The WTO’s AI for Trade programme describes tools that can identify commodities, build negotiating lists, simulate tariff reductions and compare provisions across trade agreements. Its September 15 masterclasses also demonstrate how officials can use such tools in policy analysis and negotiation. These are valuable capabilities for exploring scenarios. They are not a customs determination for a specific shipment.
That distinction can disappear quickly inside an exporting company. A market analyst produces an AI-assisted tariff scenario, a salesperson copies the lower figure into a quotation, and the customer reads it as the amount that will apply at import. The analysis may have been mathematically consistent while the commercial statement is still wrong.
A product label is not a customs classification
Sales names are designed for recognition. Classification depends on the product’s material, function, construction, use and sometimes its presentation with other items. One catalogue family can contain models or accessories that do not share the same treatment. An AI match based on a short product description can identify questions to investigate; it should not silently become the classification used for a transaction.
Origin adds another layer. Shipping from a Chinese warehouse does not by itself resolve how origin is determined for every product. Preferential treatment may depend on an agreement being in force, the relevant tariff line appearing in a schedule and the shipment satisfying documentation and origin requirements. If one of those facts is unknown, the scenario should preserve the gap rather than fill it with a confident rate.
Maintain three columns in the working file: baseline, policy scenario and currently executable position. The first two support planning and discussion. The third must point to current official material and the approved product-level assessment, with a retrieval date and owner. News, policy workshops and simulations belong in the scenario columns until the effective rule and product application are confirmed.
This approach also prevents an internal forecast from becoming a public claim. A possible reduction can inform sensitivity analysis without appearing in a product page, advertisement or customer email as if it already applies.
The quote still needs a responsibility boundary
Even a verified rate does not complete a customer quote. Incoterms, customs value, importer of record, currency, validity period, freight and brokerage responsibilities can materially change what each party pays. DDP, DAP and FOB are not interchangeable expressions of the same price.
Write the quotation so the buyer can see which costs are included, which party handles import formalities and what changes trigger recalculation. A change in destination, kit composition, quantity or delivery term should reopen the calculation. Do not hide a tariff scenario inside a single total and leave the buyer to assume that every border cost is fixed.
A practical workflow gives AI a useful but bounded role. It can assemble candidate classifications, retrieve relevant provisions, compare scenarios and identify missing inputs. A product owner confirms the item and configuration. A customs or trade specialist confirms the classification, origin and applicable rule. Sales uses only the approved version and states its validity.
Our article on testing AI quotation calculators with boundary examples explains why calculation rules need reproducible cases. Tariff work introduces an additional problem: a correct calculation can still be applied to the wrong product, origin or date. Both the formula and the applicability need evidence.
The best output from an AI tariff exercise is therefore not a single authoritative-looking number. It is a scenario sheet that exposes assumptions, unresolved fields, source dates and accountable reviewers. That sheet can improve market planning and prepare better questions for specialists. A customer quote should change only after those questions have been resolved for the actual transaction.
Archive the scenario that sales actually used. If a later customs question arises, the team should be able to reconstruct the product description, assumed origin, destination, rule date and delivery term behind the price. A live dashboard that overwrites yesterday's inputs cannot provide that explanation. Versioned evidence protects both the buyer conversation and the next internal review.
The archive should identify the reviewer and approval time as well as the source files, so an old scenario cannot be mistaken for the current executable position.
Sources
World Trade Organization, AI for Trade, accessed September 15, 2026, including the masterclass information for that date.