AI use case

AI examination of trade documents under letters of credit and collections

AI that reads the full document presentation under a letter of credit or collection (bill of lading, commercial invoice, packing list, certificates), extracts and cross checks the data, tests it against the instructions and the ICC rules (for letters of credit, the credit terms, UCP 600 and ISBP), and lists discrepancies by severity with the rule cited, so qualified examiners focus on the genuine exceptions.

By Len Debets · Last verified 27 September 2026 · 3 public deployments

USD 540,000 to USD 2.7 million
Indicative value per year
A bank examining 20,000 documentary credit presentations a year. Worked example, see how it is calculated.

What problem does it solve?

Documentary trade still runs on paper. For every presentation under a letter of credit, a trained examiner reads each document, compares names, dates, quantities, amounts, ports and goods descriptions across them and against the credit, and applies a large body of international practice to decide whether the presentation complies. A missed discrepancy can leave the bank paying against documents its client may refuse to reimburse; an unnecessary one delays the client's money.

RMB's head of trade describes the checking of numerous unstructured trade documents as manual and extremely time consuming, and Standard Bank Group's head of trade presents automation as a way to minimise repetitive tasks. Microsoft notes that traditional OCR and template based systems can struggle when layouts change or data is missing. Modern document AI plus a rules engine can do the extraction and the mechanical cross checks. In our view the main gain is examiner time freed for the genuinely ambiguous cases, as long as the contractual judgement on those stays with a qualified examiner.

How does it work?

  1. Ingest the presentation. Scanned and digital documents are classified by type and read with OCR and document AI, whatever their layout.
  2. Extract and normalise. Parties, amounts, currencies, dates, ports, goods descriptions, marks and quantities are extracted and normalised.
  3. Cross check. Data is compared across documents and against the credit terms (the MT700 fields), for example amount and currency, shipment dates, and consistency of goods descriptions.
  4. Apply the rules. A rules engine mapped to UCP 600 and ISBP tests each finding, and a language model helps with free text comparisons such as whether two goods descriptions conflict.
  5. Report discrepancies. Findings are listed by severity with the rule or credit clause cited; clean presentations go to a lighter review, exceptions to a qualified examiner who decides.
Audience
Back office
Autonomy
Supervised agent
Adoption
Early adopters
Channels
Internal tools, API and system to system

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

No public deployment has disclosed a measurable outcome yet.

Value drivers: Speed and cycle time, Employee productivity, Risk and loss reduction, Lower cost to serve.

Indicative value

A bank examining 20,000 documentary credit presentations a year

USD 540,000 to USD 2.7 million

Examiner time released, valued at loaded cost per year

How this is calculated

Formula: presentations * hoursPerPresentation * reduction * hourlyCost. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Presentations examined per year presentations, presentations per year20,00020,000The reference bank.
Examiner hours per presentation hoursPerPresentation, hours per presentation1.53Editorial assumption, replace with your own time study. Complex presentations take much longer.
Share of examiner time saved reduction, fraction of time0.30.5Editorial assumption, replace with your own pilot results; banks on this page have not published measured figures.
Loaded cost of an examiner hour hourlyCost, USD per hour6090Editorial assumption, replace with your own loaded cost.

What it leaves out: Values examiner time only. It leaves out the cost of the platform and integration, faster payment for clients, fewer missed discrepancies and the value of scaling without hiring scarce specialists.

Who already uses it?

3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

Rand Merchant Bank

South Africa · Banking · 2022

ProductionGrade B

RMB (Rand Merchant Bank), which describes itself as a leading African corporate and investment bank, announced that it had gone live on Traydstream's AI enabled trade finance platform. The platform can digitise documents related to letters of credit, collections and open account transactions for automated document checking, clause matching and rules validation with machine learning and OCR; the releases do not say which of these RMB uses it for. RMB's head of trade described the checking of numerous unstructured trade documents as manual and extremely time consuming. No outcome figures are published.

No outcome disclosed.

ANZ, HSBC and Lloyds Banking Group

Global · Banking · 2025

AnnouncedGrade C

Microsoft built a proof of concept with ANZ, HSBC and Lloyds, shown at Sibos 2025, in which an AI agent embedded in a corporate's ERP parses an incoming MT700 letter of credit, cross checks it against invoice and shipping data, flags discrepancies such as currency and amount, and sends structured data aligned to the ICC Key Trade Documents and Data Elements to the bank. The same agent answers treasury questions about compliance with the credit terms, and Microsoft says such agents can help flag references to sanctioned entities or ambiguous dual use goods descriptions. It is a demonstration, not a live service.

No outcome disclosed.

Stanbic Bank Uganda

Uganda · Banking · 2021

PilotGrade C

Stanbic Bank Uganda, part of Standard Bank Group, signed an agreement to implement Traydstream's platform to digitise the manual vetting of letters of credit for discrepancies, after trade document processing on it over the previous few months. The platform digitises the documents, checks them against trade rules and adds an aggregated compliance module; the group presented it as faster processing with more thorough trade checks and more transparent transactions. No figures are published.

No outcome disclosed.

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • A library of past presentations with the examiners' decisions, for testing
  • The bank's discrepancy taxonomy and severity rules
  • Current credit terms from the trade system in structured form

Systems to integrate

  • Trade finance processing system
  • SWIFT messaging (MT700 and related)
  • Document scanning and management
  • Trade screening, so compliance checks run on the same extracted data

Complexity: High

Document variety and quality are the main difficulty, followed by encoding UCP and ISBP practice as testable rules and integrating with the trade processing system and SWIFT messages.

  1. 1

    Build a gold standard set

    Collect past presentations with the examiners' final findings, including disputed ones, to measure extraction accuracy and discrepancy recall before go live.

  2. 2

    Separate mechanical from judgement checks

    Automate the mechanical checks (amounts, dates, names, consistency) first and route judgement calls, such as whether a variation is a discrepancy, to examiners with the evidence.

  3. 3

    Version the rules

    Keep every rule mapped to its UCP, ISBP or credit clause source, versioned, so a past decision can be reproduced with the rules in force at the time.

  4. 4

    Run in shadow mode

    Let the system check live presentations in parallel with examiners for a period and compare findings before changing the workflow.

  5. 5

    Share the extraction with compliance

    Feed the extracted data to trade screening so compliance and examination work from one version of the facts.

Guardrails

  • A qualified examiner decides on every discrepancy and signs off every refusal notice
  • Every finding cites the document, field and rule or credit clause behind it
  • Rule versions and model versions are logged per presentation
  • Low confidence extraction is shown as such and routed to manual review

KPIs to instrument

  • Examination time per presentation, clean and with discrepancies
  • Discrepancy recall and precision against examiner findings
  • Extraction accuracy per field and document type
  • Share of presentations cleared with lighter review
  • Refusals later disputed or overturned

Human in the loop

Examiners review every exception and decide on ambiguous discrepancies. Clean presentations get a lighter human review until error rates are proven low, and a sample of automatically cleared presentations is re examined every month.

Common failure modes

Missed discrepancy on a clean looking presentation
The system misses a subtle inconsistency and the presentation is waved through. Sample cleared presentations and track recall on the gold standard set.
Discrepancy noise
Too many trivial findings and examiners start ignoring them. Tune severity and suppress findings that practice treats as non discrepant.
Poor scans
Low quality images break extraction. Detect image quality and route poor scans to manual review.

What are the risks and rules?

EU AI Act

Minimal risk

Checking trade documents for compliance with credit terms is not listed in Annex III and does not decide about natural persons. AI literacy duties under Article 4 apply, and the process falls under the bank's operational resilience and model governance.

Guidance

  • ICC trade finance rules and standards (International Chamber of Commerce, Global). The ICC publishes UCP 600, ISBP and URC 522 (collections), the rule base the checks are mapped to; the examiner's judgement under them stays with people.
  • MAS Guidelines for Artificial Intelligence (AI) Risk Management (Monetary Authority of Singapore, Asia Pacific). Consultation paper of November 2025 proposing supervisory expectations for all financial institutions on AI inventories, risk materiality, evaluation and testing, human oversight and monitoring.

Controls to put in place

  • Model and rule inventory with owners, versions and validation results
  • Examiner sign off recorded per presentation
  • Monthly sampling of automatically cleared presentations
  • Retention of documents, findings and rule versions for the statutory period

Frequently asked questions

Which banks use AI to check trade documents?
RMB (Rand Merchant Bank) announced in September 2022 that it had gone live on Traydstream's AI enabled trade finance platform, which automates trade document checking, and Stanbic Bank Uganda (Standard Bank Group) signed an agreement in 2021 to implement the same platform after months of trade document processing on it. On the corporate side, ANZ, HSBC and Lloyds built a proof of concept with Microsoft in which an AI agent in the company's ERP cross checks a letter of credit against invoice and shipping data before sending structured data to the bank.
Does AI decide whether a presentation complies?
It should not decide ambiguous cases. It extracts, cross checks and lists discrepancies with the rule cited; a qualified examiner decides and signs off, because the bank stays responsible for honouring or refusing the presentation whatever tool it uses.
How fast is automated checking?
The banks on this page have not published measured figures. The RMB and Traydstream announcements speak of faster processing and improved turnaround times without numbers, so measure examination time on your own presentations in shadow mode.

How to cite this page

Blits.ai AI Use Case Library, "AI examination of trade documents under letters of credit and collections", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/trade-document-examination. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 27 September 2026: First published

Related use cases

Banking

AI screening of trade finance transactions for trade based money laundering

AI that screens every trade finance transaction for financial crime risk: it checks parties, vessels and ports against sanctions and watchlists, tests goods descriptions against dual use and controlled goods lists, compares unit prices with benchmarks for over or under invoicing, and reads trade documents and messages for laundering red flags, then prepares a case narrative for a human investigator.

Deployments
3 public, best grade C
Autonomy
Supervised agent
Cross industryGovernment and public sector

AI document intelligence for unstructured forms and documents

AI that takes documents in any format, such as scanned forms, PDFs, photos, emails and handwritten notes, splits and classifies them, extracts the required fields with a confidence score, validates them against business rules and source systems, and sends only the uncertain cases to a person before the data enters the downstream process.

Deployments
5 public, best grade B
Reported accuracy
at least 90%
Ancine, vendor claim
Cross industryBanking

AI for inbound correspondence triage and routing

AI that sorts inbound correspondence before anyone answers it: it takes every inbound letter, email, upload and secure message into one intake, identifies what it is, extracts the key fields, links it to the right customer and account, sets priority and routes it to the right team or workflow, replacing the manual sorting desk.

Deployments
6 public, best grade B
Reported accuracy
91%
Travelers, vendor claim
BankingPayments and cards

AI for payment investigations and exceptions

AI that works the payments that fall out of straight through processing: it reads the failure, repairs or enriches the message, drafts the ISO 20022 or SWIFT investigation, chases the counterparty bank and proposes a return, recall or correction, while an operator approves anything that moves money.

Deployments
2 public, best grade B
Reported automation rate
at least 10%
BNY, organization claim