ROS on your behalf before each quarterly deadline.

Security model for credentials

Three secrets are stored during the lifecycle: the Verification Code (set by the user during registration), the System Password (emailed by Revenue on approval), and the Digital Certificate with passphrase. Each is encrypted with Fernet symmetric encryption keyed to a server-side master key (DVAT_MASTER_KEY). The Verification Code last four digits are stored separately for user identification. Every reveal operation is logged to an append-only audit trail alongside all status transitions and field edits.

Effective date calculation

Revenue grants OSS registration from the first day of the quarter in which the application is submitted — they do not backdate. DeterminedAI calculates the effective date automatically and flags any gap between the date of first EU B2C supply and the effective registration date, helping users understand and address any pre-OSS exposure before filing their first return.

Conclusion

The tax technology industry has operated on the same paradigm for two decades: hire tax analysts, build rule tables, sell annual licenses. DeterminedAI challenges this model by demonstrating that AI-powered classification, combined with deterministic tax engines, delivers faster, cheaper, and more accurate results.

Claude provides the intelligence layer — understanding what is being transacted and how it should be categorized under VAT/GST law. Deterministic engines provide the precision layer — applying exact rates, place-of-supply rules, and ERP mappings that must be auditable and reproducible.

This architecture is not a compromise. It is the right tool for each part of the problem. AI for understanding. Rules for precision. Humans for approval. Together, they form a tax determination system that is more capable, more accessible, and more transparent than anything that came before.