New reciprocal tariffs imposed by President Donald Trump on April 2nd, 2025, came as a shock to many U.S. importers. The Wall Street Journal reported on April 9th, 2025 that “many companies weren’t financially prepared to absorb the additional cost of tariffs,” because none of them knew at the beginning of this year that tariffs of this magnitude were coming just four months later.

Business Impact of the New Reciprocal Tariffs

An analysis which Geoprise recently performed for one of our clients confirms that the 10% baseline tariff now in effect has already reduced their gross margins on certain imported goods by 40%, assuming that none of the extra costs are passed through to customers, and will further drive gross margins to a 1.5% loss when the suspension of Trump’s country-specific (Annex I) tariffs ends in July 2025.

Our analysis also found that all the enterprise resources planning (ERP) systems on the market today will require modification to accurately compute the new tariffs, and that no ERP vendors other than ourselves have rolled out the necessary modifications to their products. The Wall Street Journal adds that the calculation process has:

become much more complicated as Trump rapidly rolled out new tariffs—and, at times, rolled them back … a company that previously might have had only to calculate one duty rate might now have to figure out three or more levies depending on where the goods were made, whether the items contain steel or aluminum and whether the shipment is compliant with the trade pact known as the U.S.-Mexico-Canada Agreement [USMCA].

The urgency of quickly completing such modifications and incorporating them into business processes that calculate and compare landed costs of imported goods for accounting and supply chain planning purposes will presumably rise in direct proportion to the extent the tariffs are causing pain. Only the most agile of businesses will be able to do this quickly enough to minimize financial losses and seize new opportunities to gain competitive advantages.

The Promise of Artificial Intelligence

Considering recent hype, and in particular the ability to digest reams of global trade data from industry reports and official sources in near real time as well as generate program code automatically, one could hardly imagine a scenario in which artificial intelligence (AI) technology delivers greater potential time-to-live and productivity benefits than this.

Nucleus Research, a U.S. technology research and advisory firm, offers one such hypothesis that:

A full-suite, AI-driven ERP system goes beyond traditional systems of record by driving action through real-time insights … the ability of AI agents and generative AI to rapidly produce contingency plans can reduce response times to unexpected challenges by 40 to 60 percent, enabling organizations to swiftly address emerging issues and maintain operational continuity in the face of unforeseen disruptions. By investing in a modern ERP solution, organizations gain the ability to manage risk and maintain competitiveness in an evolving economic landscape.

So, we designed an experiment for proving this hypothesis.

Method of Investigation

We first performed an analysis of business requirements from official sources using the traditional Google search engine without Google’s “AI Overview” feature to locate the Web sites of the relevant government authorities. Our analysis identified fifteen business requirements as of April 15th, 2025:

• In the U.S., retrieve the applicable general (most favored nation, i.e. MFN) and special duty rates from a database using the date and 10-digit harmonized tariff schedule (HTS) code associated with the imported product
• In Thailand, retrieve the applicable general (MFN) and special duty rates from a database using the date and 8-digit HTS code based on the Association of Southeast Asian Nations (ASEAN) Harmonized Tariff Nomenclature (AHTN) associated with the imported product
• Retrieve and apply preferential rates set out in any applicable trade agreements (for e.g., USMCA when importing goods to the U.S.; ASEAN Free Trade Agreement, i.e. AFTA when importing goods to Thailand) from a database using the date, HTS code associated with the imported product, and country of origin
• In the U.S., retrieve and apply any anti-dumping, Section 301, or countervailing duties in addition to the duty rate using the date, HTS code associated with the imported product, and country of origin
• In the U.S., customs value includes the cost of the imported goods shown on an accompanying invoice but excludes insurance, freight, and adjustments
• In Thailand, customs value includes cost of the imported goods, insurance, and freight (CIF), plus royalty/license fees and the fair market value of goods and services supplied by the buyer free of charge or at reduced cost for use with the production and sale of the imported goods
• The U.S. specifies ad-valorem, specific or compound duty rates for each HTS code
• Thailand does not use compound duty rates; if a specific duty applies for an HTS code, apply whichever specific or ad-valorem rate yields the larger tax
• In Thailand, ad-valorem value-added tax (VAT) is imposed based on CIF value
• In Thailand, ad-valorem excise tax is imposed on certain product groups based on CIF value plus customs duty
• As per President Trump’s executive order of April 2nd, 2025, additional 10% baseline reciprocal tariff imposed upon all imports from MFN countries of origin to the U.S. from April 5th, 2025 (except imports of goods described in Section 3(b) of that executive order, imports of goods from Canada and Mexico that are eligible for preferential duties under USMCA, and imports of electronics described in a subsequent executive order on April 11th, 2025)
• Add Merchandise Processing Fee (MPF) in the U.S.
• Add Harbor Maintenance Fee (HMF) in the U.S.
• Include historical and future tax/tariff rates in the database with effective dates
• De minimis customs duty exemption, if applicable, except for imports to the U.S. that no longer qualify for de minimis treatment under Section 3(h) within President Trump’s executive order of April 2nd, 2025

Next, we put AI to the test by submitting the following prompt to five well-known large language models (LLM):

Please specify the exact computational procedure for determing (sic) the correct tariff amount imposed upon a given good imported into the United States on April 15th, 2025. Then, specify the exact computational procedure for determining the correct tariff amount imposed upon the same good into Thailand on the same date. You may assume that tariff rates, harmonized system (HS) codes, harmonized tariff schedules, and country-specific tariff schedules are stored in a database which may be queried in real time. Use pseudocode snippets or formulas, as needed, to specify all computational procedures.

Our prompt had one typographical error which we purposely did not correct to observe if this would affect the LLM responses.

The five LLMs we prompted were:

• OpenAI ChatGPT-4o
• OpenAI ChatGPT o4-Mini
• OpenAI ChatGPT 4o-Mini
• Microsoft Copilot
• DeepSeek DeepThink (R1)
• Google Gemini

We ran all our tests on Friday, April 25th, 2025. Only Google Gemini corrected the typographical error in our prompt.  

Results of the Experiment

Table 1, below, summarizes the experiment results.

In Table 1, an “omission” is defined as one of the fifteen above requirements not mentioned in the LLM response, whereas a “mistake” is one of those requirements which the LLM response mentions, but incorrectly. For example, the DeepSeek DeepThink (R1) LLM mentioned the Merchandise Processing Fee (MPF) in the U.S. but its response was judged to be a mistake because it quoted the wrong ad-valorem rate and maximum fee.

“Fitness for purpose” therefore is the percentage of requirements mentioned without mistakes.

Only one LLM, Microsoft Copilot, mentioned the tariffs which Trump ordered on April 2nd, 2025 but had not suspended by April 15th, 2025. However, at 363 words Copilot’s response was the briefest of all, omitting 60% of all requirements yet making 3 mistakes, so its fitness for purpose was 20%, the lowest of all models tested.

All the tested models achieved a failing grade as to their ability to satisfy all fifteen business requirements, where a passing grade is achieved with at least 60% fitness for purpose on a conventional academic grading scale (letter grades A, B, C or D). 

Conclusion

Our experiment conclusively disproves the hypothesis that an AI-driven ERP system can rapidly reduce response times by 40 to 60 percent, and maintain operational continuity, when unforeseen tariff challenges emerge. Such outcomes would only be possible if AI-driven ERP systems could detect most of the relevant tariff policies, and devise procedures to calculate most tariff costs accurately, but our experimental evidence shows that this is not the case. Current AI technology discovers less than half the relevant policies and performs incorrect calculations for over one-third of the policies it discovers.

Meanwhile, using our RADICORE rapid application development framework, Geoprise has already rolled out modifications to its GM-X ERP application suite that manage new master data requirements and accurately compute the new U.S. tariffs which meet all the business requirements enumerated above. Our clients who are exposed to business risks in consequence of the new U.S. tariffs are already benefiting from those enhancements.