US Special Cargo AI Intake: Preserve the Uncertainty in the Dimensions
A useful extraction can leave a field unresolved
An AI assistant preparing a US special-cargo enquiry should preserve what the source actually says, including missing units, provisional dimensions and conflicting configurations. A completed form is not necessarily a better form if the system fills gaps with plausible guesses. The useful product is an enquiry that a person can review efficiently and send with confidence about its factual basis.
NIST's generative-AI profile identifies confidently presented false output as a risk. Source: AI 600-1.
YC's startup advice emphasizes building for users and talking to them. Source: essential startup advice.
For a maritime product team, those ideas support a practical approach: test a narrow intake task with real users, measure whether it improves their decisions, and retain uncertainty where the evidence is incomplete. The design proposed here is an original workflow concept. It is not a claim that any named platform already performs these functions or that AI can approve cargo for carriage.
Preserve the source value before normalizing it
For every extracted dimension, keep the original text and the source location alongside any normalized value. The reviewer should be able to see whether the number came from a packing list, a product drawing or an email estimate. A normalized figure without its context may conceal a mismatch between the machine and its shipping package.
Units deserve explicit treatment. If a document states inches, a system can produce a separate converted value using a deterministic conversion, while preserving the original. If the unit is missing, the system should flag that absence. A familiar machine model or a likely range is not enough to establish the intended unit for a shipment enquiry.
The same principle applies to the order of dimensions. A source may provide three numbers without defining their sequence. The assistant should not silently label them length, width and height because that convention is common elsewhere. It can prepare a targeted question for the supplier and mark the fields pending confirmation.
Keep cargo identity connected to the extracted values. A document package may include a main unit, an accessory and an earlier design. The largest number in the package is not automatically the overall shipping dimension. An effective intake system helps the reviewer identify which item and configuration each value describes.
Test conflicts rather than only clean documents
Consider a hypothetical US machinery enquiry with a drawing showing a width of 96 inches and a later email saying the shipping width is approximately 2.5 meters. A deterministic conversion of 96 inches is 2.4384 meters. The two values differ by 0.0616 meters, or 61.6 millimeters. These are invented inputs for testing, not a statement about acceptable cargo tolerances.
The system should not average the figures or choose the later one without context. The drawing may describe the bare machine while the email describes a package, or one source may be incorrect. The correct output is a visible discrepancy, the two source references and a request to confirm the shipping configuration.
A test case should also include an apparently easy situation in which a number is absent. The assistant should demonstrate that it can leave the field unresolved instead of importing a value from a similar item. This is valuable behavior even though it reduces the percentage of automatically completed fields.
Include revised documents with unchanged filenames. The system should help the reviewer distinguish the content actually supplied, rather than assuming a familiar filename always represents the same configuration. Also test documents containing both metric and imperial figures so the assistant does not count them as two separate pieces of cargo.
Measure the quality of the human handoff
Evaluate more than extraction speed. Count incorrect values that reach the reviewer, conflicts that are missed and fields that are completed without support. Examine the time a reviewer needs to verify the output and the quality of the follow-up questions generated. A fast draft that forces a specialist to recheck every field may provide little practical benefit.
Separate harmless formatting differences from errors that change the enquiry. An extra space in a product description is different from a missing unit or an incorrect piece count. The evaluation should reflect the commercial task while avoiding unsupported claims about safety performance. Technical cargo acceptance remains outside the intake tool's authority.
Use a representative set of actual, appropriately authorized documents for the pilot, with sensitive information handled under the organization's rules. Include difficult but ordinary examples from the users' work, such as supplier revisions and mixed attachments. A demonstration using only tidy sample forms does not establish performance on the real intake queue.
Ask users whether the output makes the next action clearer. The best improvement may be a concise list of unresolved questions rather than a fully formatted carrier enquiry. That insight is consistent with a constructive YC-style learning approach: solve the user's immediate problem, observe the result and refine the workflow before expanding its scope.
Keep sending authority separate from drafting ability
An assistant can prepare an enquiry without being authorized to send it. Define who confirms the cargo identity, configuration and recipients before external transmission. A draft based on incomplete information should retain its status until the appropriate person approves the content for its intended use.
Likewise, a carrier's reply should not automatically become an instruction to book. Quotation, feasibility assessment and accepted booking can represent different stages. The proposed workflow should preserve the source and meaning of the response, and leave commercial commitments with the person authorized to make them.
Zeaclub's interest in this direction can be framed positively and concretely: help maritime teams turn fragmented cargo information into better questions and more comparable offers. The founder's stated inspiration from YC is compatible with testing that narrow problem through user feedback; it does not imply YC affiliation or a deployed autonomous approval system.
Bring Zeaclub a representative special-cargo intake problem to discuss a bounded pilot. Success should mean fewer unsupported assumptions and a clearer human decision, with source evidence retained from the first dimension through the final reviewed enquiry.
Sources
- NIST: Generative Artificial Intelligence Profile, AI 600-1, checked October 24, 2026; confabulation context only.
- Y Combinator: essential startup advice, checked October 24, 2026; user-learning context only.
Related articles
- 0992: US Project Cargo Dimensions Sheet: Prevent Unit and Orientation Errors
- 0691: A Houston Project-Cargo Desk Needs a Piece-Level Software Demo
- 0567: US Out-of-Gauge Container Quotes: Make the Cargo Envelope Unambiguous


