Most trade banks now see a role for AI in credit risk, with over 85% naming risk analysis and fraud prevention
Back to the Trade Finance Guide
Ossiano Guides · Digital trade
A trade credit decision weighs the chance that a buyer will not pay on a short deal. Pooled default data, standard document fields and machine learning help lenders see more, and the research also shows where the gains stop.
Terms like this have a quick explainer. Tap or hover on them.
01 · The decision
A trade credit decision prices the expected loss on a short exposure
Before a lender finances a trade, it asks one question: how much could it lose?
The answer is called . It is the loss a lender plans for on a deal. It combines three parts. The first is the , the chance the other side does not pay. The second is , the share of the money lost if that happens, after any recoveries. The third is , the amount owed when the default happens. The ICC Trade Register 2025 uses expected loss, loss given default and exposure at default in its risk analysis (p. 26 and p. 35).
Multiply the three parts together and you get expected loss in dollars. Better data can sharpen each part. Figure 1 shows how much a different default estimate moves the result.
Figure 1 · Try it
How a better default estimate changes expected loss
Move the sliders for default probability, loss given default and exposure, or pick a worked example. The result shows expected loss in dollars.
Worked example default estimates
Expected loss as a share of exposure
-
Change against the 1.00% estimate
-
Expected loss
-
How it works: expected loss equals probability of default times loss given default times exposure at default.
Our Research Desk used the standard expected loss formula: PD x LGD x EAD. The inputs shown are for illustration and are not Ossiano pricing or risk estimates.
This calculator explains the concept only. Lenders estimate each input with their own models and data, and results vary by obligor, product and tenor.
With the other inputs fixed, halving the default estimate from 1.00% to 0.50% halves expected loss, from $4,000 to $2,000. Doubling it to 2.00% doubles expected loss to $8,000. That is why a sharper default estimate matters.
Trade finance starts from a strong base. The ICC Trade Register 2025 finds that "trade finance and export finance represent a low-risk asset class".
02 · Pooled data
ICC pools bank default data on trillions of dollars of trade finance
One bank sees only its own customers. Pooling data from many banks shows how trade finance behaves as a whole.
The ICC Trade Register 2025 covers about $1.2 trillion of 2024 exposures. That is around 13% of global trade finance flows. Over time, ICC's trade finance dataset has grown to span over $25.7 trillion in transactions, according to its Global Trade Intelligence Report 2026 page.
ICC measures default in two ways. An weights defaults by the dollar value of the deals that failed. An counts the customers who default, divided by all customers. An obligor is simply the party that owes the money. The two views can differ, because one large default weighs more by value than by count (ICC, p. 30).
The headline findings are public. The numerical default rates in the 2025 edition are not: ICC says the full data analysis pack is available for purchase, so this guide does not quote them.
03 · Standard data
21 of 36 key trade documents already have standard electronic versions
Data is easier to use when every document records the same fields in the same way.
On April 24, 2024, the ICC Digital Standards Initiative launched its framework, short for Key Trade Documents and Data Elements. It rests on an 18-month analysis covering all 36 key trade documents. It found that 21 of the 36 already have standardized electronic versions, according to ICC.
An electronic version of a document is an : data created, sent or stored by computer. When the fields follow one standard, a lender can compare one shipment's data with the next.
Central banks are testing what standard data makes possible. The BIS Innovation Hub and the Hong Kong Monetary Authority built Project Dynamo, completed in 2023. It was a prototype for SME financing that used digital trade tokens on a public blockchain. An , the digital form of the carrier's shipping document, served as a payment trigger. Our guide to trade finance platforms covers more of this work.
04 · Adoption
Banks see AI's first uses in risk analysis, fraud prevention and capacity
Most banks in the latest global survey expect AI to help them judge risk.
The Asian Development Bank's Global Trade Finance Gap Survey, published in December 2025, found that over 85% of bank respondents see potential use of AI for risk analysis and fraud prevention. Nearly 56% of respondents are applying AI to help find ways to increase trade financing capacity, meaning the amount of trade finance they can provide.
Banks also see new kinds of data entering credit decisions. Over 59% of bank respondents say supply chain traceability around carbon emissions could factor in risk and credit decisioning, the ADB reports.
Figure 2
How banks say they use AI and data in trade finance
Each card shows one survey measure and what it counts.
Source: Asian Development Bank, ADB Brief No. 378, ADB Global Trade Finance Gap Survey, December 2025.
05 · The evidence
Machine learning beat traditional models in a stress period, with limits
One well-known study tested whether new data and new methods predict defaults better. They did, but the edge was not the same for everyone.
means computer methods that learn patterns from past data. means information beyond financial statements and credit bureau files. BIS Working Paper 834, published on December 19, 2019, combined the two. It found that a model based on machine learning and non-traditional data was better able to predict losses and defaults than traditional models in a period of stress, according to the Bank for International Settlements.
The edge had limits. The advantage tended to decline for customers with a longer credit history, the same paper found. Where a long track record already exists, traditional data does much of the work.
Note the scope. The data came from a Chinese fintech firm's lending, not from trade credit. This guide treats it as nearby evidence, not as a trade finance result.
06 · Access
Better data is aimed at the SME rejection gap
Smaller firms are turned down most often, and data gaps are part of the reason.
The ADB survey reports a trade finance of 41% for , small and medium-sized businesses. The rate is 40% for large corporates and mid-caps, and 20% for multinationals, according to ADB Brief No. 378.
Figure 3 · Interactive
Trade finance rejection rates by company size, 2025
Tap a bar to see the share of requests rejected.
Source: Asian Development Bank, ADB Global Trade Finance Gap Survey, Brief No. 378, December 2025.
Customer checks are a large part of the picture. , or KYC, means identifying and verifying customers and their owners. adds an understanding of the relationship and ongoing monitoring. Just over 15% of respondents ranked KYC concerns first among the reasons requests are rejected, the ADB found. Our guide to know your customer in trade finance sets out the steps.
The unmet demand adds up. The , the demand for trade finance that providers do not meet, stands at $2.5 trillion in the 2025 survey, per the ADB. Our guide to the trade finance gap explains how it is measured.
07 · Ossiano view
Better data widens the view of a trade counterparty
OBSERVATION 01
Banks are already testing AI on risk
Over 85% of bank respondents in ADB's 2025 survey see potential use of AI for risk analysis and fraud prevention. Nearly 56% already apply AI to find ways to increase trade financing capacity.
OBSERVATION 02
Standard fields make data comparable
ICC found that 21 of 36 key trade documents already have standard electronic versions. Standard fields let a lender compare one shipment's data with the next.
OBSERVATION 03
The evidence has boundaries
In the BIS study, the edge for machine learning narrowed for customers with longer credit histories. Our reading is that the gain is largest where credit history is thinnest. In ADB's survey, SMEs face the highest rejection rate, at 41%.
Summary
Better data sharpens each part of expected loss, and the evidence shows where it helps most
A trade credit decision rests on expected loss: probability of default times loss given default times exposure. In the worked example, moving the default estimate from 1.00% to 2.00% doubles expected loss from $4,000 to $8,000 on a $1,000,000 exposure.
Pooled data, such as the ICC Trade Register's $1.2 trillion of 2024 exposures, shows how trade finance performs as a whole. Standard fields, already in place for 21 of 36 key trade documents, make data comparable. Over 85% of bank respondents see a use for AI in risk analysis and fraud prevention. The BIS evidence shows machine learning did better in a stress period, with an edge that declined for customers with a longer credit history.
Related guides: The trade finance gap; Buyer, country and performance risk; Know your customer in trade finance; How trade finance is priced; Fraud controls in trade finance; Trade finance platforms.
Instrument cards: Trade credit insurance; Open account; Payables finance; Factoring. Every term on this page is defined in the Trade Finance Glossary.
For questions on how data and default estimates apply to existing or planned trade finance relationships, contact the Ossiano Research Desk.
Sources
- Asian Development Bank, ADB Brief No. 378, ADB Global Trade Finance Gap Survey, December 2025. Supports: over 85% of bank respondents see potential use of AI for risk analysis and fraud prevention (p. 10); nearly 56% apply AI to help identify ways to increase trade financing capacity (p. 10); over 59% say carbon traceability could factor in risk and credit decisioning (p. 9); KYC concerns ranked first by just over 15% of respondents (p. 6); rejection rates of 41% (SMEs), 40% (large corporate and mid-cap) and 20% (multinationals); global trade finance gap of $2.5 trillion; Figures 2 and 3.
- International Chamber of Commerce (ICC), ICC Market Commentary, Trade Register 2025, October 2025. Supports: expected loss and loss given default (p. 26); exposure at default (p. 35); exposure-weighted and obligor-weighted default rates (p. 30); trade and export finance a low-risk asset class; about $1.2 trillion of 2024 exposures; around 13% of global trade finance flows; full data analysis pack available for purchase.
- International Chamber of Commerce, ICC Digital Standards Initiative launches complete framework for supply chain digitalisation, April 24, 2024. Supports: KTDDE launch; 18-month analysis covering all 36 key trade documents; 21 of 36 already have standardized electronic versions.
- Bank for International Settlements, How do machine learning and non-traditional data affect credit scoring? New evidence from a Chinese fintech firm (BIS Working Paper 834), December 19, 2019. Supports: a model based on machine learning and non-traditional data predicted losses and defaults better than traditional models in a period of stress; the advantage tends to decline with a longer credit history.
- International Chamber of Commerce, ICC Global Trade Intelligence Report 2026 (report page), September 17, 2026. Supports: ICC trade finance dataset spans over $25.7 trillion in transactions.
- Bank for International Settlements, BIS Innovation Hub, Project Dynamo, 2023. Supports: prototype with the Hong Kong Monetary Authority for SME financing using digital trade tokens on a public blockchain, with an electronic bill of lading as a payment trigger.
The real economy moves through Ossiano.
