Trade finance document review is one of the highest-discrepancy, most labor-intensive workflows in banking, and it runs under a rule that leaves no room for judgment calls: strict compliance. A letter of credit either matches its documents exactly, or the bank has to refuse it. That combination, high volume, high error rate, zero tolerance for "close enough", is exactly the kind of process an AI agent is built for, and exactly the kind of process where getting it wrong is expensive.
Our guide to AI employees in banking and financial services covers trade finance in one paragraph among many regulatory topics. This guide is the deep version: how letter-of-credit review actually works, the specific discrepancies that trip up first presentations, how sanctions screening in trade finance differs from standard KYC, and what a properly governed AI deployment looks like for this specific workflow.
A letter of credit is a bank's promise to pay a seller once the seller presents documents proving the shipment happened on the agreed terms. The document set typically includes a commercial invoice, a bill of lading (the transport document), a packing list, a certificate of origin, and an insurance certificate, along with anything else the specific credit requires.
UCP 600, the ICC's Uniform Customs and Practice for Documentary Credits, is the rulebook that governs how banks examine that document set. It applies to roughly $1 trillion of trade finance a year across 175 countries, and its central rule is the doctrine of strict compliance: documents have to comply with the credit's terms and with each other, examined strictly on their face, with no de minimis exception for a small or immaterial mismatch. Article 14 gives the examining bank a hard deadline, a maximum of five banking days to review the presentation, and Article 16 requires the bank to refuse any presentation containing a discrepancy unless the applicant explicitly waives it.
That combination of a hard deadline and a zero-tolerance standard is why discrepancy rates are so high industry-wide. Estimates put the first-presentation refusal rate at 65-80%, meaning most letters of credit get rejected at least once before they clear. The rules aren't the problem; the rules are actually quite mechanical. The problem is volume: a busy trade finance desk is running this exact check, document by document, field by field, dozens of times a day, against a five-day clock.
UCP 600 doesn't work alone. International Standard Banking Practice (ISBP), currently the 2023 edition, is the companion interpretive guide examiners actually use day to day. Where UCP 600 states the rule, ISBP shows how it applies to a specific document type, exactly how a bill of lading should be signed, what counts as an acceptable description of goods on an invoice, what an insurance certificate needs to state. An AI agent doing document examination needs both texts, not just the UCP 600 articles in isolation, because ISBP is where most of the practical, document-specific detail actually lives.
Most discrepancies fall into a handful of recurring categories, and most of them are avoidable with careful, consistent cross-checking, which is exactly what makes this workflow a strong fit for an AI agent rather than a reason to worry it's too nuanced to automate.
Discrepancy type | What it looks like | How an AI agent catches it |
|---|---|---|
Date logic | Shipment date after the credit's expiry, or presentation made after the standard 21-day window from shipment | Cross-checks every date field against the credit's own shipment and expiry dates and flags anything past the deadline |
Description-of-goods mismatch | Invoice wording for the goods doesn't match the credit's exact description | Runs a term-by-term comparison between the credit and every document referencing the goods, not just a keyword match |
Missing or incorrect endorsement | Bill of lading not properly endorsed for negotiation or transfer | Checks the required endorsement type against the credit's transport terms and flags any absence |
Quantity or amount discrepancies | Invoice amount exceeds the credit's value, or a partial shipment isn't permitted under its terms | Checks invoice totals and shipped quantity against the credit's value and any stated tolerance |
Missing required document | Certificate of origin, inspection certificate, or insurance certificate omitted from the presentation | Maintains a checklist derived from the credit's own document list and flags anything missing before presentation |
Cross-document inconsistency | The same shipment described with conflicting details across the invoice, packing list, and bill of lading | Reconciles shared fields, container numbers, weights, shipping marks, across every document in the set |
Trade finance screening is a wider surface than the customer screening covered in most KYC programs. Every underlying trade transaction involves foreign counterparties by definition, and most OFAC country-based sanctions programs are built specifically around restricting trade, which means the screening obligation extends past the buyer and seller to the vessel and the port carrying the goods. OFAC publishes separate lists of sanctioned vessels and aircraft precisely because a ship, not just a company, can be the sanctioned party in a shipment.
This isn't a theoretical risk. OFAC has continued to designate vessels tied to sanctions-evasion networks throughout 2026, including a February 2026 action naming more than 30 individuals, entities, and vessels connected to Iran's shadow fleet. A trade finance AI employee screening a shipment needs to check the vessel name and, where relevant, the port of loading and discharge against these lists, not just the named parties on the credit.
Shipments involving dual-use goods, items with both civilian and military applications, add a further layer. These can trigger a separate export-control review from the State or Commerce Department on top of the standard OFAC screen, and a bank financing that shipment needs evidence that both checks happened, not just one. For the deeper mechanics of how an AI agent handles sanctions-list matching and investigation more broadly, see our guide to AI agents for financial crime investigations.
The right split of labor here follows directly from UCP 600 itself. Article 16 requires the bank to refuse discrepant documents unless the applicant waives the discrepancy, and that waiver decision, along with any judgment call at the margins of what counts as a match, has to stay with a named trade-finance officer. An AI agent's job is to surface every discrepancy with the specific evidence behind it, not to make the final accept-or-refuse call on the bank's behalf.
In Zamp's model, that split is encoded directly in an Agent Operating Procedure, a document the trade-finance desk owns and edits itself: which discrepancy types the agent flags automatically, what evidence it has to attach for each one, and which categories always route to a person regardless of confidence. That's a meaningfully different posture from a black-box model scoring a presentation and returning pass or fail. It also means the AOP can be adjusted the day a bank changes its own tolerance policy, without an engineering ticket.
Every finding also needs a decision record a correspondent bank or an examiner could actually use: which document was checked, what the discrepancy was, and which UCP 600 article or ISBP practice it's based on, not just a generic "discrepancy found" flag. That level of traceability is what turns an AI agent's output from a suggestion into something a bank can stand behind when a correspondent bank pushes back on a refusal.
Trade finance data is cross-border by nature, so where the AI system runs and where the underlying transaction data lives matters more here than in a purely domestic workflow. A deployment that can run on-prem, in a dedicated multi-tenant instance, or fully inside the bank's own cloud environment gives a compliance team an answer to data-residency questions that a single fixed SaaS region can't.
Document format is the other practical consideration. MLETR-based legislation enabling electronic bills of lading is now in place in more than a dozen jurisdictions, including the UK, France, Singapore, and China, and the ICC's Digital Standards Initiative keeps pushing adoption further. But 2026 is still a hybrid period: an electronic bill of lading that's fully valid in one jurisdiction may still need to be printed to paper at a port that hasn't adopted MLETR. A trade-finance AI employee needs to reliably ingest and examine both formats, structured electronic records and scanned paper documents, since assuming digital-only will miss a meaningful share of live presentations for the next several years.
As a bank runs more letters of credit through this process, the accumulated corrections, the edge cases a trade-finance officer waived or refused, and why, feed back into the Company Brain: a growing, reviewable record of how the desk actually handles judgment calls, so the next presentation benefits from what the last one taught the system, instead of every case starting from a blank read of UCP 600.
For the broader regulatory picture across banking, KYC, sanctions screening, and chargebacks, see our guide to . For how AI agents handle sanctions investigation and evidence assembly at a deeper level, see .
