AI Solutions · Service
AI for freight documentation and classification
The evaluation and cost layer, not another product comparison.
At a glance
What it is
“Extraction accuracy on a clean sample tells you almost nothing. The cost is in the worst tenth of the document flow.”
A principal, Kaelo AI Solutions
What’s included
Four areas of scopeDocument extraction, evaluated properly
Bills of lading, commercial invoices, packing lists and certificates arrive as scans, photographs and inconsistent PDFs. Extraction accuracy on a clean sample tells you very little; what matters is accuracy on the worst tenth of your real document flow, which is where the exceptions and the cost live.
Classification and the threshold question
Automated HS classification is useful above a confidence threshold and dangerous below it. Setting that threshold is a commercial decision, not a technical one: it depends on what a misclassification costs you in duty, delay and penalty exposure, and that varies by trade.
Human-in-the-loop, designed rather than bolted on
The realistic target is not full automation. It is routing the confident cases straight through and putting the uncertain ones in front of a licensed person with the reasoning attached, so the review is quick.
What it actually costs at volume
Inference cost per document at production volume, not at demo volume, plus the review time the exceptions consume. Most business cases in this space are built on the happy path and quietly omit the exception queue.
How the work runs
Document sample
A real sample including the difficult tail, not a curated set.
Baseline measurement
Current handling time, error rate and rework cost.
Evaluation harness
Scored against your documents, with the threshold argued explicitly.
Pilot with the loop in place
Run alongside the existing process before replacing it.
When to come to us
- 01 You process enough documents that handling time is a real cost line.
- 02 You can supply a genuine sample, including the messy tail.
- 03 You have licensed staff who can sit in the review loop.
What we do not do
- Providing customs classification as a service. We build and evaluate tooling; the classification decision stays with your licensed broker.
- Quoting accuracy figures from a vendor's marketing material.
- Promising a level of automation before we have seen your document tail.
The product layer is crowded. The evaluation layer is empty.
Search for AI in freight and you find vendor pages, marketplace listings and press releases. Neutral writing on how to evaluate any of it is owned by no agency and no consultancy.
Related reading
The golden dataset, the harness, and what the score has to mean. →
Cost per token at production volume, and the exception queue nobody prices. →
Tender design, carrier selection and cost baselining — as method, not as phases. →
The vertical practice: forwarders, brokers, carriers, 3PLs and relocation firms. →
Common questions
