Blog
AP Automation5 min12 September 2026

Why a generic AI model fails at accounting, and jurisdiction-aware validation doesn't

A model that reads French, US and UK documents equally well can still validate them equally badly — because reading and validating are different problems with different answers per jurisdiction.

A modern language model reads a French invoice, a US receipt, and a UK expense claim with roughly equal fluency — extracting vendor names, amounts, and dates regardless of language is a solved problem. The mistake is assuming that because extraction is jurisdiction-agnostic, VALIDATION can be too. It can't, and the gap between the two is where a lot of "AI accounting" tools quietly produce wrong answers with high confidence.

This is a different question from what changes operationally when a practice or BPO serves multiple markets — compliance calendars, filing deadlines, staffing. This is about what changes INSIDE the software's own validation logic, underneath whatever operational process sits on top of it.

Reading and validating are different problems

Extraction asks: what does this document say? Validation asks: is what it says CORRECT, given the rules that actually apply to a document like this, from this place? A model can answer the first question the same way everywhere. The second question has a different correct answer depending on jurisdiction, and no amount of general-purpose language fluency substitutes for knowing which specific rule applies where.

Where a generic approach actually breaks

A legal identifier isn't one universal format

A French business identifier follows one check-digit algorithm; a US employer identification number follows a different structure entirely; a UK company number follows yet another. A validation check written generically — "does this look like a valid business ID" — either accepts things it shouldn't or rejects things it shouldn't, because "valid" only means something relative to a specific format.

A tax rate that's normal in one place is a red flag in another

A VAT rate applied to a service in France follows France's own rate structure and exemption rules. Apply the same expectation to a US document — which has no VAT at all, only state sales tax with completely different logic — and a generic check either fires false positives constantly or gets disabled entirely, which defeats the point of having it.

A bank detail format isn't interchangeable

An IBAN validates one way. A US routing number plus account number validates a completely different way, with no IBAN in the picture at all. A fraud check built assuming every payment reference is an IBAN simply doesn't fire on a domestic US payment — not because the check failed, but because it was never applicable to begin with, and a system that doesn't know that silently skips a real control.

Reasonable amounts differ by context, not by a fixed number

What counts as an unusually large, worth-flagging expense varies by market and by document type — a fixed global threshold either flags too aggressively in a lower-cost context or misses genuinely anomalous amounts in a higher-cost one.

What "jurisdiction-aware" actually means in practice

It means the SET OF CHECKS that runs on a document is determined by that document's own detected jurisdiction — not a single universal checklist applied everywhere, and not three separate products bolted together. A French invoice triggers French-specific checks; a US document triggers US-specific ones; and checks that are genuinely universal (does the arithmetic reconcile, is the date in the future) run everywhere because they actually are jurisdiction-independent.

Check categoryFranceUnited StatesUnited Kingdom

|------------------|--------|----------------|------------------|

Legal identifierBusiness ID with its own check-digit algorithmEmployer identification number, different structure entirelyCompany number, a different structure again
Consumption tax logicVAT rate tied to the type of supplyNo VAT — state-level sales tax, different logic entirelyVAT, but its own rate structure and thresholds
Bank detail formatIBANRouting number + account number, no IBAN involvedSort code + account number
Arithmetic/date sanityUniversal — applies identically everywhereUniversalUniversal

The practical test for any vendor claiming multi-market support

Ask what changes, mechanically, when the same vendor processes a French invoice versus a US invoice versus a UK invoice — not whether both get read correctly (they likely do), but whether the VALIDATION LOGIC itself is different, and specifically what's different about it. "We support multiple languages" is not the same claim as "we validate against multiple jurisdictions' actual rules," and the two get conflated constantly.

Related reading

FAQ

Does jurisdiction-aware validation require a separate model per country?

No — the underlying extraction model can stay the same; what needs to differ is the RULE SET applied after extraction, based on the document's detected jurisdiction. That's a configuration and logic difference, not a separate model per market.

How is a document's jurisdiction actually determined?

Typically from a combination of signals already present in the document itself — currency, language, legal identifier format, address — cross-checked against the account's own market rather than assumed from a single weak signal like language alone (a French-language document can still be issued from a French subsidiary of a US company, for instance).

What happens to a document whose jurisdiction can't be confidently determined?

A defensible system defaults to the narrower, more general check set rather than guessing a specific jurisdiction's rules and applying them incorrectly — an unclear jurisdiction should trigger conservative handling, not a confident wrong answer.

Ready to try DOXALIO?

Free trial. No credit card required.

Get started for free
Why a generic AI model fails at accounting, and jurisdiction-aware validation doesn't — DOXALIO Blog