Detect. Determine. Draft.
AI-Native VAT Automation.

Executive Summary

DeterminedAI is a VAT/GST automation platform built for cross-border SaaS, marketplaces, and digital businesses. One product covers the work up to the moment you file: detect, determine, draft.

• Detect: Stripe-native exposure check and CSV ingest surface every jurisdiction where a seller has crossed a registration threshold, across 60+ nonresident regimes, before a tax authority finds it first.
• Determine: AI classifies any transaction into 25 supply types with confidence scoring; deterministic engines apply place-of-supply and rates across 103 jurisdictions.
• Ask: the AI Tax Assistant is the front door to determination. Ask any VAT/GST question (treatment, invoicing requirements, place of supply) and get an advisor-grade answer with the rule that applied, scoped to the jurisdictions you subscribe to, with saved history.
• Draft: upload a period's transaction data (CSV, Excel, or the invoices themselves) and get every box of that country's return computed, an XLSX workpaper with full workings, a draft return PDF laid out on the authority's form structure, and an AI review that flags issues to check. You review the draft and file it yourself, through your usual channel.
• Free public tools: Multi-country tax-ID validator (VIES, HMRC, Brønnøysund), filing-deadline tracker, and global e-invoicing mandate tracker.

Legacy engines (Avalara, Vertex, Thomson Reuters ONESOURCE) require manual mapping of every SKU to a proprietary tax code, multi-month implementations, and annual contracts in the five-to-six figure range. They sell the engine but leave exposure analysis, registration, and filing to an army of consultants.

DeterminedAI monitors exposure, answers determination questions in real time through a conversational Tax Assistant, and turns uploaded transaction data into full return workings and a reviewed draft return, from one product, one contract, one price sheet. Onboarding is hours, not months. The Compliance subscription bundles the engine, monitoring, and the Tax Assistant at $99 per jurisdiction per month (EU OSS counts as a single jurisdiction covering all 27 EU countries). Filing itself stays with you: DeterminedAI prepares, you review and submit. No SOWs, no implementation partners.

Why Legacy Tax Engines Fail Modern SaaS

Manual mapping is the bottleneck. Avalara has thousands of AvaTax codes; Vertex has its own taxonomy; ONESOURCE has another. Every SKU must be mapped before a single determination runs. A 500-product catalog across 25 EU countries is a multi-month project before go-live.

Rule engines don't understand context. A keyword matcher sees "software training" and picks one code. It can't distinguish a live instructor-led course (education service, possible reduced rate) from a pre-recorded video (electronically supplied service, standard rate). Wrong assignment, silent miscalculation, every transaction.

They sell determination, not compliance. Legacy engines stop at "what's the rate." Registration in each jurisdiction, return preparation, e-filing: that's all sold separately by their professional services arm or punted to a third-party partner. ViDA, KSeF, FATOORA, MTD: each is a different SOW.

Pricing assumes enterprise. Annual contracts, implementation fees, and per-transaction overage charges scale poorly for early-stage SaaS that just needs to file an EU OSS return on $40k of B2C revenue.

Exposure & Nexus Monitoring

Determination only matters once you know where you owe. For nonresident sellers, the hardest question isn't “what's the rate”: it's “which jurisdictions have I already crossed a registration threshold in, and didn't notice?” Avalara, Vertex, and ONESOURCE assume the customer already knows. DeterminedAI starts a step earlier.

Inputs

• Stripe sync: Connect a restricted API key. We pull Charges, Invoices, and Customer location data (billing address, tax ID, IP geolocation) and reconstruct a per-country revenue and headcount picture.
• CSV upload: For non-Stripe billing stacks (Chargebee, Recurly, Paddle, in-house). Up to 100,000 rows per upload, mapped via a column-picker UI.

Threshold engine

A version-controlled threshold registry covers 60+ nonresident regimes: EU OSS Non-Union (no threshold), UK (£0 for digital B2C), Norway VOEC (NOK 50,000), Singapore OVR (SGD 100,000 global + SGD 100,000 to Singapore), Japan JCT (JPY 10M), Australia Simplified GST (AUD 75,000), India OIDAR (no threshold), and the long tail of LatAm, EMEA, and APAC digital-services regimes. Each threshold tracks its window (rolling 12-month, calendar year, prior year, or trailing four quarters) and the metric type (gross revenue, taxable revenue, transaction count, or a combination).

Output

A per-jurisdiction status board with four states: safe (under threshold with comfortable margin), approaching (within 20% of threshold), crossed (over threshold, registration required), and overdue (crossed more than 30 days ago without an active registration). Each row links to the relevant Country VAT guide and an estimated back-tax exposure if the threshold was crossed in a prior period.

System Architecture

A modular pipeline. AI does what AI does best: understand descriptions. Deterministic engines do what must be exact: rates and place-of-supply.

A second pipeline runs in parallel for exposure analysis (same threshold registry and jurisdiction data, different ingestion path):

The AI never sets tax rates. It classifies what a transaction is; deterministic engines decide how it's taxed. Rates stay auditable and reproducible; classification gains real understanding.

AI-Powered Transaction Classification

A frontier large language model classifies transactions into VAT/GST supply types from natural-language descriptions plus context (countries, B2B/B2C, direction). Domain-specific prompting, structured JSON output, confidence scores, and human-in-the-loop approval.

A keyword matcher sees "software" and picks one code. The model distinguishes:

• SaaS subscription: electronically supplied service
• Perpetual license on USB: goods
• Custom software development: consulting (B2B reverse charge)
• Live-instructor training: education (potential reduced rate)

25 supply types across 7 categories

Confidence scoring & caching

Every response is structured JSON with a 0.0–1.0 confidence score and reasoning. High-confidence classifications auto-approve; medium and low are flagged for review. Once approved, the (product_code → supply_type) mapping is cached. The AI is not called again for that SKU. AI is onboarding acceleration, not runtime cost.

Determination Pipeline

Place of Supply

A deterministic engine maps (seller_country, customer_country, customer_type, supply_type) to the taxing jurisdiction with full rule traceability. Implements EU VAT Directive Arts. 44/45/58, the €10,000 micro-business threshold, B2B reverse charge, TBE rules, destination-based goods, and analogous rules outside the EU. Output feeds OSS reporting directly: one return covers all 27 EU member states.

Rate Lookup

5,100+ rate entries across 103 jurisdictions in version-controlled YAML, sourced from the EY Worldwide VAT, GST and Sales Tax Guide 2026. Indexed by country × supply type × treatment (standard, reduced, zero-rated, exempt, reverse charge). Every rate change is a tracked diff. No opaque vendor release notes.

AI-Assisted Return Workings & Drafts

Determination is table stakes. The work that actually consumes finance teams is the return itself: pulling a period's transactions together, computing every box, and producing a workpaper a reviewer can trace. DeterminedAI automates that preparation and hands back a finished draft. Filing stays with you. You review the draft and submit it through your usual channel; DeterminedAI never lodges anything with a tax authority.

From upload to computed boxes

Choose a country and period, upload the period's transaction data, and the engine computes every box of that jurisdiction's VAT/GST return from a version-controlled per-country return spec: which rows feed which box, how rates and directions map, how output and input tax net off. Box values are always computed in code from your rows, never generated by a model.

AI in the loop, deterministic at the core

The AI that classifies transactions carries through to the returns layer. Three things happen with a language model in the loop, each backed by deterministic checks so a tax figure is never produced by the model alone:

• Any-layout ingestion. Upload a CSV or Excel export in any column shape. A synonym table maps the obvious columns deterministically; when it cannot find a date plus the amount columns a return needs, the model maps the remaining column names to fields. Cell values are always read from the file in code, so the model never transcribes a number.
• Invoice reading. Upload the actual invoice PDFs or photos and the model extracts each line, shaped to the return's row format. Every line carries a confidence score; each invoice is reconciled in code against its own stated totals, and low confidence, a failed reconciliation, or a missing field flags the invoice for human review before any of its rows feed a return.
• Draft review. Before you file, the model reviews the computed draft against a PII-free, code-computed breakdown of the rows (aggregated by jurisdiction, rate, direction, and B2B/B2C) and raises issues as plain-language questions: rows taxed in a jurisdiction other than the return's country, an unexpected or missing rate, in-period rows that mapped to no box, an all-zero return, mixed currencies. The figures it cites are computed in code; the model phrases the questions, it does not invent the numbers.

All three degrade safely: with no model available the ingestion falls back to deterministic column mapping and the draft review falls back to a deterministic issue checker, so the return still builds.

The deliverables

• On-screen draft: the computed return, box by box, free to view as soon as the data is in.
• XLSX workpaper: four sheets. Cover, Summary, box-by-box Workings, and the Raw Data behind every figure, so a reviewer can trace any box back to the rows that produced it.
• Draft return PDF: a branded draft laid out on the authority's own form structure, so what you check is shaped like what you will enter in the filing portal.

A VAT Return Pack credit unlocks the workings XLSX and the draft return PDF for one country, year, and period, and a stored record of each return generated remains retrievable afterward. Submission is yours: once the draft survives review, you file it with the tax authority the way you always have.

Free Public Tools

No login, no rate limits, no upsell wall. Used as lead generation but genuinely useful in their own right, and not offered by Avalara or Vertex.

• Tax-ID Validator (/tools/validate): Batch VAT/GST/ABN validation across ~20 countries via VIES (EU), HMRC (GB), Brønnøysund (NO) plus regex format checks. CSV upload, downloadable results.
• Filing Deadline Tracker (/tools/deadlines): Searchable VAT/GST filing frequency and due dates by jurisdiction, sourced from primary tax-authority publications.
• E-invoicing Mandate Tracker (/tools/einvoicing): Filterable global compliance status: SdI (IT), KSeF (PL), ZATCA FATOORA (SA), SII (CL), MyDATA (GR), Peppol rollouts, and EU ViDA timelines.
• Stripe Exposure Check (/exposure): Connect a Stripe restricted key, get a per-jurisdiction exposure snapshot from live Charges plus Customer location data. Snapshot persistence is a $29/mo upsell; the check itself is free.

DeterminedAI vs. Avalara, Vertex & Thomson Reuters

The Tax Assistant

The determination engine is delivered through a conversational interface: the AI Tax Assistant. Instead of mapping products to codes or wiring an integration, a user describes a transaction or asks a question in plain language and gets an advisor-grade answer in the chat, with the reasoning and the rule that applied shown alongside. Every substantive answer carries an informational, not formal advice, disclaimer.

It answers treatment questions for a transaction type (rate, place of supply, reverse charge, OSS/IOSS, OIDAR), what a compliant invoice must show in a given country, registration thresholds and timing, and the mechanics of a filing. The assistant runs on the strongest available Claude model, scopes country-specific advice to the jurisdictions on the customer's Compliance subscription, and saves every conversation to the account so prior answers and their reasoning stay retrievable.

Onboarding workflow

1. Check exposure: connect Stripe or upload a CSV; the platform surfaces every jurisdiction where a registration threshold has been crossed.
2. Declare nexus: mark each jurisdiction where you are registered or need to be, so answers and return drafts are scoped correctly.
3. Ask and draft: use the Tax Assistant for treatment and invoicing questions across your subscribed jurisdictions, and generate return workings and a draft return when each period is due. Review it, then file through your usual channel.

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.

New jurisdictions are added by authoring rate entries in version-controlled YAML. No code changes required. Source: EY Worldwide VAT, GST and Sales Tax Guide 2026, supplemented by primary tax-authority publications.

Security & Compliance

• Encryption in transit: TLS 1.2+ for all API traffic.
• Encryption at rest: AES-256.
• Authentication: Supabase Auth (Google OAuth, magic link) or SHA-256 hashed API keys.
• No model training on customer data: Anthropic zero-retention applies to determinations.
• Audit trail: every determination and every return generated logged to append-only storage.
• GDPR: data processor, SCC Module 2, DPA on request, 30-day access/correction/deletion SLA.
• Row-Level Security on all multi-tenant Postgres tables (Supabase RLS).
• Infrastructure: Vercel (app), Supabase/AWS (Postgres), Anthropic (LLM), Stripe (billing).

Scope & Limitations

In scope today

• AI classification across 25 supply types with confidence scoring and human-in-the-loop review.
• Rate and place-of-supply determination across 103 VAT/GST jurisdictions.
• AI Tax Assistant for VAT/GST questions, scoped to subscribed jurisdictions, with saved chat history.
• Return workings and draft returns from uploaded data: every box computed from a version-controlled per-country return spec, delivered as an on-screen draft, an XLSX workpaper (Cover, Summary, Workings, Raw Data), and a draft PDF on the authority's form structure. The in-app country picker shows current return coverage.
• AI-assisted return prep: any-layout spreadsheet ingestion with LLM column mapping, invoice-PDF reading with reconciliation + confidence gating, and AI draft review (all with deterministic fallbacks); stored record of every return generated.
• Tax-ID validation across ~20 countries (VIES, HMRC, Brønnøysund + regex).
• Free public tools: validator, deadline tracker, e-invoicing tracker, Stripe exposure check.
• Stripe-native ingest and audit trail with configurable retention.

Out of scope (current release)

• US state sales tax: a different regime (origin vs. destination, product taxability matrices, state-by-state nexus). Separate product on the roadmap.
• Submission to any tax authority: DeterminedAI does not file returns or submit registrations anywhere. The user reviews the workings and the draft, then files through their usual channel.
• E-invoicing & real-time reporting (SdI, KSeF, FATOORA, ViDA): tracked, not yet emitted.
• Customs duties / import VAT calculation on cross-border goods.

Conclusion

Avalara, Vertex, and Thomson Reuters built tax engines for a world where the customer paid consultants to map products and accountants to file returns. DeterminedAI was built for a world where neither budget exists.

The Tax Assistant is the interface. AI is the intelligence layer. Deterministic engines are the precision layer. Return workings and AI-reviewed drafts are the compliance layer, and filing stays in your hands. One product, no SOWs.