Before AI copies a shipping invoice, check what the numbers mean.

The template already separates the quantities.

Page 16 of DHL Express’s 2026 Hong Kong guide includes a sample commercial invoice with separate fields for “Total Units” and “Number of Pieces”. It also distinguishes net and gross weight. The document gives those values different jobs. DHL also asks senders to supply both the invoice file and its data electronically. [1 · p16]

The same guide describes existing digital services, including MyDHL+ for shipment tasks and MyGTS for cross-border information. My proposal concerns the preparation and checking of a shipper’s source data before it enters a system. [1 · p18]

Now imagine an AI extraction gives a company one field called quantity. It fills that field with 120 from the invoice, then encounters 10 on a packing record. The source documents may be clear. The company’s chosen data structure has made them ambiguous.

That is the design decision I would examine first. A field that loses the unit of measurement can turn an accurate transcription into a misleading record. Reading the number correctly is only part of the job.

Every quantity needs its object and its source.

For this example, I would preserve four things: the value, the unit, what it refers to and the passage it came from. An invoice’s count of product units belongs to a product line. A count of cartons describes packaging. The relationship between them needs its own evidence.

Suppose a packing record states that ten cartons each hold twelve units of the same product. A fixed calculation can reconcile that with an invoice for 120 units. The AI may help locate and interpret the passages; the multiplication can be ordinary code.

If the record gives only the carton count, dividing the invoice total by ten would produce a plausible pack size. It would not establish how the goods were packed. I would keep that field unresolved and ask for the missing detail.

The same principle applies when a supplier uses “pieces” to mean individual products. A word alone should not decide which system field receives the value. The document type, product reference and packing information matter.

This is also where an interactive explanation earns its place. A reader can remove the packing detail and see why an apparently simple match becomes unresolved.

Change the evidence. Watch the result.

The fictional shipment contains one product, packed uniformly, with one carton treated as one shipping piece. Those are explicit assumptions for this demonstration. Mixed cartons, multiple products and pallets would need a richer record.

Give the agent a defined place in the process.

I would start with a bounded job: prepare a quantity-reconciliation record for one type of shipment. The system would receive approved source documents and keep each extracted value attached to its source passage. A reviewer could correct the interpretation without losing the original text.

The AI’s proposed role is document interpretation: locate the product lines, suggest what a quantity refers to and draft a clear request when information is missing. Fixed rules would validate the required fields and calculate the packing totals. A reviewed mapping would determine which value belongs in which destination field.

For a team using an ERP or a shipping platform, the next step is to verify that destination’s actual field definitions. A field labelled “pieces” may have a different meaning in another application. The receiving system’s contract needs to be explicit before anything is written to it.

I would first make the output an internal draft. If an integration were later added, it would need a defined write permission, a stable shipment reference and a record of whether an earlier attempt succeeded. Retrying a failed request should not create a second shipment.

The pause needs to be useful too. A reviewer should see the conflicting values, their sources and the precise unanswered question. If the invoice and packing record disagree, the workflow should ask which source needs correction. It should never choose the more convenient figure.

None of this assumes that an enterprise needs a new platform. The first implementation question is whether its current systems already offer the required extraction, validation and review controls, and where a specific gap remains.

Test the exceptions before connecting the output.

I would assemble a small, permissioned set of representative documents with the operations team. Each example needs an agreed interpretation: which product is being counted, how it is packed and which sources establish the relationship.

The test set should include missing pack sizes, mixed products, conflicting references and unclear scans. Record whether the proposed workflow preserves the right meaning, asks for the right missing evidence and leaves unresolved cases open.

Review effort belongs in the evaluation. Count the corrections a person must make and measure the time needed to inspect an exception. A fast extraction can still leave a slow review if nobody can find the source passage.

Those would be measured results from a future pilot. This page demonstrates the design of one check. It does not report a model accuracy score, a reduction in errors or time saved.

A practical enterprise question inside a trade document.

Hong Kong’s current AI discussion includes how businesses connect AI to everyday operations. A July 2026 Legislative Council exchange specifically addressed support for import and export SMEs adopting AI. HKT’s August commentary on its business survey also discussed agentic workflows, integration and access controls. [2] [3]

There is already a local example of AI applied to trade-document matching. Hong Kong’s Census and Statistics Department describes using language processing to link declarations and cargo manifests that its earlier rules could not pair. That work concerns record linkage. This demonstration examines the meaning of quantities within a proposed shipment record. [4]

Those are market signals, rather than search-volume measurements. They make this a relevant problem to examine; they do not establish demand for a particular solution.

For Enzwa, I want the work to be specific enough that a trading, logistics or operations team can recognise the decision. Bring one repeated document task, identify the meaning that gets lost between systems, and make the proposed check visible. That is a useful starting point for an AI workflow brief.