Machine learning for NMFC codes and freight class: what it does, and what density still decides

Freight class tells LTL carriers how costly a shipment is to move. There are 18 classes, from 50 to 500, based on four factors: density, stowability, handling and liability. All else equal, a lower class generally costs less to ship per pound.

An NMFC item is the commodity entry the class comes from. You find the item that describes your product, and the item gives the class, either as a single class or, for density-based items, a scale of classes by pounds per cubic foot. The NMFC is published by the National Motor Freight Traffic Association (NMFTA), and the official lookup is its ClassIT+ subscription.

Getting it wrong costs money. When a carrier inspects a shipment and decides it belongs in a different class, it bills the difference and may add a reclassification or reweigh fee.

Docket 2025-1 took effect on 19 July 2025. NMFTA describes the main changes as follows:

• Thousands of freight items were reclassified.
• Many commodities moved from commodity-based classes to a single density-based scale, for freight with no special handling, stowability or liability concerns.
• The density scale grew from 11 sub-provisions to 13, adding classes 50 and 55 for the densest freight.
• Over 2,000 items moved to full-scale density-based classification, and many generic headings were consolidated.

Two things follow for anyone automating freight class. Weight and dimensions now matter more than ever, so accurate measurement is the first job. And item numbers your systems saved before July 2025 may be cancelled or consolidated, so a lookup table built before the change needs checking.

Source: NMFTA’s 2025 changes. Check the current item in ClassIT+ before using historical classifications.

A shipper or 3PL usually has years of history: product descriptions next to the NMFC item and class that were actually used. That is labeled data, and a model trained on it learns how your team maps your own descriptions to items. Useful jobs for it:

• New SKUs. Suggest the NMFC item for each new product in the catalog before the first shipment, so the item number is on the bill of lading from day one.
• Messy descriptions. Supplier descriptions, abbreviations and part numbers vary. A model that reads meaning copes with different wording better than a keyword table.
• Many shippers, one team. A 3PL or broker sees thousands of descriptions from different customers. A model trained on its own past decisions gives every coordinator the same first suggestion.
• Checking what is already in the system. Run the model over existing SKU records and look at the rows where its answer differs from the saved item. Those are the records most worth a second look after the 2025 changes.

What it does not do: read a tape measure. The class for a density-based item still comes from the weight and dimensions of the actual handling unit, calculated against the item's density scale. A model can tell you the item is density-based; your scale and dimensioner give the class.

We have not benchmarked NMFC items. The NMFC is a subscription publication, and we do not have a public set of product descriptions labeled with NMFC items to test on, and we won't publish a number we haven't measured.

The closest task we have measured is HS code classification: a short product description in, one code out of thousands. In our HS code classification benchmark we trained on 5,621 U.S. Customs (CBP) rulings and tested on 1,167 later ones. The training rulings covered 2,014 different 10-digit codes, and 1,129 of those appeared only once.

The AI Classifier Model was at 0.9 confidence or higher on 6.7% of the test rulings, and 96.2% of those were right to all 10 digits.

Read these as a hard case, not a forecast. Each model saw only a one-line ruling subject, and many codes had a single training example. A shipper's own NMFC history is usually more repetitive: the same products ship again and again, and descriptions come from one catalog. That should help, but the only way to know is to test on your own records, which AMT does before you use the model.

Confidence is what makes it usable

Alongside overall accuracy, check how many answers the model is sure of, and how often those sure answers are right. Every answer from StayCharted AI Model Trainer (AMT) comes with a confidence from 0 to 1, and the test report shows, for each cutoff, the share of answers above it and how many of those were right. You pick the cutoff. Your workflow can use answers above the cutoff to populate the bill of lading. On Essentials and Business, uncertain answers go to the Review Queue for a person to check in ClassIT+. Corrections become training examples when you retrain and publish the next version. Confidence does not guarantee a correct item. Read more on confidence scores and human review.

Most teams end up using all three: the lookup table for known SKUs, a trained model for everything else, and a person with ClassIT+ for whatever the model is unsure of.

If you ship a few dozen products that rarely change, a lookup table maintained by one person is simpler and cheaper. Machine learning earns its place when new SKUs arrive every week, when descriptions come from many suppliers or customers, or when several people classify and give different answers for the same product.

1. Export your history. One row per product or bill of lading line: the description, and the NMFC item your team used. Add the sub and class columns if you keep them. A few dozen examples per item is a reasonable start; see how many examples you need.
2. Check the items against the 2025 changes. Rows filed under cancelled or consolidated items should be updated before training, or the model learns the old item.
3. Train and read the test report. Start with the AI Classifier Model, with one included on Free, and look at accuracy per item and the confidence table before you use it.
4. Fill new rows where they live. Fill a CSV or Excel export, use the Google Sheets add-on, call the API, or fill straight from a Snowflake or Databricks table.
5. Keep measuring. Weigh and measure the handling unit, apply the density scale, and use the Review Queue for uncertain model answers on Essentials and Business.

Start free · See plans · How AMT works