Sales is pushing to get the shipment out. The logistics team has already given AI the product name, materials, and intended use to organize, and the system has recommended a tariff classification code. The hardest question comes next: Who confirms the recommendation and takes responsibility for the result? If the product description omits a critical specification or the wrong code is selected, the company may face document revisions, customs clearance delays, additional costs, or penalties. Labeling the result as an AI recommendation does not transfer that responsibility to AI.
The approaches taken by FedEx and DHL, as reported by technology media, illustrate how this responsibility should be divided. FedEx has incorporated product descriptions and U.S. Harmonized Tariff Schedule (HTS) code lookup into its existing shipping process. DHL’s MyGTS integrates regulatory information, document templates, Landed Cost calculations, and Free Trade Agreement (FTA) information. These tools help businesses prepare shipping data, but shippers remain responsible for verifying the declared information. We believe AI can assist with customs documentation, but companies must still understand the basis for every recommendation, establish approval checkpoints, and retain records that can be reviewed later.
Let AI prepare the data, but do not let it submit declarations directly
We recommend using AI first to organize the product name, materials, intended use, model number, and destination. Based on the information provided by the company, it can then identify potentially applicable classification codes and regulatory requirements. The person handling the shipment can compare each item against the product master data, previous customs declarations, and the requirements of each market. Once AI has made a recommendation, a person must still make the final decision. The company must determine whether the product’s materials, intended use, or specifications affect its classification and whether previous declarations apply to the current shipment.
Companies can begin by mapping their existing process and identifying at least four checkpoints: product description, classification code, destination requirements, and declaration documents. Each checkpoint should specify who is responsible for verification, what information must be available before the shipment can proceed, and which circumstances require the case to be returned for additional information. For example, if the product description does not identify the primary material, the AI Agent should not be allowed to fill it in independently. If two candidate classification codes are identified, the decision should be referred to someone familiar with both the product and customs rules. During the initial implementation, companies can start with one standardized product category and a single export market. They should confirm that all four checkpoints work properly before attempting to automate the entire process.
AI must explain the basis for its recommendation and stop when data is insufficient
Companies should first check whether the information provided to the AI Agent is clear. If the product master data contains only an internal product name and item number, with no details about materials, intended use, composition, or specifications, the person handling the shipment will have difficulty verifying the result even if AI produces a seemingly complete product description. Previous customs declarations should not be applied directly to a new shipment if they do not identify the relevant market, product version, and period of applicability.
This information can be managed in three categories. Product master data explains “what this is.” Classification rules and historical customs documents show “how it was classified before.” Destination requirements and information from logistics providers confirm “what this shipment must comply with.” Every source should identify the applicable market, product scope, version date, and responsible department. The AI Agent should also cite the sources supporting its recommendation.
If no verifiable basis can be found, different sources conflict, or current regulations do not cover the product, the system should not provide a conclusion that can be applied directly. It should refer the case for human review. Before implementation, companies can audit the data for their most frequently shipped products, fill in missing information about materials, intended use, and versions, and then begin connecting the process to declaration documents.
Human approval cannot be reduced to an “Approve” button
If approvers can see only the code recommended by AI, but not the product data, cited regulations, or alternative candidates, clicking “Approve” effectively requires them to accept full responsibility without sufficient information. At a minimum, the approval screen should display the shipment’s critical fields, the data sources used by AI, the recommendation, and any unresolved questions. It should also allow approvers to return the case for additional information or change the determination.
Companies should also assign approvers according to shipment risk. If a shipment contains a new product, involves a product being exported to a market for the first time, has conflicting information, or carries a relatively high value, the company can assign someone with expertise in classification and regulations to review it. Even for repeat shipments with complete data, a designated person must approve the release in accordance with company policy.
EgentWrX can connect multiple tasks into a workflow and require human approval before handoff. After the AI Agent completes data preparation and the initial lookup, the workflow pauses and moves to the next stage only after the responsible person has approved it. When companies adopt this type of mechanism, they should write the release criteria directly into the task rather than mentioning them only during training.
Before implementation, the customs, logistics, product, and internal control teams should jointly review the four checkpoints and assign a primary approver and backup approver for each stage. This prevents the approval process from being bypassed simply to meet a shipping deadline when the usual person in charge is absent.
When a classification is disputed, companies must know what information was reviewed and who made the decision
The company remains responsible for its customs declarations, so retaining only the final submitted declaration is not enough. If customs authorities, logistics providers, or internal auditors question a product classification, the company must be able to identify which version of the product data was used, what supporting information the AI Agent found, what it recommended, which fields were changed by the person handling the shipment, and who approved the decision at what time.
These records help compliance and internal control personnel reconstruct the decision process, determine whether the problem originated in the product master data, classification rules, or approval criteria, and correct the relevant data or conditions accordingly.
EgentWrX provides searchable and exportable audit logs and can verify whether records have been altered. Actions taken by administrators to download audit records are also logged. The implementation team can incorporate AI-assisted customs declarations into the company’s existing spot-check program by first defining the criteria for high-risk shipments and then periodically checking whether the data sources, AI Agent processing records, and human approval records form a complete chain.
Acceptance testing should not focus solely on whether AI can produce a classification code. The team should also simulate a classification dispute and confirm that it can retrieve the data used at the time, the AI recommendation, the human changes, and the approval records.
When implementing AI-assisted customs declarations, companies should first ensure that every recommendation is supported by traceable data, every submission has a clearly accountable person, and the processing records can be retrieved if a dispute arises. Define the four checkpoints, the situations in which AI must stop, and the approver for each stage before deciding how many systems to integrate. This is more consistent with actual accountability requirements than pursuing full automation from the outset.
FAQ
Can an AI-recommended tariff classification code be used directly?
Using it directly is not recommended. The person handling the shipment should first verify the product’s materials, intended use, specifications, destination, and cited sources, then submit the recommendation to the company’s designated responsible person for confirmation. If multiple candidate codes are identified, the data conflicts, or critical fields are missing, the case should be referred for human determination.
How complete must product data be before AI can assist?
At a minimum, the information must allow the person handling the shipment to understand what the product is. It should include a consistent product name, item number, primary materials, intended use, and any necessary specifications, along with the data version and responsible department. If only an internal abbreviation is provided, the approver will have difficulty verifying whether the description generated by AI is correct.
At which stages should human approval be required?
Human approval should be required, at a minimum, before applying the classification code, confirming destination requirements, generating declaration documents, and formally submitting the declaration. Companies can adjust the approval level according to shipment risk, but every checkpoint should clearly specify the release criteria, approver, and process for handling cases returned for additional information.
What records should be retained when a classification is disputed?
Companies should retain the version of the product data used at the time, the cited regulations, the AI Agent’s recommendations and processing details, the fields changed by a person, the approver, and the approval time. These records must all be linked to the same shipment so that compliance or internal control personnel can reconstruct the decision process.
References
- How are logistics giants strengthening pre-shipment services as FedEx uses AI-assisted tariff classification and DHL integrates global trade information? — TechOrange’s overview of FedEx and DHL’s pre-shipment compliance tools and the boundaries of responsibility.
