Credit Decisioning Model: Rules, Scorecards or ML?

A credit decisioning model is the logic that turns a new customer's data into a credit decision: approve, decline, refer to a person, and at what limit. B2B teams use three kinds, rules, scorecards, and machine learning, and most of the good programs combine them.

What Is a Credit Decisioning Model?

A credit decisioning model takes inputs such as bureau scores, payment history, financial ratios, and public records, applies a defined method, and returns a decision plus a credit limit. The method can be a list of if-then rules, a weighted points system called a scorecard, or a statistical model trained on past outcomes. The choice sets how fast you decide, how well you can explain the decision, and how much history you need.

This guide covers the model itself. For the software that runs it, see the credit decisioning platform guide and the credit decisioning software guide.

How Do Rules, Scorecards, and ML Models Differ?

A Worked Scorecard Example

This example uses assumed weights to show the mechanics. It is not a calibrated model, and your factors and cutoffs should come from your own loss history.

Decision bands: 75 or more approves at the tier limit, 55 to 74 approves at a reduced limit with a 60-day review, and under 55 goes to a person.

Take an applicant with 3 years in business (10 points), a PAYDEX of 82 (25), references that run 5 days beyond terms (20), a current ratio of 1.2 (10), and no public records (15). The score is 80, so it approves at the tier limit. A rules-only policy that says "decline anyone under 5 years in business" would have turned this customer away. That is the cost of rigid rules, and the reason a scorecard lets strong factors offset a weak one.

The PAYDEX bands follow D&B's published tiers: its PAYDEX runs from 1 to 100, with 0 to 49 marked high risk of late payment, 50 to 79 moderate, and 80 to 100 low. D&B states the score measures past payment performance based on the D&B Data Cloud. It describes history. It does not tell you what the customer's finances look like this month.

When Does Machine Learning Make Sense for B2B Credit?

ML earns its place when you have thousands of past accounts with clear outcomes, a firm definition of a bad outcome such as 90 days past due or a write-off, and a team that can monitor the model after launch. Most mid-market credit teams do not have that history in their own books. They can still use ML-based scores from outside providers as one input into a scorecard, which keeps the decision explainable.

A sound way to adopt ML is to run it in shadow mode next to your scorecard for two quarters. Record where the two disagree and what happened to those accounts. Promote it only if the disagreements favor the model.

For how AI fits the wider credit workflow, see AI in credit management, and for outside data that widens thin files, see alternative data for trade credit.

Do Credit Decisioning Models Have to Be Explainable?

For business credit, the answer depends on the applicant's size and your process. Regulation B, which implements the Equal Credit Opportunity Act, applies to business credit. For applicants with $1 million or less in gross revenues, the creditor must follow the notice rules that call for a statement of reasons with an adverse decision, with listed exceptions. For applicants over $1 million in gross revenues, and for trade credit, the creditor owes a written statement of reasons only if the applicant requests it in writing within 60 days (12 CFR 1002.9(a)(3)). Either way, you need to be able to say why a decision came out as it did. A rules list or scorecard makes that simple. A black-box ML score makes it hard. This is general information and not legal advice, so confirm your process with counsel.

What Should a Credit Team Actually Build?

A layered model works for most B2B teams:

1. Hard-stop rules first for conditions that no score should override, such as a sanctions match or an active bankruptcy.
2. A scorecard in the middle that sets approve, reduce, or refer, with weights recalibrated against your own losses at least once a year.
3. ML as a challenger once you have the history to test it.
4. A human on the edge cases, where judgment is worth the time.

The model is the smaller problem. Most programs run it once, at onboarding, and then freeze the inputs. A customer approved at a score of 80 in January can be a different risk by summer, and nothing in the model has changed because nothing re-ran it. Credit Pulse feeds the same decisioning logic with research agents that refresh the inputs continuously, so the score moves when the customer's situation moves. The same scoring also shows which accounts have earned a higher limit, which is how credit data supports growth instead of only limiting loss.

To see where this fits in a full program, read about credit management software and the credit application process.

Frequently Asked Questions

What is a credit decisioning model?

A credit decisioning model is the method that converts customer data into an approve, decline, or refer decision and a credit limit. It can be a set of rules, a points-based scorecard, or a machine learning model.

Which credit decisioning model is best for B2B?

Most B2B teams do best with hard-stop rules plus a scorecard. The scorecard is explainable and lets strong factors offset weak ones. Machine learning fits teams with large accounts history and the staff to monitor the model.

Do I need machine learning for credit decisions?

No. A well-built scorecard handles most mid-market volume. Add ML as a challenger model once you have enough past accounts with known outcomes to test it fairly.

Does Regulation B apply to business credit decisions?

Yes. Regulation B covers business credit, with different notice rules for applicants at or under $1 million in gross revenues and for those above it. Confirm the details with counsel.