Incoterms, Multi-Country Sourcing, and the Tax Code Problem

Nobody buys from just their own country anymore. A mid-sized machinery manufacturer sources castings from Poland, electronics from Taiwan via a Dutch distributor, software licenses from Ireland, and packaging from a German supplier who in turn imports from the Czech Republic. Every purchase order carries a three-letter clause: EXW, FCA, FOB, CIF, DAP, DDP. And somewhere at the end of the chain sits a person who has to assign exactly the right tax code to that one invoice.

That's where it routinely goes wrong. Not because people are careless, but because the task is structurally underestimated.

What Incoterms Are — and What They Explicitly Are Not

Incoterms govern three things: when risk transfers, who bears which transport costs, and who handles customs clearance. It's a commercial, logistical agreement. Nothing more.

What Incoterms do not govern: the VAT treatment. There is no tax code mapped to an Incoterm clause. "DDP" is not a tax scenario. And yet that's one of the most common reasoning errors in practice — reading the clause off the order and mentally building a 1:1 mapping to a tax code.

The truth is subtler: the Incoterm is not a decision-maker, but an important indicator. It influences who owes import VAT, in which country the place of supply lies, whether a registration obligation arises in the destination country — but always only in combination with other facts. A DDP supplier from a non-EU country shipping to Germany owes German import VAT and may need a German registration. The same supplier under DAP — and suddenly the buyer is the importer, with entirely different input VAT recovery and a different tax code.

Same goods. Same supplier. One letter changed. Different tax code.

Why Multi-Country Sourcing Escalates the Problem

For a simple domestic invoice, the tax code is usually trivial. The complexity doesn't come from any single country — it comes from the combination:

• Chain transactions. A in Italy sells to B in Germany, B sells to C in France, and the goods move directly from Italy to France. Which leg of this chain is the "moving" supply? That determines who has the exempt intra-community supply and who pays VAT where. The Incoterm between the parties is one of the indicators for assigning the transport.
• Triangular transactions and their simplification rule — which only applies if every single condition is met exactly.
• Consignment and call-off stock (introduced with the 2020 Quick Fixes) — the goods move, but the taxable event happens at a different time.
• Non-EU imports with EU interim warehousing, where the Incoterm decides whether there's an intra-community event at all or an importation.
• Construction and other cross-border services, where reverse charge applies — or doesn't, depending on establishment and place of supply.

Each of these constellations requires six to eight individual facts that live in completely different places: the Incoterm on the order, the VAT IDs in master data, the actual goods movement on the delivery note, the type of service in the contract, the position in the supply chain in the heads of the procurement team. Nobody sees all these facts in one place. And that's exactly why people default to "whatever worked last time."

Why Generic AI Doesn't Solve This

The obvious 2026 thought: "Let's just ask a large language model." That doesn't work reliably, for the same reasons we've described in detail elsewhere — generic models sound competent but get specific tax cases dangerously wrong. An LLM will explain Incoterms straight out of the textbook. It will tell you with full confidence that this is a triangular transaction — while missing that one of the three parties doesn't hold a valid VAT ID in the right country.

This matches exactly the pattern we see with AI in accounting generally: strong at interpretation, dangerous at unsupervised decisions. And it's the same structural problem as the classic SAP tax code tables: a rigid mapping can't capture the context, and a hallucinating model captures it wrong.

Where a Specially Trained Small Language Model Actually Helps

The answer isn't a bigger AI. It's a division-of-labor architecture with a small, domain-trained model at its center.

The key insight: don't ask the model which tax code it should be. Let it do the hard part that AI is genuinely good at — fact extraction and normalization:

• Read the Incoterm off the order and interpret it correctly — including spelling variants, free-text entries, and the question of between which parties it applies.
• Reconcile establishment and VAT-ID status of all parties from master data and the invoice.
• Distinguish goods from services, reconstruct the actual goods movement from the delivery note, identify the position in a possible chain.

These normalized facts then feed into a deterministic rule engine that derives the tax code — traceable, audit-proof, with no room for hallucination. No creativity needed where tax law is a decision tree.

Why a small, specially trained model rather than the largest one available?

Sensitive procurement and supplier data is among the most confidential information a company has. A model small enough to run in your own environment solves the data privacy problem as a side effect — and is cheap enough to apply to every single incoming invoice, not just the suspicious ones.

What This Means for Finance Teams in Practice

Three pragmatic conclusions:

• Never read the tax code off the Incoterm alone. The Incoterm is evidence, not a verdict. Treating it as a tax scenario systematically produces audit findings.
• Capture the underlying facts, not just the result. A documented path — which VAT ID, which goods movement, which rule — is worth more at the next tax audit than a "fitting" code with no justification.
• Start narrow. Not "automate all of sourcing," but solve one recurring, painful scenario cleanly — say, non-EU imports under DAP/DDP — learn, then expand.

AI replaces no one here. It takes the reconstruction work off the accountant's plate and hands them a justified, verifiable recommendation. Responsibility stays with the human — but they work with the context instead of against it.

This exact combination — AI-assisted fact extraction plus rule-based, audit-proof tax logic — is what we work on at HybridAI, including VAT Intelligence for cross-border scenarios. Not a generic chatbot guessing tax codes. A system that understands there's more between an Incoterm and a correct tax code than a lookup.