Cross-border diligence fails in a specific, repeatable way, and it is not translation. It is that a provision which is unremarkable in one jurisdiction is material in another, and a screening system tuned on one legal tradition will confidently rank it as routine. The document was read correctly. The risk was assessed against the wrong baseline — and nothing in the output shows that happened.
Four distinct problems, usually treated as one
“Cross-border” collapses four separable difficulties. They have different fixes and different residual risks, and teams that treat them as a single “does it handle international deals” question end up buying against the easiest one.
| Problem | What actually goes wrong | Tractable? |
|---|---|---|
| Language | Source text is not in English | Largely — the most solved of the four |
| Legal-concept mismatch | A term has no clean equivalent across systems | Partially, and only if the system says when it is unsure |
| Materiality baseline | Routine in one market, serious in another | Only with per-jurisdiction configuration |
| Regulatory surface | Filings, approvals and thresholds differ entirely | Rules-based, not model-based |
The third row is where cross-border deals actually go wrong, and it is the one no demo tests.
Why the materiality baseline is the real risk
Consider an employment provision that is standard in one market and creates substantial transfer liability in another. Or a security interest that perfects differently. Or a shareholder consent right that is boilerplate in one jurisdiction and a genuine transaction blocker in another.
In each case the model reads the clause accurately. It extracts the right text. It may even quote it correctly. Then it assigns a severity — and severity is where jurisdiction lives. A system whose sense of “serious” was formed on one legal tradition will rank that clause low, and the output will look exactly like a correct low-risk finding, because that is what it is under the wrong baseline.
It also explains why translation-quality questions dominate vendor conversations while producing little protection. Translation failures are visible. Baseline failures are invisible, which makes them the ones that reach closing.
What to require, in order of value
1. Jurisdiction declared per document, not per deal
A cross-border transaction is not one jurisdiction. The target's customer contracts may sit under three governing laws and the employment documents under a fourth. If the system applies a single deal-level jurisdiction setting, it is applying the wrong baseline to a large share of the set by construction.
2. Per-jurisdiction scoring configuration
The same clause type must be able to score differently depending on the governing law. If severity is a fixed global property of a clause type, the tool cannot represent the problem in the first place, and no amount of tuning will fix it.
3. Original-language source text in the output
Every finding should carry the source sentence in the language it was written in, alongside any translation. This matters for a reason beyond verification: local counsel reviewing the finding needs to read what the document says, not a round-trip through English. It also makes translation errors visible rather than baked in.
4. Explicit uncertainty on untranslatable concepts
When a legal concept has no clean equivalent, the correct behaviour is to flag it as requiring local review — not to map it to the nearest English term and proceed. Ask a vendor directly what their system does here. Many map silently, and it will not show up in a demo run on English documents.
5. Coverage reporting by language and jurisdiction
You need to know which parts of the data room were examined under which baseline, and which languages were handled poorly or skipped. Absence of findings in a language block is not evidence of clean documents.
Where screening earns the most on cross-border deals
Cross-border is, in fact, the setting where a scoring stage is worth the most — for a reason unrelated to speed.
The constraint on international diligence is local counsel time. It is expensive, engaged late, in another time zone, and it is the bottleneck on the critical path. Teams manage this by sending local counsel a document set chosen by whatever the deal team could tell was relevant — which, across a language barrier, is a weak filter.
A screening pass that scores every document against every category, per jurisdiction, changes what gets sent. Instead of “here are the 200 documents we think matter,” local counsel receives a ranked set with the source clause attached and a reason for the rank. Three consequences follow:
- Local counsel time goes to the highest-ranked material rather than to whatever the deal team happened to recognise.
- The ranking itself becomes reviewable. Local counsel can say “that clause is routine here, drop it” or, far more valuably, “this one you ranked low is a blocker in this jurisdiction” — which is precisely the baseline correction the system cannot make on its own.
- The correction is recordable. Overrides captured against a category become the configuration for the next deal in that jurisdiction, so the baseline improves through use rather than through vendor updates.
That third point is the durable one. It reframes local counsel from a reviewer of documents into a calibrator of the baseline, which is a much better use of the scarcest resource on the deal.
The regulatory layer is not a model problem
Merger control thresholds, foreign investment screening, sector-specific approvals, filing deadlines — these are rules, not judgments, and they should be handled by rules.
Asking a language model to determine whether a filing is required is a category error. The correct architecture is a deterministic check against maintained thresholds, with the model's role limited to extracting the inputs — turnover, asset values, shareholding percentages, sector — from the documents and quoting where each came from. Where a threshold is close, the output should say so and escalate, rather than resolving it.
This division matters commercially, because a missed filing is among the few diligence failures that can unwind a completed transaction. It is not a place for probabilistic output.
What stays human, permanently
- Local law materiality. Whether a provision is serious in a given market is a question for someone qualified in that market. A score is a prompt for that question, not an answer to it.
- Enforceability. A clause that reads identically in two jurisdictions may be enforceable in one and void in the other, and the document does not say which.
- Commercial and cultural practice. What is customary is frequently unwritten, and no system reads what was never documented.
- Anything not in the data room. This bites harder cross-border, where local disclosure norms differ from what the buyer's team expects to be given.
Bottom line
Translation is the problem everyone asks about and the least dangerous of the four. The one that reaches closing is a clause read perfectly and scored against the wrong jurisdiction's sense of what is serious — invisible in the report, because a wrongly-calibrated low score looks exactly like a correct one.
So require per-document jurisdiction, per-jurisdiction severity, original-language source text, and explicit uncertainty on concepts that do not map. Then use the ranking to point expensive local counsel at the right material first, and capture their corrections — because the baseline you need cannot be bought, only accumulated.
Sources
This page makes no external statistical claims. Its arguments are structural and can be tested against your own deals. Where we cite figures elsewhere on this site, we cite only sources we have retrieved and read — see our methodology. Nothing here is legal advice; cross-border materiality is a question for counsel qualified in the relevant jurisdiction.