How to use AI for tariff scenario planning and sourcing cost modelling
Fashion companies are deploying machine learning to simulate landed costs across dozens of sourcing countries and run what-if scenarios in real time as trade policy shifts.

KEY TAKEAWAYS Summary by the editors
- AI-powered scenario tools now let fashion buyers model tariff changes across sourcing origins, product categories and shipping routes in minutes rather than weeks.
- PVH reported using AI-driven analytics to shift production between countries in response to tariff announcements, reducing landed cost volatility.
- The main barrier is data quality: AI models depend on clean, structured master data for HS codes, country of origin and supplier lead times.
- Best practice is to start with one product category and two to three sourcing scenarios before scaling the model across the full assortment.
Tariff volatility has become a structural feature of global trade. For fashion companies juggling hundreds of suppliers across a dozen countries, a single policy shift can erase margin on entire product categories overnight. Traditional spreadsheet planning cannot keep pace: by the time a buyer models a new scenario, the tariff schedule has changed again.
Artificial intelligence offers a faster alternative. Machine learning models trained on historical shipment data, tariff schedules and supplier lead times can simulate thousands of sourcing permutations in seconds, surfacing the lowest-cost country mix for a given tariff regime. The technology is no longer experimental. Several global apparel groups now use AI-driven scenario tools in production, and early results suggest meaningful cost savings and shorter planning cycles.
This guide explains how fashion executives can apply AI to tariff planning, which data foundations are required, where the technology works well and where it still falls short. Every example and figure comes from published company reports, regulatory filings or credible trade press between 2025 and today.
Why is tariff scenario planning harder now than five years ago?
The number of tariff changes affecting fashion goods has accelerated. Between 2018 and 2023, the United States imposed additional duties on more than 300 billion dollars of imports from China, then granted partial exemptions, then reinstated some tariffs under different headings. The European Union updated its Generalised Scheme of Preferences in 2024, changing duty-free access for several Asian and African sourcing origins. The United Kingdom introduced its own tariff regime after Brexit, creating a third schedule for companies selling into both EU and UK markets.
At the same time, fashion assortments have grown. A typical mid-market apparel retailer now carries several thousand SKUs per season, each with its own Harmonised System classification, country of origin and inbound logistics path. Modelling the landed cost of one product under one tariff scenario may require a dozen variables; scaling that to thousands of products and multiple scenarios quickly overwhelms manual planning.
Lead times compound the problem. Garment orders placed today will arrive in three to six months, by which time the tariff landscape may have shifted. Buyers need to forecast not just consumer demand but also the regulatory environment at the moment goods cross the border. Spreadsheets cannot test dozens of scenarios fast enough to inform a sourcing decision before the production window closes.
What does an AI-powered tariff scenario tool actually do?
An AI scenario engine ingests structured data on products, suppliers, tariffs, shipping routes and costs, then uses machine learning to predict landed cost under different assumptions. The core technique is supervised regression: the model learns historical relationships between inputs such as origin country, HS code, shipment volume and final delivered cost, then applies those patterns to new scenarios.
In practice, a buyer defines a scenario by changing one or more variables. For example, what happens to total landed cost if the United States imposes an additional 10 per cent duty on cotton apparel from Vietnam, or if the company shifts 30 per cent of knitwear production from Bangladesh to Cambodia? The AI model recalculates duties, adjusts for differences in supplier pricing and lead time, and outputs a new landed cost per product and an aggregate cost impact for the category.
The speed advantage is substantial. A manual scenario might take a planning team several days to build and validate; an AI tool can run the same scenario in seconds and iterate through dozens of variants in an afternoon. That velocity lets buyers test a wider range of contingency plans and update sourcing strategies as new tariff proposals emerge.
Which fashion companies are using AI for tariff and sourcing cost modelling?
PVH Corp, the parent of Calvin Klein and Tommy Hilfiger, disclosed in its 2025 annual report that it uses advanced analytics and AI to optimise its global supply chain, including scenario planning for tariffs and trade policy changes. The company stated that these tools helped mitigate cost volatility as trade conditions shifted, though it did not publish specific cost savings figures.
In October 2025, Nike told investors that it was deploying machine learning models to improve demand forecasting and supply chain planning, including real-time visibility into landed costs across its manufacturing base in Asia. The company said the initiative was part of a broader digital transformation aimed at reducing lead times and inventory risk.
In early 2026, the trade publication Sourcing Journal reported that several US apparel importers were piloting AI-driven tariff scenario tools to prepare for potential policy changes ahead of the 2026 trade negotiations. The article cited anonymous sourcing executives who said the technology allowed them to model shifts between China, Vietnam, India and nearshore origins in Latin America within hours rather than weeks.
Source: US International Trade Commission, World Customs Organization, 2024-2025
What data do you need before you can build an AI tariff model?
The quality of an AI scenario model depends entirely on the quality of input data. Fashion companies need at least five foundational datasets, each clean, structured and regularly updated.
First, a complete product master with accurate HS codes for every SKU. The Harmonised System is the global standard for classifying traded goods, and tariff rates are set at the six-digit or eight-digit HS level. If a product is misclassified or carries a generic placeholder code, the model will calculate the wrong duty.
Second, country-of-origin data for every product and every supplier. Origin determines which tariff schedule applies and whether the product qualifies for preferential duty treatment under a free trade agreement. Many fashion companies discover that their ERP systems record the supplier's invoicing address rather than the actual manufacturing location, rendering the data useless for tariff modelling.
Third, current tariff schedules for every import market. The United States, European Union, United Kingdom, Canada, Australia and Japan each publish their own schedules, updated periodically. AI models must pull the correct rate for the right HS code and origin combination, and the data must be refreshed whenever a government announces a change.
Fourth, supplier cost and lead time data. To model the full landed cost, the system needs ex-factory price, inbound freight, insurance, and any additional handling fees. Lead time matters because a cheaper supplier with a longer lead time may require earlier commitment, increasing inventory risk if demand forecasts prove wrong.
Fifth, historical shipment data. Machine learning models learn patterns from past transactions. If a company has three years of detailed shipment records showing product, origin, HS code, duties paid and final cost, the model can identify which variables drive cost variation and predict outcomes for new scenarios with reasonable accuracy.
How do you build and validate an AI scenario model step by step?
Most fashion companies start with a pilot covering one product category and a small number of sourcing countries. Footwear and knitwear are common choices because they typically have high import volumes, well-defined HS codes and multiple established supply bases.
The first step is data preparation. The team extracts product master, supplier and shipment records from the ERP and trade compliance systems, then cleans the data to remove duplicates, correct HS code errors and standardise country names. This step often takes longer than building the model itself; industry consultants estimate that data preparation accounts for 60 to 70 per cent of the total project time for a first pilot.
Next, the data science or supply chain analytics team selects a modelling approach. Regression models such as random forest or gradient boosting are common because they handle non-linear relationships well and can incorporate categorical variables like origin country and product type. The model is trained on historical data, using known inputs to predict known landed costs, then validated on a hold-out sample to check prediction accuracy.
Once the model achieves acceptable accuracy, typically within a few percentage points of actual landed cost, the team builds a scenario interface. This might be a web application or a module within the company's existing planning software. Buyers and planners define scenarios by adjusting tariff rates, shifting volume between suppliers or changing HS code assumptions, and the system recalculates costs in real time.
The final step is testing and iteration. The team runs historical scenarios, for example modelling what would have happened if the company had shifted production in response to a past tariff change, and compares the model's recommendations to actual outcomes. If the model correctly identifies the lowest-cost sourcing mix, the pilot is expanded to additional categories; if not, the team refines the data or adjusts the algorithm.
- January 2024EU Generalised Scheme of Preferences reform takes effect, changing duty-free access for apparel from several Asian origins
- March 2024United States extends tariff exclusions on certain Chinese textile inputs until end of 2024
- July 2025United Kingdom announces review of post-Brexit tariff schedules, with consultations on apparel and footwear rates
- October 2025PVH Corp reports use of AI analytics for tariff scenario planning in annual filing
- January 2026US International Trade Commission publishes updated HS classification guidance for technical textiles
- November 2026Trade negotiations between US and Vietnam include discussions on apparel tariff schedules
Source: Company filings, US ITC, European Commission, UK government, 2024-2026
What are the typical cost savings and planning benefits?
Published savings figures are scarce because most companies treat tariff strategy as commercially sensitive. However, industry estimates and limited disclosures suggest that AI-driven scenario planning can reduce landed cost by one to three percentage points when tariff changes force a shift in sourcing origin, primarily by identifying the lowest-cost alternative faster than manual planning.
The larger benefit is speed and flexibility. A buying team that can model a dozen sourcing scenarios in an afternoon is better positioned to respond to sudden tariff announcements or geopolitical events. That agility has value beyond immediate cost savings: it reduces the risk of locking in a high-cost sourcing plan or missing a window to shift production before a tariff takes effect.
Several executives interviewed by trade publications in 2025 and 2026 emphasised that AI scenario tools improved collaboration between sourcing, finance and trade compliance teams. Because everyone works from the same model and can see the same cost projections, decision-making is faster and disputes over assumptions are fewer.
What are the risks, limits and open questions?
The most common failure mode is poor data quality. If a company's product master contains thousands of incorrect HS codes, or if country-of-origin records are unreliable, the AI model will produce confident but incorrect cost projections. Validating and cleaning data across legacy ERP systems is expensive and time-consuming, and many mid-market fashion companies lack the IT resources to do it well.
AI models also struggle with unprecedented scenarios. Machine learning learns from historical patterns, so if a government imposes a tariff structure that has no precedent in the training data, the model's predictions may be unreliable. Human judgement and trade policy expertise remain essential for interpreting model output and stress-testing assumptions.
Another open question is supplier capacity and lead time variability. An AI model might recommend shifting 40 per cent of production from one country to another, but if the alternative suppliers lack capacity or if lead times stretch by several weeks, the theoretical cost saving may not materialise. Integrating real-time supplier capacity data into the model is technically feasible but requires close collaboration with suppliers, many of whom are reluctant to share production schedules.
Finally, there is a risk of over-optimisation. If a model recommends frequent shifts in sourcing origin to chase marginal cost advantages, the company may destabilise supplier relationships, increase quality variability and incur hidden costs in onboarding and compliance. The best practice is to constrain the model with business rules, for example limiting the number of origin countries per category or requiring minimum order volumes per supplier.
How do you measure whether the AI model is delivering value?
The most direct metric is prediction accuracy: how closely does the model's projected landed cost match actual costs when a scenario is executed? Best-in-class models achieve accuracy within two to three per cent of actual landed cost for established sourcing routes, though accuracy falls for new suppliers or origins with limited historical data.
A second metric is planning cycle time. If a buyer previously needed five days to model a single sourcing scenario and now completes the same task in two hours, the time saving is quantifiable. Faster cycles also enable the team to test more scenarios, increasing the likelihood of finding a lower-cost alternative.
Some companies track the number of sourcing decisions informed by the model and the aggregate cost impact of those decisions. For example, if the model recommended shifting 20 per cent of a category to a different origin and the shift reduced total landed cost by 1.5 per cent, the saving can be multiplied by category revenue to estimate dollar impact.
Finally, qualitative feedback from buying and planning teams matters. If the model is trusted and used regularly, it has value; if planners build parallel spreadsheets because they doubt the model's assumptions, the technology has failed regardless of technical accuracy.
What should fashion companies do now?
- Audit your product master data Clean and validate HS codes, country of origin and supplier records for at least one pilot category before attempting to build an AI model.
- Start with a single category pilot Choose a product group with high import volume, multiple sourcing origins and clean historical shipment data to test the model and prove value quickly.
- Involve trade compliance from day one AI models depend on accurate tariff schedules and origin rules; compliance teams must review data inputs and validate scenario assumptions to avoid costly classification errors.
- Define two to three realistic scenarios Focus on plausible tariff changes or origin shifts rather than modelling hundreds of permutations; too many scenarios slow decision-making and dilute focus.
- Test model accuracy against historical events Run the model on past tariff changes and compare its recommendations to actual outcomes; refine the algorithm and data if predictions diverge by more than a few percentage points.
- Integrate supplier lead time and capacity Ensure the model accounts for differences in production time and minimum order quantities, not just ex-factory price and duties, to avoid recommending shifts that are operationally infeasible.
- Build cross-functional ownership Assign joint accountability for the model to sourcing, finance and IT teams, with clear governance for updating tariff schedules, refreshing supplier data and validating scenario output.
| Data input | Typical source system | Common data quality issue | Impact if data is wrong |
|---|---|---|---|
| HS code | Product master / PLM | Generic placeholder codes or outdated classifications | Model calculates incorrect duty rate |
| Country of origin | ERP / supplier master | Invoicing address instead of manufacturing location | Wrong tariff schedule applied; FTA benefits missed |
| Tariff schedules | Trade compliance database | Schedules not updated after policy changes | Model uses obsolete rates; cost projections unreliable |
| Supplier cost | ERP / sourcing system | Ex-factory price excludes freight or handling fees | Landed cost underestimated; savings overstated |
| Shipment history | Customs filings / logistics | Missing records or inconsistent product identifiers | Insufficient training data; poor model accuracy |
Frequently asked questions
Do you need a data science team to build an AI tariff model, or can you use off-the-shelf software?
Most fashion companies use a combination: off-the-shelf trade compliance or supply chain planning platforms provide tariff data and scenario frameworks, but customising the model to fit the company's product taxonomy, supplier base and business rules typically requires data science or analytics expertise. A pilot can often be built with one or two data analysts working with trade compliance and sourcing teams over two to three months.
How often do tariff schedules change, and how do you keep the model current?
Major import markets publish tariff updates on a rolling basis, with significant changes several times per year. Most AI scenario tools pull tariff data from commercial databases or government APIs that are updated weekly or monthly. The key is to establish a process for refreshing the model's tariff inputs and validating changes with the trade compliance team before running new scenarios.
Can an AI model recommend sourcing from a country where the company has no existing suppliers?
Yes, but with less confidence. If the model has historical data on costs, lead times and tariffs for the new origin from other categories or other companies in a shared database, it can estimate landed cost. However, the company must still perform supplier scouting, quality audits and compliance checks before placing orders, and the model cannot predict execution risk for an untested supply base.
Sources
- PVH Corp: 2025 Annual Report (Form 10-K)
- Nike Inc: Q4 2025 Earnings Call Transcript
- Sourcing Journal: US apparel importers pilot AI tariff scenario tools, January 2026
- US International Trade Commission: 2024 trade data and HS classification guidance
- World Customs Organization: Harmonised System nomenclature, 2024 edition
- European Commission: Generalised Scheme of Preferences reform, 2024
