What problem does it solve?
Trade is a well known channel for moving criminal money and evading sanctions: goods can be over or under invoiced, misdeclared, traded through shell companies, or routed through restricted jurisdictions, sanctioned ports and vessels. The evidence is spread across letters of credit, invoices, bills of lading, SWIFT messages and vessel tracking data, much of it unstructured.
Much of this checking has been manual. United Bank Limited in Pakistan, for example, relied on manual processes for vessel screening, tracking and dual use goods identification before it selected an automated platform in 2020. Done by hand, the work means looking up vessels, reading goods descriptions against control lists and searching for counterparties one by one, which is slow and open to errors and delays. Regulators expect a risk based, documented process, and compliance gaps can lead to penalties.
- LexisNexis Risk Solutions states, citing an outside source, that trade based money laundering represents up to 80% of capital flight from developing nations.Targeting Trade-Based Money Laundering in APAC (2025)
How does it work?
- Extract. Document AI and message parsing pull parties, goods descriptions, quantities, prices, ports, vessels and dates from the letter of credit, trade documents and SWIFT messages.
- Screen names and routes. Parties, banks, vessels and ports are screened against sanctions and watchlists, and vessel history and tracking are checked for suspicious port calls or transponder gaps.
- Check the goods. Goods descriptions are classified against dual use and controlled goods lists, including vague or unusual descriptions that need a closer look.
- Test the economics. Unit prices and quantities are compared with benchmarks and the client's usual trade pattern to spot over or under invoicing and unusual routes.
- Prioritise and explain. Alerts are scored, duplicates and known false positives are suppressed, and a case narrative with the evidence is drafted for the investigator, who decides and, where needed, files a suspicious activity report.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Emerging
- 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: Risk and loss reduction, Compliance quality, Speed and cycle time, Employee productivity.
Indicative value
A trade bank handling 40,000 trade finance transactions a year
USD 26,667 to USD 341,333
Compliance review time released, valued at loaded cost per year
How this is calculated
Formula: transactions * alertRate * minutesPerAlert / 60 * reduction * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Trade finance transactions per year transactions, transactions per year | 40,000 | 40,000 | The reference bank. |
| Share of transactions that need a manual compliance review alertRate, fraction of transactions | 0.2 | 0.4 | Editorial assumption, replace with your own alert volumes. |
| Minutes of review per alerted transaction minutesPerAlert, minutes per alert | 20 | 40 | Editorial assumption, replace with your own time study. |
| Share of review time saved by extraction, suppression and drafted narratives reduction, fraction of time | 0.2 | 0.4 | Editorial assumption, replace with your own pilot results. |
| Loaded cost of a compliance analyst hour hourlyCost, USD per hour | 50 | 80 | Editorial assumption, replace with your own loaded cost. |
What it leaves out: Values review time only. Possible further value, such as fewer missed red flags and lower regulatory penalty risk, is not quantified here and is not yet shown by public evidence, nor are faster turnaround for clients or the cost of data feeds.
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.
ANZ, HSBC and Lloyds Banking Group
Global · Banking · 2025
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
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.
United Bank Limited
Pakistan · Banking · 2020
United Bank Limited in Pakistan selected an automated trade screening platform in 2020 to move its trade compliance checks from mostly manual work to automation and meet the Pakistan Single Window regulatory directive. One interface now runs customer sanctions screening and risk assesses the trading activity itself, including dual use goods identification, the countries involved and vessel history and tracking, with a full activity history for audits. The vendor's case study reports faster turnaround and fewer false positives but gives no figures, and it describes automated screening against sanctions, trade and vessel data, not AI.
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
- Sanctions, watchlist, vessel and port data kept current
- Dual use and controlled goods lists for the relevant jurisdictions
- Price benchmarks per commodity and a history of client trade patterns
- Labelled past alerts and investigation outcomes to tune thresholds
Systems to integrate
- Trade finance processing system
- SWIFT messaging
- Sanctions screening engine
- Vessel tracking and maritime data provider
- Financial crime case management
Complexity: High
Needs good trade document extraction, vessel and commodity data feeds, price benchmarks, and integration with the trade processing and case management systems, all under model risk and financial crime governance.
- 1
Map the red flags you must cover
Start from the typologies your regulators and industry bodies list (pricing anomalies, dual use goods, vessel and route risks, unusual parties) and map each to data and a check.
- 2
Automate extraction and list checks first
Begin with the checks that are still manual: data extraction from documents and vessel, port, party and goods screening. United Bank Limited, for example, replaced manual vessel screening, tracking and dual use goods identification with automated checks.
- 3
Add anomaly detection with explanations
Add price and pattern anomaly models only with clear reasons per alert, so investigators can act on them and examiners can follow them.
- 4
Draft, do not decide
Let the system draft the case narrative with evidence links; the investigator decides and signs any suspicious activity report.
- 5
Monitor as model metrics
Track list freshness, false positive rates and missed cases found later as model performance metrics with an owner.
Guardrails
- Humans decide on every suspicious activity report and on declining or exiting a transaction
- Screening lists and vessel data freshness monitored, with alerts on stale feeds
- Full data lineage and reasoning retained for every alert and decision
- Suppression rules for false positives are documented, approved and reviewed periodically
KPIs to instrument
- False positive rate and alerts per thousand transactions
- Time from presentation to compliance clearance
- True positives and escalations, including those found later by other means
- Freshness of screening lists and vessel data
- Analyst hours per alert
Human in the loop
Trade compliance analysts review every alert above threshold and every sanctions or dual use match. Investigators decide on escalation and suspicious activity reports. The financial crime function approves suppression rules and thresholds, and model risk validates the models.
Common failure modes
- Over suppression
- Tuning to cut false positives also hides real risk. Test suppression rules against past true positives and sample suppressed alerts.
- Vague goods descriptions slip through
- Generic descriptions ("machinery parts") evade keyword checks. Flag vague descriptions for review rather than passing them.
- Stale data
- A sanctions list or vessel feed stops updating unnoticed. Monitor feed timestamps and block clearance when data is stale.
What are the risks and rules?
EU AI Act
Minimal risk
Financial crime screening of trade transactions is not listed in Annex III. It still processes personal data of individual parties, so GDPR applies, and supervisors expect it to be governed like any financial crime model.
Rules that apply
Guidance
- MAS Guidelines for Artificial Intelligence (AI) Risk Management (Monetary Authority of Singapore, Asia Pacific). Consultation paper of 13 November 2025 (consultation closed 31 January 2026; no final Guidelines were found on the MAS consultation page when checked in September 2026) proposing supervisory expectations for AI oversight, inventories, life cycle controls, human oversight and monitoring, relevant for models that prioritise or suppress financial crime alerts.
- ICC trade finance rules and financial crime risk controls (International Chamber of Commerce, Global). ICC's trade finance hub links the documentary credit rules (UCP 600, eUCP) with its financial crime risk control guides on dual use goods, price checking and vessel checking, and with the Wolfsberg Group, ICC and BAFT Trade Finance Principles.
Controls to put in place
- Model inventory, validation and ongoing performance monitoring
- Documented red flag coverage mapped to typologies
- Audit trail of data, rules and model versions behind each alert
- Periodic independent testing of screening effectiveness
Frequently asked questions
- Can AI replace trade compliance analysts?
- No. It removes manual look ups and drafts the case, but humans decide on escalation and on every suspicious activity report. The gains reported so far are faster turnaround and fewer false positives, claimed by a vendor for automated screening without AI; any gain in detection should be measured in your own pilot.
- What does a real deployment look like?
- The clearest production example is automation, not AI: United Bank Limited in Pakistan automated vessel screening and tracking and dual use goods identification in one platform to meet the Pakistan Single Window directive, and its vendor's case study, which does not mention AI, reports faster turnaround and fewer false positives without figures. On the AI side, Microsoft, ANZ, HSBC and Lloyds showed a proof of concept at Sibos 2025 in which an AI agent parses letters of credit, and Microsoft says such agents can help flag references to sanctioned entities and ambiguous dual use descriptions. We found no bank that has published results for AI trade crime screening in production.
- Is this high risk under the EU AI Act?
- No, financial crime screening is not in Annex III. It still needs strong governance because supervisors treat these models as part of the bank's financial crime controls.
How to cite this page
Blits.ai AI Use Case Library, "AI screening of trade finance transactions for trade based money laundering", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/trade-finance-crime-screening. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published