A score you can trust. No custom AI to train.
Every document gets a confidence score from a general model you never have to train — then dozens of freight-native checks (prefix, route, HS, weight math) cross-examine each field against the rules air cargo runs on.
Corrections pile up in the low scores and all but vanish as it rises — so you set one cutoff. At or above it, documents clear on their own; below it, they route to review.
30 free scans to start · no card required · no model to train
Corrections vanish as confidence rises.
When you plot how often a reviewer later corrected a field against the score it was given, the shape is consistent: corrections cluster in the low bands and thin out toward the top. That's what lets you route on a single number — you set one cutoff. At or above it, documents clear on their own; below it, they route to review.
Illustrative figures. Correction rate = how often a reviewer later corrected documents in each score band — example numbers, not a guarantee. Everything inside the box auto-clears; everything to its left goes to review.
One dial, not a rules engine.
The cutoff is the only thing you tune. It trades review effort against how much clears on its own — and you can move it any time as you watch how your own documents behave.
Tighter control
Push the cutoff up and more fields route to review. Fewer documents auto-clear, but a person sees more of the borderline ones first.
More throughput
Bring the cutoff down and more documents auto-clear. Fewer reach a person, so the stream keeps moving with less hands-on review.
A sensible default
85 is where most workspaces start — a practical balance where the correction rate above the line has already thinned out in the illustrative shape above.
See it before you commit
Because the score is built the same way every time, you can look back at how a new cutoff would have routed your recent documents before you change it.
The number shown as “Popular” and the correction-rate figures above are illustrative examples to explain the mechanic — your own numbers depend on your documents. US-only at launch.
How the score behaves.
Do I have to train a model on my documents?
No. A general-purpose model reads the page — you never train it on your data — and then fixed, freight-native checks (MOD-7, prefix and route lookups, HS lookup, weight recompute, consolidation detection) cross-examine each field the same way for every document.
What actually feeds the confidence number?
Each check contributes to the confidence of the field it touches — a check digit that computes, a prefix that resolves, weights that reconcile. Those per-field results roll up into one document confidence, which is what you route on.
Are the correction-rate figures real numbers?
They're illustrative examples used to show the shape — corrections concentrated in the low bands, thinning as confidence rises. They are not a benchmark or a guarantee; your own figures depend on your documents.
Can I change the cutoff later?
Yes — it's a single dial. Raise it to send more fields to review, lower it to auto-clear more, and adjust any time as you watch how your own documents score.
Route on a score you can explain.
Start with 30 free scans and set your own cutoff. No card, and no model to train.
Related: Air waybill validation · Pricing