A US food supplier's CFO should score the credit of each business customer on a model whose factors the credit team can list before granting longer payment terms, and should use machine learning only to check that model's scores.
Atradius, which sells credit insurance and debt collection, surveyed accounts receivable contacts at 240 US companies, 80 of them in agri-food, in mid 2025. In the report, published 17 September 2025, agri-food respondents said 40% of B2B invoices are overdue, and 53 percent of those firms expected the insolvency risk of their business customers to rise over the next 12 months. The 2026 edition, published 16 September 2026 from 662 businesses in Canada, Mexico and the United States on an updated panel, says comparisons with previous reports are not possible and carries no agri-food figures in its report or statistical appendix. Atradius's food industry report of 4 September 2026, after a line on larger US retailers, reports a negative impact on the credit risk situation of smaller food retailers.
Regulation B, under the Equal Credit Opportunity Act, defines credit to include the right to purchase property or services and defer payment therefor. For an extension of trade credit, a creditor must notify the applicant of its action within a reasonable time and give reasons for adverse action if the applicant asks in writing within 60 days of that notice. Refusing credit on substantially the terms requested in an application is adverse action unless the applicant uses or expressly accepts a counteroffer, and action on an account over its inactivity, default or delinquency is excluded. Where adverse action rests on a credit scoring system, the official interpretation says the reasons disclosed must relate only to those factors actually scored in the system.
A Bank of Italy working paper of December 2019, published as preliminary results, tested one-year forecasts of Italian non-financial firms defaulting on bank credit from 2011 to 2017. Adding credit behaviour indicators, such as delinquencies with a bank, raised the models' ability to rank future defaulters above sound firms by about 10 percentage points. Machine learning beat logistic regression, a standard statistical model, on financial ratios and firm characteristics alone; its advantage diminishes when high quality information, such as credit behavioral indicators obtained from the Credit Register, is also available, and becomes negligible when the dataset is small. Its small dataset was a tenth of about 300,000 firms' observations. In a credit allocation test, machine learning ratings implied lower credit losses than statistical models; the authors say machine learning may benchmark more transparent statistical models.
In the first month the credit manager exports two years of ledger for every customer on terms and marks each customer that reached 90 days past due or was written off. In the second, the credit manager and the FP&A head fit a logistic regression on days paid beyond terms, share of invoices disputed and use of the credit limit, holding back a fifth of customers to test it; the controller approves the factor list. In the third, a machine learning model trains on the same record, any customer the two models rank far apart goes to the credit manager, and the CFO decides each longer-terms request with its scored factors attached.
If the machine learning model ranks the held-back customers no better than the logistic regression, the CFO stops funding it.
The official interpretation adds that if an application passes a scoring system and a person refuses it, the reasons disclosed must relate to the factors that person reviewed; if the score leaves it in a gray band, the reasons must come from both.
